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.1005087 | A Parsimonious Model of the Rabbit Action Potential Elucidates the Minimal Physiological Requirements for Alternans and Spiral Wave Breakup | Elucidating the underlying mechanisms of fatal cardiac arrhythmias requires a tight integration of electrophysiological experiments, models, and theory. Existing models of transmembrane action potential (AP) are complex (resulting in over parameterization) and varied (leading to dissimilar predictions). Thus, simpler models are needed to elucidate the “minimal physiological requirements” to reproduce significant observable phenomena using as few parameters as possible. Moreover, models have been derived from experimental studies from a variety of species under a range of environmental conditions (for example, all existing rabbit AP models incorporate a formulation of the rapid sodium current, INa, based on 30 year old data from chick embryo cell aggregates). Here we develop a simple “parsimonious” rabbit AP model that is mathematically identifiable (i.e., not over parameterized) by combining a novel Hodgkin-Huxley formulation of INa with a phenomenological model of repolarization similar to the voltage dependent, time-independent rectifying outward potassium current (IK). The model was calibrated using the following experimental data sets measured from the same species (rabbit) under physiological conditions: dynamic current-voltage (I-V) relationships during the AP upstroke; rapid recovery of AP excitability during the relative refractory period; and steady-state INa inactivation via voltage clamp. Simulations reproduced several important “emergent” phenomena including cellular alternans at rates > 250 bpm as observed in rabbit myocytes, reentrant spiral waves as observed on the surface of the rabbit heart, and spiral wave breakup. Model variants were studied which elucidated the minimal requirements for alternans and spiral wave break up, namely the kinetics of INa inactivation and the non-linear rectification of IK.The simplicity of the model, and the fact that its parameters have physiological meaning, make it ideal for engendering generalizable mechanistic insight and should provide a solid “building-block” to generate more detailed ionic models to represent complex rabbit electrophysiology.
| Understanding and preventing life-threatening irregular electrical heart rhythms includes basic experimental, numerical, and theoretical research. Computer models of the electrical dynamics of cardiac cells and impulse propagation throughout the heart are essential tools of this research. For example, simulations of electrical activity such as rotating ‘spiral waves’ have been used to understand how irregular heart rhythms are maintained. However, existing models are derived exclusively from sub-cellular data under a variety of environmental conditions and species. These models tend to be exceedingly complex including hundreds of variables and parameters which make them difficult to validate and analyze. Using experimental data from one species (rabbit) under nearly identical physiological conditions, we developed a simple model of the electrical activity of a cardiac cell derived from sub-cellular, cellular, and tissue experimental data. This model reproduces cellular excitability and its recovery as well as several important “emergent” phenomena including beat-to-beat cellular alterations and unstable spiral waves. Under some conditions, unstable spiral waves in this model give rise to continuous formation of new spiral waves (i.e., “spiral wave breakup”) which is thought to be the underlying cause of cardiac fibrillation and sudden cardiac death.
| The simulation of action potentials (APs) and their propagation has been an integral part of the field of electrophysiology for over half a century.[1] Over a hundred cellular AP models from many regions of the heart encompassing a variety of species have been published. These “ionic” models have been derived almost exclusively from voltage clamp data from a variety of species, often recorded under conditions that are not physiological. Ionic models are increasingly complex, and are comprised of numerous sub-models (mostly transmembrane currents), and contain tens of variables and hundreds of parameters. For example, a recent meta-analysis found that 50% of the data used in the development of the ten-Tusscher-Panfilov [2] and Iyer [3] human ventricular models were of non-human origin [4]. While their complexity has many advantages, there are a variety of limitations including: over-parametrization[5]; non-uniqueness/multi-stability [6, 7]; dissimilar predictions [8–12]; limited validation outside their development domain [13] (e.g. using a cell model derived from voltage clamp data to study reentrant arrhythmias); difficulty in representing natural “physiological redundancy” (e.g. repolarization reserve [14]); interpreting the relative roles of model sub-components on model predictions; and generalizing insight from “model-specific” simulations results.
An alternative, and complementary approach compared to these ionic models, is to use “phenomenological” models which are designed to represent specific, often macroscopic phenomena (e.g., rate dependence of action potential duration)[15–19]. Phenomenological models are designed to reproduce one (or two) specific phenomena(on) very well, and are: simplistic, computationally efficient, and sometimes amenable to analytical approaches. However, unlike ionic models, phenomenological models do not provide a direct link to physiologically meaningful model parameters derived experimentally. Therefore phenomenological models provide only limited mechanistic insight, and are not amenable for reproducing numerous phenomena, nor can they be easily extended.
The two most important membrane currents for cardiac excitability are the rapid sodium current (INa) and the rectifying potassium current (IK1), which are both highly non-linear functions of transmembrane potential (Vm). INa is responsible for the rapid all-or-none depolarization of the AP upstroke, and IK1 is responsible for maintaining the resting potential near -85 mV. In fact, it has been shown recently that transfecting cells with genes encoding for cardiac INa, IK1, and gap junctions Cx43 (only) into unexcitable somatic cells can transform monolayers into excitable tissue capable of supporting propagating “cardiac” action potentials including reentrant “spiral” waves![20]
Due to the ethical and practical limitations involved in studying human physiology, animal experiments are necessary to understand and develop clinical strategies aimed at treating the pathological mechanisms underlying heart disease. The isolated rabbit heart has been shown to share certain electrophysiological characteristics as the human, including repolarization reserve [21] and ventricular fibrillation dynamics.[22] as well as the effects of many pharmaceuticals.[23]
All existing ionic models of the rabbit ventricular AP [24–27] incorporate the INa equations from the LR1 guinea pig model,[28] albeit with different conductance values. The LR1 INa equations incorporate the gating kinetics of fast activation, fast inactivation and slow inactivation. The fast activation and fast inactivation equations were derived from voltage clamp (VC) data from spherical clusters of eleven day-old embryonic chick heart cells,[29] and slow inactivation parameters were based on VC results from sheep and pig ventricular trabeculae.[30] Obviously, there is a need to update rabbit AP models to incorporate an INa model based on rabbit data, but this is not a simple task. To date, INa activation parameters have been derived exclusively from VC protocols; however, since the INa current in the adult myocyte is too large to allow adequate voltage control under physiological conditions,[31] these VC protocols are carried out at low temperatures with low extracellular sodium concentrations or in HEK cells or oocyctes transfected with the SCN5A gene which encodes the cardiac INa channel (Nav 1.5).
The first aim of this paper is to develop a “parsimonious” model for the rabbit myocyte AP based only on data recorded from the rabbit under physiological conditions. A model is considered parsimonious if it accomplishes a desired level of explanation or prediction with as few parameters as possible (although, as far as we are aware, there is no definitive and scientifically rigorous definition of parsimonious). For this paper, the desired level of prediction is the following well-known and important electro-physiological phenomena, measured from rabbit ventricular myocytes/tissue under nearly identical and physiological conditions: 1) steady-state inactivation as determined from voltage clamp experiments [31]; 2) action potential depolarization in single cells; 3) recovery of AP excitability in single cells[32]; and 4) action potential depolarization dynamics during propagation in the whole heart. [33] For calibration and evaluation of the model we coupled this INa model with various models of IK1, and also explored a range of cellular and tissue behavior in an AP model incorporating a previously published phenomenological two parameter model of repolarization.[33] The overall model will be referred to as the parsimonious rabbit (PR) model. The second aim of this paper is to demonstrate that this simple and parsimonious model is sufficient for reproducing important “emergent” complex phenomena such as alternans and spiral wave breakup, which suggests a novel physiological mechanism for arrhythmia instabilities.
The equations of INa are of the form pioneered by Hodgkin-Huxley (HH) [1]
INa=gNam3h(Vm−ENa),
(1)
where Vm is the transmembrane potential, gNa is the maximal conductance of INa, m (fast activation) are gating variables, and h (fast inactivation), and ENa is the Nernst equilibrium potential for sodium. Beeler and Reuter [34] introduced a slow inactivation gate (j)
INa=gNam3hj(Vm−ENa),
(2)
which has been incorporated into all of the most recent cardiac HH INa models.[24–27].
The HH equations for the gating variables are of the form:
dydt=αy(1−y)−βyy≡y∞−yτy,
(3)
where y represents each gating variable, αy(Vm) and βy(Vm) represent the on and off rate constants for gating (respectively) which are voltage dependent, y∞(Vm) and τy(Vm) are the steady state fraction of activation or inactivation and time constant (respectively) and are functions of Vm. Hodgkin and Huxley [1] fit their voltage clamp data to empirically determine the functions αy(Vm) and βy(Vm) using only 9 parameters; in contrast the LR1 INa equations contain 31 parameters. Since 1952 there has been significant advancement in the derivation of gating equations based on first principles.[35, 36] The simplest functional form for the rate equations of voltage dependent gating based on thermodynamics and chemical reaction rates are:[37]
αy=αy0exp(−δyVmky),βy=βy0exp((1−δy)Vmky),
(4)
where the constants αy0, βy0, and 0 ≤ δy ≤ 1 are positive, while ky < 0 for activation and ky > 0 for inactivation. Using Eq 3 we can derive the equations for y∞(Vm) and τy(Vm) from Eq 4:
y∞(Vm)=11+exp[(Vm−Ey)ky],
(5)
and
τy(Vm)=2τy0exp[δy(Vm−Ey)ky]1+exp[(Vm−Ey)ky].
(6)
There are 4 parameters per gate (Ey,ky,τy0,δy) in this formulization. The main limitation of this formulization is that τm(Vm)→0 away from Em which is not realistic, because gating transformations cannot be instantaneous. For example, the LR1 equations predict τm(−85mV) = 0.0055 ms and τm(+30mV) = 0.041 ms at 37 C, however these values are below the resolution of voltage clamp measurements. We address this problem by removing the Vm dependence of τm (making it a constant, thus eliminating parameter δm). To summarize, we define the novel PR INa submodel using Eqs 1 and 3, with m∞ and h∞ given by Eq 5, τh given by Eq 6, and τm equal to a constant.
All model parameters are provided in Table 1. Values for Eh and kh where taken directly from Table 1 in [31]. We set ENa = 65mV, and all other parameters were determined via manual fitting to the experimental data described below using no more than 2 significant digits; τhmax and δh were computed via “inverting” Eq 6 given the two value pairs τh(−80mV) = 4.0ms and τh(0mV) = 0.45ms.
Simulations required for the calibration were carried out in single cells and a one-dimensional cable by integrating the three ordinary differential equations using exact integration for the gating variables and forward Euler integration for Vm with dt = 0.001 ms which is actually a very conservative choice made because of the low computational cost. The fact that τm is constant provides for a well-defined minimum time constant of the model which allows for the use of larger values of dt and computational efficiency compared to other models. [38] Cable simulations were performed using a central difference approximation of the Laplacian in a 4 cm long cable using Iion = INa + IK1L (in order to establish a steady-state resting membrane potential) with diffusion coefficient (D) equal to 0.001 cm2/ms and dx = 0.01 cm, where IK1L is the formulation of IK1 provided by Livshitz and Rudy [39]. INa and Vm signals were recorded at the center of the cable to construct the I-V curves for calibration to the experimental data as described below.
The current during the AP upstroke is predominately INa, therefore the precise time course of Vm(t) contains a wealth of information regarding INa kinetics.[40–42] In isolated myocytes the time course of INa can be estimated as Iion=−CmdVmdt (after the stimulus current is turned off), where Cm is the specific membrane capacitance which we assume is equal to 1 μF/cm2. During propagation, however, the total transmembrane current is not proportional to dVmdt; but we recently developed a methodology and justified the computation of INa as Cm(D(CV)2d2Vmdt2−dVmdt) during the AP upstroke during planar wave propagation, where CV is the conduction velocity [33]. These relationships were used to calibrate our model to dynamic I-V curves from glass microelectrode Vm measurements during AP upstrokes from isolated rabbit ventricular myocytes[43] and from the epicardial surface during planar propagation on the surface of the isolated Langendorrff-perfused heart.[33] In isolated myocytes Vm traces from APs stimulated with near threshold stimuli with a latency of 2.5 ms (range of 1.2 to 3.7 ms) were chosen from a previous study[43], because this latency provided datum during the early decrease INa of near –40 mV. Similar to Joyner et al. [32] we found a slight decrease (5.2 ± 0.5 mV/ms per ms) of (dVmdt)max as a function of latency which corresponds to only a 3.4% decrease compared to the traditional latency value of 1 ms.
It is well known that (dVmdt)max is decreased during the relative refractory period of the AP, and this is thought to be a cellular phenomenon [32]. Here we assume that this phenomena (“recovery of AP excitability”) is controlled only by the voltage and time dependent INa current and the AP shape. To reproduce the experimental results of recovery of AP excitability in isolated myocytes,[32] we developed a novel “virtual protocol”. This protocol involves simulating an action potential clamp of INa using APs measured during pacing from six isolated myoctes recorded in a previous study.[43] During the 11th beat the simulations were switched to current-clamp mode and stimulated with a 2 ms stimulus at various coupling intervals; the amplitude was adjusted such that the latency was 1 ms. The resulting peak INa was normalized to that recorded during the 10th beat and presented as a function of coupling interval.
Since the INa equations in all previous rabbit models are identical to LR1, except for scaling (conductances), we present previous model predictions using LR1 with two conductance values (23 mS/μF and 8 mS/μF) to cover the entire range (grey lines in figures). In addition we present results from the Ebihara-Johnson (EJ) model [29] (dashed grey lines) which is identical to LR1, except it does not contain slow inactivation.
To create our parsimonious rabbit (PR) cell model, we combined our model of INa with a previously published phenomenological model of repolarization (IK) designed to reproduce rabbit AP repolarization during propagation:
IK=gKe−b(Vm−EK)(Vm−EK),
(7)
where gK is the maximal conductance of IK, b is a parameter controlling AP shape, and EK is the reversal potential for potassium (set equal to the average resting potential of the six cells in Fig 1 of -83 mV); the nominal parameters values for IK are provided in Table 1, as previously determined.[33] We define our PR cell model via a total ionic current of INa + IK in Eqs 1 and 5–7 with parameters in Table 1.
The experimental traces of Vm during AP depolarization of isolated rabbit ventricular myocytes are shown in Fig 1A; the traces are aligned to (dVmdt)max. The average dynamic I-V curve during the AP upstroke for these traces are plotted as symbols (● representing the mean ± standard error of the mean) in Fig 1B. The calibrated PR results are shown as a thick black line. Previous model predictions (grey lines) bracket the range of peak INa but do not accurately capture the initial inward current deflection of the experimental I-V relationship (see box).
The experimental dynamic I-V curves during the AP upstroke during propagation from [33] are shown as symbols (●) in Fig 2A. Similar to above, previous model predictions (grey lines) span the range of peak INa but do not accurately capture the initial inward current deflection of the experimental I-V curve (see box). Although the PR model fits the initial decrease of INa and maximum Vm well, it does not match the experimental dynamic I-V curve during the second half of the upstroke.
It is important to note that the experimental dynamic I-V curves for isolated myocytes (Fig 1B) is different than that obtained during propagation (Fig 2A). To quantify these differences we fit both curves to fifth order polynomials and identified the regions in which the 95% confidence intervals of these fits did not overlap (see Fig A in S1 Text). The inward current during the upstroke for myocytes was more negative compared to during propagation for -41 mV < Vm < -22 mV and for Vm > +4.5 mV.
The normalized magnitude of peak INa as a function of recovery time are shown in Fig 3B; experimental results from [32] are shown as symbols (●), PR model results are shown as a thick black line, and previous model predictions shown in grey. Simulation results (i.e., lines) include standard error bars because the virtual protocol includes six different APs and thus include the effect of normal variation of AP shape. Unlike previous models with slow inactivation (Eq 2), the PR and EJ80 models which include only fast inactivation (Eq 1) are consistent with the experiments.
The voltage dependence of steady state INa activation and inactivation as well as the corresponding time constants are shown in Fig 4 for our PR model, as well as LR1, and EJ models. Our PR model exhibits a steeper activation slope at slightly less depolarized voltages compared to LR1 and EJ (panel A). As described above, τm in the PR model does not exhibit voltage dependence but is similar to the mean value of LR1 and EJ (panel B). The steady state INa inactivation for PR is more negative compared to LR1 and EJ (panel C) and the maximum value of τh for PR is less than that for τh for LR1 and EJ80 and is considerably less than τj for LR1 (panel D).
The PR results in Figs 1 and 2 represent the results from for our novel INa sub-model only; the results in Fig 3 include coupling to the IK1L sub-model. To assess the effect of repolarization current on the PR results shown in Figs 1–3 we reran all simulations with cell models incorporating IK1L or the phenomenological model of repolarization Eq 7 (PR cell model). The results from these simulations were nearly identical to the PR INa sub-model only; in fact the results are shown as thick dashed (IK1L) and dotted (IK) lines in Figs 1–3 but cannot be seen because they superimpose upon the solid black line.
Simulations using our PR cell model resulted in APs with the following characteristics: INamin=−233μA/cm2, Vmmax=34mV and APD = 197ms for single cells; and INamin=−289μA/cm2, Vmmax=24mV, CV = 55 cm / s and APD = 142ms for propagation in a cable. With the exception of APD, these values were not significantly altered (< 2%) compared to simulations using IK1L [39] or IK [33] indicating that the predictions of our PR INa sub-model are quite insensitive to the specific choice of repolarization current.
We performed parameter sensitivity analysis by quantifying the effect of a 1% variation of the eight INa gating parameters on eight quantities of interest; results are provided in the Table A in S1 Text as an 8x8 matrix. We also characterized the effect of AP shape on cell and tissue-level behavior for our new PR model (INa + IK) by varying repolarization parameters gK and b over the entire “physiological range” for all mammals: reported values of gK range from 0.1 to 0.5 mS/μF and the range of b values chosen, 0.03 ≤ b ≤ 0.05 mV-1, correspond to APDs ranging from 23 ms to 516 ms. The results are displayed in the Figs B and C in S1 Text. Over this range of gK and b, INamin varied 1.6% and Vmmax varied 1.2% in single cells, while during propagation the following quantities exhibited a fairly small dependence on IK parameters: INamin: 1.8%; Vmmax: 7.5%; CV: 4.5%. The fact that action potential depolarization characteristics are fairly insensitive to repolarization in single cells and during propagation provides support for the robustness of our INa model.
To confirm that the data used in the model calibration is representative, we compared several model results to experimental data from previously published studies. Table B in S1 Text demonstrates that the following model results are within the range found in the literature: resting membrane potential, action potential amplitude, (dVm/dt)max, and CV.
In addition, we compared the predictions of spiral waves in the PR and Mahajan (M08) models (5 cm x 5 cm) to those recorded in the whole rabbit heart by Schalij et al.[44] Schalij et al found that only fibrillation was observed in the intact heart but that, after freezing the inside of the heart, stable reentry occurred around a linear line of conduction block of ~ 2 cm with a cycle length of 130 ± 11 ms.[44] Simulation results for PR and M08 are shown in Fig 5. The cycle length (~ 150 ms) and line of block (~ 3.2 cm) for PR using the nominal parameter values were larger than the experiments. Incidentally, decreasing parameter b by only 4% decrease improved the correspondence considerably (cycle length: ~125 ms; line of block: 2.5 cm as shown in the middle, right panel of Fig 6 below). This change is justified because the PR model does not include any repolarization kinetics and is not developed to reproduce APD shortening with rate. In contrast to the experimental results of Schalij et al., [44] the line of block in our simulations was not static, but rotated which is typical of spiral wave simulations.[2] Simulations with M08 resulted in non-sustained unstable reentry with average cycle length of length (~ 150 ms); the tip trajectory traversed the majority of the region, and eventually terminated by hitting the top boundary. There was evidence of wave break (see green and blue trajectories in Fig 5B consistent with previous results indicating SWB for this model in a 7.5 cm x 7.5 cm region.
The level of validation required for a model depends on its “context of use” and the consequences of incorrect model predictions[13]. Hence further validation is expected to be required for each specific context. The preliminary validation performed here is appropriate for using the model to gain insight into underlying physiological mechanisms as demonstrated below.
During “functional” spiral wave reentry the depolarization and repolarization processes are interdependent. As shown in Fig 6, the dynamics of spiral waves in our model varied considerably as a function of repolarization (i.e., gK and b). Spiral waves were not stable (i.e., rotationally symmetric) for any parameter pairs (gK, b) in the range studied. For 8 of 9 parameter pairs, the spiral wave tip followed typical non-circular trajectories, while spiral wave breakup (SWB) occurred for one parameter set (gK = 0.5, b = 0.035). The frequency of activity for single spiral waves ranged from ~3 Hz to ~18 Hz and the frequency of activity during SWB was ~30 Hz.
Since SWB was observed (albeit for parameters corresponding to APD = 23ms which is not consistent with normal rabbit physiology) we investigated whether cellular alternans occurred in our PR model. We paced the PR cell model with a 2 ms stimuli of -20 pA/pF and found that alternans occurred for cycle lengths between 202 and 210 ms, as shown in Fig 7, which is within the range (190–240 ms) reported previously for isolated rabbit myocytes.[26]. Since the PR model contains no repolarization kinetics, we hypothesized that APD alternans occurred as a result of alternations of peak INa, which resulted in alternans in Vmmax, which caused alternans in APD. Since INa is negligible except during depolarization, APD is a function of only Vmmax (and can be computed by integrating Eq 7 with initial condition Vm=Vmmax). Since alternations in the variable h were more pronounced than m (see Fig 7B), we hypothesized that only INa inactivation kinetics are necessary for cellular alternans. A “reduced” two variable (Vm,h) model, in which variable m3 was replaced with function m∞3(Vm) and gNa was reduced to 5.8 mS/μF (to maintain CV = 55cm / s), confirmed that INa inactivation by itself was capable of generating cellular alternans in our PR model.
Traditionally, cardiac AP models have been developed from equations and parameters derived from voltage clamp experiments. Many currents have voltage-dependent gating variables whose characterization via voltage clamp provides intuitive and meaningful parameters. However, in order to obtain such parameters it is necessary to isolate individual currents either chemically, functionally, or genetically. In addition, the relationship of voltage-clamp derived parameters to AP characteristics in single cells and during propagation is complex, non-intuitive, and often involves the interaction of multiple variables.
Here we provide an alternative approach in which we calibrate our model in such a way as to reproduce certain AP behavior under physiological conditions while including all available voltage clamp data. In this way, our PR model is both physiologically meaningful and parsimonious. It is physiological because it incorporates HH-like equations and parameters with only one phenomenological parameter b which controls AP shape. It is parsimonious because it was constructed was constructed to reproduce the specific set of experimental data sets (Figs 1–3) that capture important electrophysiological behavior with as few parameters as possible. In addition, we have recently shown that this model is mathematically identifiable (i.e., not over parameterized)[45]. Incidentally, as described by Biktashev et al. [46] FitzHugh–Nagumo models fail to reproduce some features of cardiac ionic models such as: slow repolarization, slow sub-threshold response, non-Tikhonov asymptotic properties of excitability [47], wave front dissipation [48], and different action potential amplitude in single cells versus propagation [9] which our model captures (see Fig 2).
Physiological models are developed using a set of experimental calibration data, which for our model is for rabbit tissue and myocytes at physiological temperature and heart rate. Our approach should be relatively straight-forward to implement for other species, although this will require the high-fidelity measurement of the dynamic I-V curve during propagation [33] and in single cells as well as the recovery of excitability in isolated myocytes [32]. As far as we are aware, our PR model is the first to include dynamic I-V curves during propagation into its development (i.e., calibration) which we believe is superior to simply adjusting gNa to match CV. We believe that proper extensions of our PR model will include additional constraints obtained using appropriate data sets during recalibration. One reason we think our model is useful is because it has only one inward current which is non-zero only during depolarization, and only one outward current. Unlike traditional models in which multiple combinations of currents can give rise to the same AP shape [49], in our model repolarization is characterized only by a time-independent, voltage dependent current (IK), and therefore, IK parameters completely determines AP shape [33]. In addition, each current typically incorporates one or two variables (often gates). Despite the non-uniqueness of the relationship between model parameters and AP shape, model developers have largely ignored the interaction of the various currents and variables (components) in simulations. For a model with n > 1 components, the total number of possible interactions between two or more components, excluding the full model of all n components, is 2n − n − 2. Traditional models have at least 10 components, which results in over 1000 interactions! Characterizing these interactions appropriately both physiologically and mathematically continues to be a significant challenge.
[2, 31, 50, 51]The simplicity of our model allows unique physiological insight regarding the interaction of depolarization and repolarization processes. We found that IK conductance (gK) is related to the cycle length of reentry (see Fig 6) consistent with the experimental findings of Warren et al. in guinea pig [52]. In addition, our PR model exhibited the expected relationship between APD and spiral wave behavior [53]; specifically, the frequency of activity for single spiral was inversely related to APD, and also inversely related to the region covered by the spiral wave tip.
Surprisingly, our PR model (comprised of only a HH INa current with only two gating variables and a time-independent, voltage-dependent IK current) exhibited important “emergent” behavior including cellular alternans (Fig 7), unstable spiral waves (Figs 5 and 6), and even spiral wave breakup (Fig 6, bottom left panel). We initially thought that alternans and spiral wave breakup would require dynamic repolarization variables with time constants with values close to the APD capable of generating APD alternans. We only discovered spiral breakup serendipitously while performing a sensitivity analysis of the spiral wave meandering behavior on the IK repolarization parameters. It is important to note, however, that the parameter set for which breakup occurred (gK = 0.5,b = 0.035) resulted in a rate (~30 Hz) which is twice that observed in experiments.[54, 55]
While our PR model reproduces important physiologically meaningful behavior, it does have limitations. It should be noted that the alternans resulting from the interaction of INa and IK are the ‘minimal’ requirements, other currents such as Ito and ICaL may either enhance or suppress this behavior. Our PR model does not include intracellular calcium dynamics, nor the majority of repolarization membrane currents. It should be noted that the Tyrode’s solution used to perfuse the hearts from which the propagation I-V experimental data were obtained (Fig 2) in the presence of 15 mM diacetyl monoxime, however the longitudinal propagation velocity was 59 ± 3 cm/s which is within the range reported for the rabbit without any uncoupling agents (see Table B in S1 Text). The INa equations are of HH type, Markov models are required to replicate certain features of voltage clamp experiments including drug binding kinetics.[50] Importantly, the APD in our model is dependent on Vmmax which depends on the stimulus current, which is different in single cells compared to in tissue during propagation. When the PR model is stimulated to unphysiologically large Vm the APD is unphysiologically long; to avoid this in tissue simulations we initiated propagation in the cable by holding Vm at the end (1 mm) at 0 mV for 2 ms.
One significant difference between our INa sub-model and previous rabbit INa sub-models is that we did not include a slow inactivation gate. While slow inactivation of INa can be justified from voltage clamp experiments [31], we have found that it can induce unphysiological post repolarization refractoriness as shown in Fig 3 and [51]. In addition, the INa sub-model in the existing rabbit models are the same as LR1 in which the slow inactivation kinetics were derived from sheep and calf trabeculae muscles in which spatial homogeneity of Vm and VC control are lacking. We believe that much more research is needed to appropriately incorporate slow inactivation of INa (as well as late INa) into AP models, especially those used to simulate propagation. Incidentally, ten Tusscher et al. interpolated experimental time constants of fast and slow activation as a function of voltage over the range -80 to -40 mV in which data is not available to values more than four times values outside of this range (see Fig 1 D & E in [2]).
Our novel PR model provides unique and physiological generalizable insight into the relationship among gating kinetics and AP behavior in single cells and during propagation. We show unequivocally that only two currents (INa,IK) and three variables (Vm,m,h) are required to reproduce cellular alternans and spiral wave breakup. In fact, only two variables (Vm,h) are required to produce alternans and implicate a destabilizing effect resulting from the interaction of the non-linear rectification of IK [56] and the voltage and time dependence of h. We believe that our PR model is an ideal tool to quantify excitability and propagation related to a variety of important physiological and pathophysiological issues. For example our model can be used to study the effects of drugs (our model predicts that an 83% decrease of gNa is required for propagation failure), genetic mutations of SCN5A such as LQT3, and the differences between atrial and ventricular tissue. In addition we believe that it provides a solid “building-block” to develop more realistic rabbit AP models by including additional sub-models (e.g., ICaL, Ito, IKr and intracellular calcium handling).
All experimental data and PR simulation results presented in this paper (with the exception of the videos) will be made available online at Research Gate. The model will be uploaded to CellML.
|
10.1371/journal.pmed.1002675 | Climate change and African trypanosomiasis vector populations in Zimbabwe's Zambezi Valley: A mathematical modelling study | Quantifying the effects of climate change on the entomological and epidemiological components of vector-borne diseases is an essential part of climate change research, but evidence for such effects remains scant, and predictions rely largely on extrapolation of statistical correlations. We aimed to develop a mechanistic model to test whether recent increases in temperature in the Mana Pools National Park of the Zambezi Valley of Zimbabwe could account for the simultaneous decline of tsetse flies, the vectors of human and animal trypanosomiasis.
The model we developed incorporates the effects of temperature on mortality, larviposition, and emergence rates and is fitted to a 27-year time series of tsetse caught from cattle. These catches declined from an average of c. 50 flies per animal per afternoon in 1990 to c. 0.1 in 2017. Since 1975, mean daily temperatures have risen by c. 0.9°C and temperatures in the hottest month of November by c. 2°C. Although our model provided a good fit to the data, it cannot predict whether or when extinction will occur.
The model suggests that the increase in temperature may explain the observed collapse in tsetse abundance and provides a first step in linking temperature to trypanosomiasis risk. If the effect at Mana Pools extends across the whole of the Zambezi Valley, then transmission of trypanosomes is likely to have been greatly reduced in this warm low-lying region. Conversely, rising temperatures may have made some higher, cooler, parts of Zimbabwe more suitable for tsetse and led to the emergence of new disease foci.
| Tsetse flies are blood-feeding insects that transmit pathogens causing fatal diseases of humans and livestock across sub-Saharan Africa.
The birth and death rates of tsetse are influenced by environmental conditions, particularly temperature.
Since 1975, mean daily temperatures at Rekomitjie, a research station in the Zambezi Valley of Zimbabwe, have risen by c. 0.9°C and temperatures in the hottest month of November by c. 2°C. These increases in temperature may have impacted tsetse populations and the diseases they transmit.
Since the 1960s, wild tsetse have been caught regularly from cattle for insecticide tests conducted at Rekomitjie.
Prior to 1990, catches were on average >50 flies per animal per afternoon; however, in 2017, teams frequently failed to catch a single fly.
A mathematical model of tsetse population change, which included temperature-dependent rates for births and deaths, suggests that the decline in tsetse is related to rising temperatures.
Our findings provide a first step in linking the effects of increasing temperatures to the distribution of diseases caused by tsetse.
If the effect extends across the Zambezi Valley, then tsetse-borne disease is likely to have been reduced across the region. Conversely, rising temperatures may have made some higher, cooler areas more suitable, leading to the emergence of new disease foci.
| Tsetse flies (Glossina spp.) transmit protozoa of the genus Trypanosoma that cause sleeping sickness—human African trypanosomiasis (HAT)—in humans. The initial phase of HAT is characterised by intermittent fever and joint pains; thereafter, there are sleeping difficulties and confusion. Without treatment, the disease is fatal. Parasites of this genus also cause nagana—animal African trypanosomiasis (AAT)—in livestock.
In 2015, HAT was responsible for c. 202,000 Disability-Adjusted Life Years (DALYs) [1]. The most recent global estimates indicate that AAT kills approximately 1 million cattle per year [2], with approximately 55 million cattle at risk [3].
In addition to the DALYs resulting from HAT, AAT also has substantial impacts on human health by reducing the supply of meat and milk, as well as animal draft power for crop production. These losses affect not only human nutrition but also the agricultural incomes that allow access to education and healthcare [4]. A study in 1999 indicated that the annual economic losses from meat and milk production alone were c. US$1 billion at current prices [5].
In Africa, there has been an increase in temperature of c. 1.5°C between 1900 and the 1990s [6]. However, the effects of recent and future climate changes on the distribution of tsetse and other vectors, and their associated diseases, remain poorly understood [7,8]. There is disagreement, for example, about whether the resurgence of malaria in the East African highlands in the 1990s was caused by rising temperatures or by increasing levels of drug resistance and decreasing control efforts [9–13]. Resolution of the debate is made more complex by the apparent absence of data on changes in vector population levels and biting rates.
Increases in global temperatures since the late 1800s [14] have led to shifts in the ranges of many animals [15]. Insects in particular are sensitive to changes in temperature, with consequences for the transmission of vector-borne pathogens [7,16,17]. Mechanistic models, capable of explaining how recent climate change [14] has affected vector distribution and abundance, could be used to predict future disease risks [16], but existing studies often rely instead on statistical correlations [8,18–20].
In general, the ways in which climate change will affect infectious disease burden in sub-Saharan Africa is poorly understood because of a lack of empirical evidence [21]. It has been suggested that requirements for accepting a ‘causal’ relationship between climate change and changes in human health outcomes for vector-borne diseases should, as a minimum, include (i) evidence of biological sensitivity to climate, (ii) meteorological evidence of climate change, and (iii) evidence of entomological and/or epidemiological change in association with climate change [8].
For vector-borne diseases, the difficulty is the ability to separate climatic effects from those originating from other environmental, ecological, and sociological changes influencing the population dynamics of parasites and vectors. Contributing to this difficulty is a lack of contiguous data on vector abundance and detailed records of local climate. Work on tsetse and trypanosomiasis carried out at Rekomitjie Research Station in the Mana Pools National Park, Zimbabwe over the last 59 years provides a valuable exception to this rule, producing long-term datasets for both vector abundance and climate profile.
The study site is located >10 km inside a protected area (S1 Fig). According to the World Database on Protected Areas (https://protectedplanet.net/), the Mana Pools National Park, together with its adjoining Sapi and Chewore Safari Areas and the adjacent Hurungwe Safari Area, has a total area of 9,660 km2. It has been free of agricultural settlement since 1958, when the people living there were relocated [22]. Since then, the combined area has been protected against settlement, agriculture, and illegal hunting and logging and was designated a UNESCO World Heritage Site in 1984. In this area, HAT occurs, and tsetse populations have not been exposed to any form of control. In addition, being situated in a protected area, the region has not been subject to other deliberate environmental or sociological change. Analyses by Hansen and colleagues [23] show that this area experienced <0.5% woodland loss between 2000 and 2010, with the majority of the 30 m × 30 m pixels in the Hansen and colleagues dataset within Mana Pools consisting of at least 10% wooded cover (S1 Fig). In addition, an aerial survey for elephant and buffalo in 2014 [24] indicated that, in the 200 km2 around Rekomitjie, there was an average of 1.6 elephants and 7.3 buffalo per km2. Vale and colleagues [25] showed that c. 2 elephants per km2 can provide about half of the diet of savanna species of tsetse and can support a population of flies even when alternative hosts are heavily depleted.
The data available therefore provide the possibility of developing temperature-driven models for tsetse population dynamics. Such models could be used to predict the present and future distribution of tsetse in Africa. Given that there is never any cyclically transmitted African trypanosomiasis without the presence of tsetse, such models will provide a more powerful approach for estimating changes in the distribution of human and animal trypanosomiases.
Tsetse flies are poikilotherms, and their development and mortality rates are dependent on temperature [26–30]. We aimed to use data on temperature and tsetse abundance at Rekomitjie to test whether observed increases in temperature over recent years are sufficient to explain the observed decline in the local tsetse population since the 1990s. To do this, we used a mechanistic model of tsetse population dynamics that incorporates the effect of temperature on adult and pupal mortality and rates of larval deposition and pupal development, established from laboratory and field studies [26–30]. We fitted the model to a 27-year dataset of Glossina pallidipes numbers caught from bait oxen.
The methods for the production of data for tsetse and climate were not guided by an analysis plan for the present study. Instead, the climate data were produced as a standard procedure at the research station over the past 59 years, and the tsetse data were obtained from previous studies [31–36].
Daily records of rainfall and minimum and maximum temperature have been kept at Rekomitjie since 1959. Staff at Rekomitjie, operating in accord with directions from the Zimbabwe Department of Meteorological Services, made recordings at 7:00 AM each day from maximum and minimum mercury thermometers housed in a Stevenson screen. The location of the screen is at 16° 10′S 29° 25′E, altitude 503 m. To quantify changes in the mean temperature over time, we first calculated mean monthly temperatures between October 1959 and June 2017. Then, using a reference period between January 1960 and December 1989, we calculated monthly temperature anomalies by subtracting the reference mean from the actual mean. We smoothed the temperature anomaly data using a 5-year running mean, as done for previous analyses of regional and global changes in temperature [37–39]. In addition, a time series linear regression model was fitted to the mean monthly temperature data using the ‘tslm’ function from the forecast package [40]—a wrapper for fitting linear models allowing for a trend variable. We subsequently employed the fitted trend to estimate the change in monthly temperature, from the peak in 1975 to the peak in 2017, and the 95% prediction intervals, using the forecast function in R [41].
Sampling of tsetse at Rekomitjie, in pursuit of various ecological and behavioural studies, has suggested a decline in tsetse abundance in the last two decades. It is difficult to interpret the catches confidently because they have been made using widely different methods at irregular intervals. From 1966, however, fed female G. pallidipes have been collected from stationary oxen at Rekomitjie, with the sole original aim of providing test insects for bioassays [31–36]. Because these collections were made using a single sampling system, run at approximately the same time each day, the change in the numbers collected offer an indication of the extent of the decline in tsetse abundance over recent decades.
Catches were made for 3 hours in the afternoon during the period of peak tsetse activity [42]. Each collection team comprised two hand net catchers and an ox, operating within 2 km of the research station. Each team operated at least 200 m from other teams, in areas chosen to maximise catches in accord with seasonal changes in the distribution of tsetse between vegetation types [43]. In the 1960s, it was usual for each team to take enough tubes to collect a maximum of about 50 flies each day. This quota was set in consideration of the minimum expected catch at that time and has been maintained at this level ever since, even though it has proved impossible to meet the quota in the last two decades. Daily records are available from 1990 for the number of catching teams employed, and for the catch of each team. The monthly averages of the number of flies caught per team per day are taken as indices of fly abundance. Prior to 1990, tsetse catches regularly reached the upper limit of 50 flies; thereafter, this hardly ever occurred. Fitting the model only to catch data for the period after 1990 ensured that there was no truncation of data used in the fitting procedure.
Tsetse females give birth, approximately every 7 to 12 days [28], to a single larva, which immediately burrows into the ground and pupates, emerging 30 to 50 days later as a full-sized adult [44]. Female adult flies can live for up to 200 days [45]. As quantified by researchers in the laboratory and field, larviposition and pupal emergence rates are dependent on temperature, as are the mortality rates of both pupae and adults [26–28,30]. Preliminary analyses suggested that the inclusion of temperature-dependent mortality rates was sufficient to model the observed decline. In response to suggestions from peer reviewers, we also reanalysed the data using models that included functions for temperature-dependent larviposition and pupal emergence rates.
Hargrove [29,30], using data from mark-recapture experiments on Antelope Island, Lake Kariba, Zimbabwe, showed that for G. pallidipes, female adult mortality increases with temperatures above 25°C. In our ordinary differential equation (ODE) model of tsetse population dynamics, described below, we therefore model female adult losses per day (μA) due to temperate-dependent mortality using
μA={a1T≤25a1exp(a2(T-T1))T>25
(1)
where T is temperature in°C. T1 is not a parameter but is a constant set to 25 to ensure that a2 is in a convenient range.
For pupae, the relationship between mortality rate per day (μp) and temperature was quantified by Phelps [27] in the laboratory. The data from these experiments show that pupal survival to adulthood is highest for temperatures between about 20°C and 30°C. As temperatures depart from this range, the mortality rises sharply, resulting in a U-shaped curve. This form of relationship has also been documented for various other insects [46] and, for tsetse, can be suitably modelled using the sum of two exponentials:
μp=b1+b2exp(-b3(T-T2)+b4exp(b5(T-T3))
(2)
where T is temperature in°C. T2 and T3 are not parameters but are constants chosen to ensure that the coefficients b3 and b5 are in a convenient range and were set to 16°C and 32°C, respectively.
Phelps also quantified the daily rate of pupal development (β) in G. m. morsitans as a function of constant temperature in the laboratory, fitting the data using the function [29]
β=c1/(1+exp(c2+c3T)
(3)
where, for females, the fitted estimates were c1 = 0.05884, c2 = 4.8829, and c3 = −0.2159.
The effects of temperature on pupal development and mortality rates in the laboratory are supported by work showing similar effects in the field [44,47–49].
Lastly, using ovarian dissection data from marked and released G. m. morsitans and G. pallidipes at Rekomitjie, Hargrove [28] showed that the larviposition rate per day (ρ) increases linearly between 20°C and 30°C. We therefore assume a linear increase in larviposition rate with temperature using the equation
ρ=d1+d2(T-T4)
(4)
where T4 was set to 24°C. The time taken for a female tsetse to produce her first larva is longer than for subsequent larvae. Accordingly, the values for d1 and d2 in Eq 4 are lower for nulliparous females (d1 = 0.061 and d2 = 0.002 [ρn]) than for parous females (d1 = 0.1046 and d2 = 0.0052 [ρp]) [30].
Considering the above temperature-dependent processes, and using the outlined functions for the five parameters μA, μP, β, ρn, and ρp, we model changes in the numbers of G. pallidipes female adults (A) and pupae (P) using three ODEs:
dPdt=ρnAn+ρpAp-(β+μP+δP)P
(5)
dAndt=β2P-(μA+ρn)An
(6)
dApdt=ρnAn-μAAp
(7)
Pupae are produced by nulliparous (An) and parous (Ap) adult females at rates ρn and ρp, respectively. Losses from the pupal stage are due to pupae emerging as nulliparous adults (An), at rate β/2, to density-dependent mortality, with coefficient δ and mortality μP. Losses from the nulliparous adult stage are due to first larviposition at rate ρn and mortality (μA), assumed equal for both nulliparous and parous females.
As initial starting estimates for the parameters in the model described in Eqs 5–7, we used the published [26,28,30] fitted values for larviposition and pupal emergence rates as described above (Eqs 3 and 4). For adult and pupal mortality, we fitted the functions in Eqs 1 and 2 to published data—described above and elsewhere [27,29,30]—using nonlinear least squares regression.
It was not necessary to vary all parameters in the ODE model to get a reasonable fit to the bioassay catch data. The only parameter in the population dynamic model for which we did not have an initial starting estimate from published data was the density-dependent mortality coefficient (δ). For model fitting, therefore, we first allowed only this parameter to vary while keeping all other parameter values fixed. We then fitted the model to the average monthly tsetse catches allowing just δ and the parameters for adult mortality (a1 and a2) to vary, followed by those for just pupal mortality (b1 to b5) and lastly for δ and both mortality functions. For pupal mortality, it was only necessary to fit b1, b3, and b5 in the ODE model. Model fits to the data were compared using Akaike Information Criterion (AIC).
As a preliminary to the data fitting procedure, the model was run for 5 years prior to the start of the first month of available temperature data in October 1959 using the average monthly temperatures from October 1960 to September 1961 because we did not know initial starting values for numbers of pupae, relative to the numbers of fed female adults caught. The initial number of parous adults and pupae at time t = 0 was set to 100 and the number of nulliparous adults to 25, and the model was solved at monthly time steps for comparison with values from the bioassay catch data. We fitted the model to the data using maximum likelihood, assuming the data were Poisson distributed. For 80% of months, the variance to mean ratio for the daily catch data was less than 1.5, and was between 1.5 and 4.0 for 20%, indicating that, in most circumstances, the variance was approximately equal to the mean. For each set of parameters, we first estimated parameter values using the stochastic simulated annealing algorithm [50] and then used updated parameter estimates in a final fit using Nelder-Mead [51].
A penalty was incurred for parameter estimates of a1 greater than 0.04 or less than 0.01, ensuring baseline adult mortality was within biologically reasonable limits, by stopping the function before computing the likelihood and automatically assigning a high negative log likelihood value [45]. A penalty was also incurred for model fits for which, on average, there were fewer than 50 tsetse between January 1965 and December 1984 because, during that period, sampling teams consistently obtained their quota of 50 flies in an afternoon. Confidence intervals (95%) were calculated for fitted parameters using the Fisher information matrix.
A peer reviewer noted that these confidence intervals allow for no uncertainty in the fixed parameters. To explore the importance of this, we refitted the model using the upper and lower limits of the 95% confidence intervals of the fixed parameters b1, b3, and b5 of the function for the temperature dependence of pupal mortality. All analyses were done in R [41] and are available online, with all the data required to reproduce figures, at the following website: https://github.com/jenniesuz/tsetse_climate_change.
Although there is considerable seasonal and interdecadal variation in temperature (Fig 1A), our analyses indicate an increase of c. 0.9°C from the peak in 1975 to the peak in 2017 (Fig 1B). This increase is not even across the year, being greatest in November when temperatures are already highest. During this month, mean daily temperatures increased by c. 2°C between 1975 and 2017 (Fig 2). In addition, the number of consecutive years in which the hottest mean monthly temperature was above 30°C has increased since 1990 (Fig 1A).
Tsetse flies are poikilotherms, and their development and mortality rates are dependent on temperature [26–30]. We used four temperature-dependent functions, with starting parameters estimated from fits to published data for pupal and adult mortality, larviposition, and pupal emergence rates (Fig 3, Table 1), in an ODE model of tsetse population dynamics (Eqs 5–7).
The observed decline in catches of fed female G. pallidipes has continued to the present day, and the rate of decline has accelerated since 2010 to the point that teams now sometimes fail to catch a single fly in an afternoon (Fig 4). If these catches scale roughly with the population density of tsetse around Rekomitjie throughout the study period, the data suggest a steady decline in numbers over the last 27 years.
To simulate this decline, the model was run using mean monthly temperatures between October 1959 and June 2017. Model fits to the monthly tsetse catch data for 1990 to 2017 (Fig 4), varying δ and only the adult mortality parameters a1 and a2, while keeping all other parameters fixed, gave the lowest AIC of 1609 and provided a reasonable fit to the data (Fig 4). By comparison, varying only δ, or varying δ and parameters in Eq 2, or varying δ and parameters in both Eqs 1 and 2, produced AIC values of 2867, 1789, and 1764, respectively. Fixed and fitted parameter estimates for each function are summarised in Table 1. From 1959 until the mid-1980s, fitted model numbers of tsetse fluctuated between about 50 and 100 and then declined from c. 50 in 1990 to <1 in 2017, in good agreement with observed data. In addition, the fitted parameters a1 and a2 for adult mortality as a function of temperature (Eq 1) were similar to estimates from fits to the published mark-recapture data (S2 Fig, Table 1). The main difference was a higher baseline mortality for adults and a slower increase with temperatures above 25°C. When we carried out the sensitivity analysis, the upper and lower bounds for the coefficients for adult mortality were a1 = 0.024 and 0.030, and a2 = 0.145 and 0.198 (S1 Table).
Between 1959 and 2017, the pupal mortality rate was usually <0.005 in the fitted model. Prior to the 1990s, mortality was higher than this in October–December in 14 months over 30 years. In the 27 years since 1990, the pupal mortality rate was higher than this in 31 months during the hot-dry season. The hot-dry season is also the time of year when the modelled adult mortality was highest: >0.05 day−1. Adult and pupal mortalities in the fitted model were both above these levels in October and November more frequently in years after 1990. This is consistent with the idea that increasing temperatures during the hot-dry season are primarily responsible for the observed decline in numbers of tsetse at Rekomitjie since the 1990s, and particularly since 2000. Indeed, increases in mean daily temperatures have been most pronounced in November when temperatures are already highest (Fig 2).
The results of the preliminary analyses using constant values for larviposition and pupal emergence rate parameters are presented in S1 Text. The model with no temperature-dependent parameters did not provide a good fit to the data and had an AIC of 6762 compared to 2523 when adult temperature-dependent mortality was included (S1 Text).
While there are statistical models relating climate change to changes in vector populations [8,18–20], mechanistic models that relate climate change to data for the population dynamics of an important vector of human and animal pathogens are much less common. Our mechanistic model, incorporating the effects of temperature on mortality, larviposition, and emergence rates was sufficient to explain the observed decline in numbers of tsetse. The >99% decline in numbers reported here is comparable to the effects of successful large-scale tsetse control operations conducted in Zimbabwe.
Hargrove and Williams [52] found, similarly, that temperature was an indispensable factor in modelling tsetse population growth on Antelope Island, Lake Kariba, Zimbabwe. They had access to a wide range of measures of meteorological variables but found that once temperature had been included in their model, the addition of any further candidate variables—including rainfall, humidity, saturation deficit, and cloud cover—did not result in any improvement in the fit to the data. At Rekomitjie, over the whole study period, we had data only on temperature and rainfall. Nonetheless, the Antelope Island study suggests that we were unlikely to be missing other important climatological variables.
Our results provide evidence that locations such as the Zambezi Valley in Zimbabwe may soon be too hot to support populations of G. pallidipes. Similarly, G. m. morsitans populations at Rekomitjie are declining and might also be close to local extinction within the next few decades [53].
There are several biologically feasible reasons to expect increased tsetse mortality at high temperatures. Tsetse are poikilotherms, and their metabolic rate increases with temperature: adult tsetse therefore utilise their blood meal more rapidly at elevated temperature and must feed more frequently. But feeding is a high-risk activity, and increased feeding rates will likely result in increased mortality [54,55]. Tsetse use artificial refuge sites when ambient temperatures exceed 32°C [56], thereby reducing the temperatures they experience by up to 6°C during the hottest times of the day [57]. This behaviour reduces their metabolic rate but also reduces their chances of feeding. Hence or otherwise female tsetse have reduced fat levels and produce progressively smaller pupae as temperatures increase [58,59]. This has a knock-on effect on pupal mortality because smaller pupae can suffer very high mortality at elevated temperatures [47].
As temperatures increase, rates of pupal fat consumption increase linearly with temperature, whereas pupal duration decreases exponentially. The interplay between these rates results in fat levels at adult emergence being highest for pupae experiencing temperatures of about 27°C and progressively lower as temperatures increase above this level [26,47]. Reduced fat levels at adult emergence prejudice the chances of a teneral fly finding its first meal before fat reserves are exhausted and the fly starves or suffers excess mortality as a consequence of taking additional risks in attempting to feed [60]. The rate at which fat is used by teneral flies increases with temperature, exacerbating the above problems for the fly. There are also direct effects of high temperature on pupal mortality such that, when they are maintained at a constant level >32°C, no pupae emerge (Fig 3B) and all are found to have died before they utilised all of their fat reserves [47].
Few studies of vector-borne disease have been able to show a clear link between climate change and a change in either vector or pathogen population dynamics and subsequent disease burden [8]. Studies are frequently confounded by other environmental, ecological, and sociological factors, or the necessary empirical data are too difficult to collect. Although we acknowledge that this study presents only a first step in linking the effects of climate change to changes in the risk of acquiring a trypanosome infection, it suggests that climate change is already having effects on the density of disease vectors. In this respect, our study contributes vital analysis of long-term (>10 years) data in a region where temperatures have increased and where, as a consequence, the dynamics of a disease vector have also changed [8].
If these effects extend across the Zambezi Valley, then transmission of trypanosomes is likely to have been greatly reduced in this region. Conversely, rising temperatures may have made some higher, and hence cooler, parts of Zimbabwe more suitable for tsetse and led to the emergence of new disease foci. There is a pressing need to quantify the magnitude and spatial extent of these changes on tsetse and trypanosomiasis.
While there are no data on annual incidence of HAT from Zimbabwe to compare with the long-term data on tsetse populations, in the last 20 years, cases have been reported from the vicinity of Makuti [61], at the relatively high-altitude of c. 1,500 m, where tsetse populations are close to their low temperature limit. We are unaware of trypanosomiasis being reported from this area prior to the 1990s. This circumstantial evidence of the emergence of HAT in cooler regions of Zimbabwe raises the possibility of the resurgence of tsetse populations, and then of T. brucei infections, in parts of Southern Africa where they have been absent since the rinderpest epizootic of the late 1890s, apparently because the areas were too cold. Because tsetse dispersal is thought to arise through random movement [62], such a resurgence would come about where diffusion took tsetse to areas that are now climatically more suitable than they were in the recent past. Tsetse populations could only become established if, in addition, there were sufficient numbers of host animals and suitable vegetation to support tsetse. Hwange National Park in Zimbabwe and Kruger National Park in South Africa are examples of such areas, where suitable hosts and habitat for tsetse are abundant, and where tsetse did occur in the 19th century.
HAT is one of several vector-borne diseases for which detecting human cases is difficult even in countries with relatively strong health systems. In Uganda, for instance, it is estimated that for every reported case of Rhodesian HAT, another 12 go undetected [63]. In remote parts of the Democratic Republic of Congo (DRC), Central African Republic, and South Sudan, finding cases is even more difficult. In these circumstances, prospects for gathering data to detect or predict the impact of climate change on HAT seem poor. Models to predict where vectors are abundant, supported by xenomonitoring of tsetse populations for pathogenic trypanosomes [64], seem a more likely means of assessing the impact of climate change.
In general, if the temperature increase seen at Rekomitjie is reflected more broadly in the region, large areas that have hitherto been too cold for tsetse will become climatically more favourable and could support the flies if adequate hosts were available there [65].
In any region where the climate becomes more suitable for tsetse, there must, however, be adequate vegetation cover to provide shelter for the flies. The clearing of land for agricultural development, which is occurring at an accelerating pace in many parts of Africa [66], will reduce the vegetation cover and the densities of wild hosts in what has been termed the autonomous control of tsetse [67]. Any future predictions of the effects of climate change on tsetse populations and/or trypanosomiasis should consider these other confounding effects, as has been done for malaria [68]. Gething and colleagues [68] demonstrated that any future predicted changes in malaria due to climate would likely be magnitudes smaller than changes due to control and other anthropogenic factors.
Most (>95%) cases of HAT occur in Central and West Africa, where the important vectors are subspecies of G. palpalis and G. fuscipes, which are riverine tsetse. These species have very similar physiology to the savanna species of East and Southern Africa, including G. pallidipes, and hence we would expect that populations of riverine tsetse would decline if they were exposed to the temperature increases reported in the Zambezi Valley of Zimbabwe.
Over the past decade, c. 90% of all reported cases of Gambian HAT occurred in the DRC [69]. For the tsetse-infested regions of the DRC where HAT occurs (e.g., Provinces of Mai Ndombe, Kwilu, and Kasai), there are no data to suggest that climate change has had an impact on tsetse and HAT. For HAT foci in West Africa, however, there is some evidence that climate change has had an impact. First, Courtin and colleagues [70] describe a 100 km shift southwards in the northern limit of tsetse that they attribute to drought, rising temperatures, and increased human density. Regions where tsetse appear to be absent include areas in which sleeping sickness occurred in the 1930s. Second, Courtin and colleagues [71] report that, across West Africa, the more northerly foci of HAT located in Senegal, Mali, Burkina Faso, and Niger have ceased to be active. The authors attribute this change to increased densities of humans, anthropogenic destruction of tsetse habitat, and climate change. Medical surveys conducted between 2000 and 2006 did not detect any cases of HAT north of the 1,200 mm isohyet, and comparison of the 1,200 mm isohyet for the periods 1951–1969 and 1970–1989 show that it had shifted south.
For tsetse-infested areas of West Africa, Courtin and colleagues suggested that it is difficult to disentangle the effects that changes in land cover, host populations, rainfall, and temperature have on tsetse populations and sleeping sickness [70,71]. Studies are further confounded by the impact of large-scale medical interventions that have led to a decline in the annual incidence of Gambian HAT across Africa [72]. With such interpretive problems, there is a need for more studies of the present sort in which long-term measurements of tsetse abundance are made in wilderness areas where there is little change in land cover and host populations.
A limitation of our study is that the estimated confidence intervals for model-fitted parameters, such as those for adult mortality, are underestimates, in part because they incorporate only the uncertainty resulting from fitting the model with fixed values for other parameters and do not incorporate uncertainty in those fixed parameters. Another limitation is that we did not have sufficient data to test the predictive power of our fitted model.
Our deterministic model does, however, provide a good fit to available data for the change in tsetse abundance since 1990. Such models are less satisfactory for assessing if and when a population will actually go extinct because they predict that populations go to zero only as time goes to infinity. Ideally, therefore, future modelling should adopt stochastic approaches to predictions about tsetse extinction, but these would require detailed knowledge of population dynamics at very low density, such as the probability that male and female tsetse will meet in sparse populations. Unfortunately, our current knowledge of dynamics relates only to populations that are dense enough for convenient study. Nonetheless, present modelling does raise the possibility of the extinction of the Rekomitjie tsetse populations, particularly if temperatures increase further. Future research could also make use of the fitted model to make spatially explicit predictions about tsetse population dynamics for other regions of Zimbabwe and East and Southern Africa under future-predicted climate change scenarios.
|
10.1371/journal.pcbi.1005032 | Structure Prediction of RNA Loops with a Probabilistic Approach | The knowledge of the tertiary structure of RNA loops is important for understanding their functions. In this work we develop an efficient approach named RNApps, specifically designed for predicting the tertiary structure of RNA loops, including hairpin loops, internal loops, and multi-way junction loops. It includes a probabilistic coarse-grained RNA model, an all-atom statistical energy function, a sequential Monte Carlo growth algorithm, and a simulated annealing procedure. The approach is tested with a dataset including nine RNA loops, a 23S ribosomal RNA, and a large dataset containing 876 RNAs. The performance is evaluated and compared with a homology modeling based predictor and an ab initio predictor. It is found that RNApps has comparable performance with the former one and outdoes the latter in terms of structure predictions. The approach holds great promise for accurate and efficient RNA tertiary structure prediction.
| RNA is an important and versatile macromolecule participating in a variety of biological processes. In addition to experimental approaches, computational prediction of 3D structure of RNAs and loops is an alternative and important source of gaining structure information and insights into their functions. The prediction of RNA loop structures is of particular interest since RNA functions often reside in the loop regions and about 46% of nucleotides in an RNA chain remain unpaired. For this purpose, we develop an approach RNApps based on a probabilistic coarse-grained RNA model. The probabilistic nature of the model, together with a sequential Monte Carlo (SMC) growth algorithm, allows a natural and continuous sampling of structures in 3D space, making the approach unique. The coarse-graining nature of the model further increases the efficiency. Here we test this new approach with various types of loops, including hairpin loops, internal loops, and multi-way junction loops, and make comparisons with other structure predictors.
| RNAs are a type of macromolecule of crucial and versatile biological importance. In addition to their long-discovered functions of carrying genetic information and acting as a part of translation machinery, recently they were found to be able to participate in the regulation of gene expressions and protein synthesis, and to act as scaffolds for higher-order complexes and transmit signals between cells, etc [1–4]. To fully understand the function of RNAs, knowledge of their three-dimensional (3D) structure is often required. Although the most reliable sources of RNA structural information are experimental measurements from X-ray crystallography, nuclear magnetic resonance (NMR) spectroscopy, and cryoelectron microscopy, such experiments are costly or technically challenging due to the physical chemical nature of RNAs. As a result, computational prediction of RNA structures provides a valuable alternative source of gaining structural information. Many programs have been developed to assist RNA 3D structure prediction, including YAMMP [5], NAB [6], ERNA-3D [7], MANIP [8], S2S [9], FARNA [10], MC-Fold/MC-Sym [11], RNA2D3D [12], iFoldRNA [13, 14], NAST [15], Assemble [16], HiRE-RNA [17], FARFAR [18], RNABuilder [19], ModeRNA [20], OxRNA [21], 3dRNA [22], a coarse-grained model that includes salt effect [23], our pk3D [24], etc. However, this is not a complete list due to the rapid development in the field.
The prediction of the tertiary structures of RNA loops, as part of the effort of structure prediction, merits specific attention. First, this is because RNA functions often reside in loop regions and about 46% of nucleotides in an RNA chain remain unpaired, according to Dima and her colleagues’ research [25]. Some people mistake loop regions as unstructured. Some people mistake loop regions as structured. There are stacking interactions between neighboring bases, and other non-canonical contacts. These enthalpies sometimes dominate the entropic components so that order dominates. Other times entropy dominates. Second, the prediction accuracy of the loop regions is usually lower than that of the helical regions, mostly due to the high flexibility of loops and the difficulty in calculating the energy of non-canonical base pairs and base triples frequently observed in loops. Previously, several methods have been developed for loop structure predictions. For example, Das and co-workers developed a deterministic stepwise assembly (SWA) method, and with the Rosetta statistical potential this brute-force method either reaches atomic accuracy or exposes flaws in the energy function for a testing set of 15 RNA loops [26]. Liu and Chen designed a set of dinucleotide-based statistical potentials for RNA loops and junctions and combined them with their Vfold model, then they made predictions for the coarse-grained (CG) 3D structures of both RNA loops and junctions [27]. One unique advantage of their approach is its ability to go beyond the native structures by accounting for the full free energy landscape, including all the non-native folds. Frellsen and colleagues developed a program BARNACLE (BAyesian network model of RNA using Circular distributions and maximum Likelihood Estimation) to remove some important limitations associated with the discrete nature of fragment assembly methods [28]. BARNACLE is a probabilistic RNA model based on dynamic Bayesian network and is able to efficiently sample conformations in a natural, continuous 3D space. It is shown that it captures several key features of RNA structure, and readily generates native-like conformations for 9 out of 10 test structures, solely using coarse-grained base-pairing information.
In this work we develop an approach that is specifically designed for predicting the tertiary structures of RNA loops or missing fragments which can be part in helical regions and part in loop regions. The loops can be hairpin loops, internal loops, or multi-way junction loops. The approach includes a probabilistic CG RNA model, a sequential Monte Carlo growth method, a simulated annealing strategy, and a statistical potential used for scoring the generated structural candidates. The probabilistic model is inspired by BARNACLE. However, the original BARNACLE model considers all the seven torsional angles in the RNA chain, as is not always necessary since some dihedral angles are relatively rigid. The advantage of the new model developed here is multifold. First, the probabilistic nature of the model allows a continuous sampling in the conformational space and therefore is able to cover all the relevant conformations, as is difficult for fragment assembly methods or our previous one based on a discrete-state model [24, 29]. Second, by adjusting the probabilistic functions in the model, we are able to deliberately enhance the sampling in a specific phase space. Third, the coarse-grained model further increases the efficiency, making it possible to deal with large RNAs. We name this new approach as RNApps, short for RNA structure Prediction with a Probabilistic Sampling strategy. We test this approach with a benchmark set including nine loops, a 23S ribosomal RNA, and a non-redundant RNA 3D structure dataset containing 876 RNAs [30]. We also compare the performance of our approach with other predictors.
The atomistic structure of an RNA is defined by seven torsional angles, as illustrated in Fig 1A. However, the usage of such an all-atom model is expensive for structure modeling, particularly when high-throughput structure prediction is needed. In our approach, an RNA molecule is described by a CG model that includes a virtual bond model for the backbone and a reduced model for the nucleobases. The virtual bond model can be traced back to Olson [31], where the RNA backbone includes only two types of atoms, i.e., P and C4’. The model has been used in much work [32–35] and is also used here to describe the RNA backbone. For the representation of the RNA bases, three beads are used, including the atoms N1 and C2 in the purine (or pyrimidine) ring, and a virtual bead Bc (short for base center) at the geometric center of the six-membered ring containing atoms N1, C2, N3, C4, C5, and C6. In total there are five beads for each nucleotide in the CG model.
In the current model all the bond lengths and bond angles are fixed. Five torsional angles are defined based on the CG model, including θ, η, μ, ϕ, and ω, as illustrated in Fig 1B. Note that in the implementation there are actually two sets of such angles used for the purpose of growing RNA chains. The first set is defined in a “forward” way, i.e., from the 5’-end of the RNA to its 3’-end; and the other is defined in a “reverse” way, i.e., from the 3’-end to the 5’-end. These two sets are denoted by the subscript “+” and “−”, respectively. More specifically, from the 5’-end to the 3’-end, θ+ is calculated as the torsional angle formed by Pi-C4’i-Pi+1-C4’i+1, η+ by C4’i-Pi+1-C4’i+1-Pi+2, μ+ by C4’i-Pi+1-C4’i+1-Bci+1, ϕ+ by Pi+1-C4’i+1-Bci+1-N1i+1, and ω+ by C4’i+1-Bci+1-N1i+1-C2i+1; from the 3’-end to the 5’-end, θ- is calculated as the torsional angle formed by C4’i+1-Pi+1-C4’i-Pi, η- by Pi+2-C4’i+1-Pi+1-C4’i, μ- by C4’i+1-Pi+1-C4’i-Bci, ϕ- by Pi+1-C4’i-Bci-N1i, and ω- by C4’i-Bci-N1i-C2i.
Once all the torsional angles along the RNA chain are known, the tertiary structure of the RNA molecule can be built in a “growing” way by calculating the coordinates of the to-be-grown nucleotide based on the values of the torsional angles and the previously grown nucleotides. The growing process is realized in both the “forward” way and the “reverse” way, to take advantage of the position constraints as much as possible.
In the five torsional angles defined above, only three are independent, i.e., θ, η, and ϕ. This is because μ is linearly dependent on η (if grown in a forward direction) or θ (if grown reversely), and ω is approximately linearly dependent on ϕ, as supported by the statistics given in Fig 2 and S1 Table. The statistics are based on the RNA09 database compiled by Murray and co-workers by applying quality-filtering techniques (using resolution, crystallographic B factor, and all-atom steric clashes) to the backbone torsional angle distributions from a large RNA database [36]. Based on the above results the task of growing an RNA structure is reduced to determining the three torsional angles θ, η, and ϕ, and it is done in an iterative and probabilistic way, as illustrated in Fig 3. For example, if we are growing a loop from the 5’-end to the 3’-end, for a given θ + i, the value of η + i can be obtained probabilistically from the conditional probability p ( η + | θ + = θ + i ). Similarly, the value of ϕ + i can be calculated from the just obtained η + i and the conditional probability p ( ϕ + | η + = η + i ). The value of θ + i + 1 of the next nucleotide can be obtained from η + i and p ( θ + | η + = η + i ), so on and so forth. The loop may also be grown in a reverse direction, i.e., from the 3’-end to the 5’-end, and the procedure is similar but with a different set of conditional probabilities defined in a reverse way.
The conditional probabilities are calculated from the loop structures in the RNA09 database. Before the calculation, we first strip the helices in the structures. If two consecutive base pairs are formed, they are determined to be a helix. Three types of base pairs are considered, including A-U, G-C, and G-U. Take p(η+|θ+) as an example. We first calculate all the (θ+,η+) pairs in the loops in the database and plot the data points in a two-dimensional surface, as shown in Fig 4A. We then equally divide both torsional angles into K bins and calculate the conditional probability numerically as follows
p ( η + ( j ) | θ + ( i ) ) = n i j N i ( i , j = 1 , 2 , … , K ) , (1)
where θ+(i) and η+(j) are the center of the i-th and j-th bins, respectively, Ni is the number of data points with their θ+ values falling into the i-th bin, and nij is the number of data points with their θ+ values falling into the i-th bin and η+ falling into the j-th bin simultaneously. K is set to 72, and thus the bin width is Δ = 360°/K = 5°. The other conditional probabilities are similarly calculated. The conditional probabilities between η+ and θ+ and between ϕ+ and η+ are shown in Fig 4B and 4C, respectively. The latter are classified into four types according to the nucleotide to which ϕ+ belongs, while the former are not. This is because one (θ, η) pair involves two nucleotides and thus there are 16 types of dinucleotides; to classify the data into 16 types will lower the statistical quality since the number of data is limited.
Based on the probabilistically generated torsional angles, we build the CG structure of RNA loops/fragments by adding nucleotides one by one on the previously grown nucleotides with a sequential Monte Carlo (SMC) method [37], which is employed to prune the growing tree and bias the conformations to those having low energies and satisfying specific constraints. The constraints may include the excluded volume effect, the requirement of chain connectivity, experimental information, etc. The specific algorithm is similar to that used in our previous works [24, 29, 32, 38–41], but is improved here by growing the chain in two directions alternatively.
Specifically, for a loop sequence X1 X2⋯Xn whose structure is to be predicted, we grow it alternatively in two directions in 3D space from two anchor nucleotides to which this loop is attached. The algorithm also works if there is only one anchor, which happens when the loop is at one end of the RNA. To be more specific, for the i-th nucleotide to be grown, we randomly determine D = 10 possible positions of attaching it to the (i − 1)-th nucleotide (if grown in a forward way) by probabilistically generating the corresponding θ, η, and ϕ values based on the algorithm described in Fig 3 and the conditional probabilities exemplified in Fig 4. In order to increase the success rate of connecting the growing loop to the anchor nucleotide, we exclude the newly grown nucleotide whose P atom is farther from the P atom in the target anchor than a threshold, which is related to the number of nucleotides between these two nucleotides. This early pruning of the unlikely partial chain will also save a lot of computational time. After appending the D random configurations, we obtain a total of N = L × D partial chains, where L is the number of chains before the growth of the i-th nucleotide. If N is greater than a threshold M, we do a resampling procedure by randomly choosing M partial chains out of N with the probability of choosing the m-th one proportional to exp(−Em/RT1), otherwise we keep all the N chains. Here Em is the statistical energy of the m-th partial chain that will be described later. The temperature T1 is used to control the relative probabilities between the candidates and RT1 is set to 300 RASP energy unit.
The above procedure works for a single chain. For internal loops or multi-way junction loops with more than one chain, the procedure is similar. Take three-way junction loop as an example. We first build M conformations for the first chain using the SMC method described above. Next for each in M conformations, we grow D conformations for the first nucleotide in the second chain and perform resampling to select M conformations out of N = M × D ones. We repeat these steps for the rest nucleotides in the second chain and then for the third chain. Finally we get M chains after the whole three-way junction loop is built.
The parameters M and D are chosen as a compromise between the accuracy and efficiency of the SMC algorithm. A large value increases the accuracy while deteriorates the efficiency, whereas a small value does the opposite. We use M = 80 and D = 20 for reasonable run time. These values are the default in this work unless otherwise stated.
The growth procedure based on sequential Monte Carlo gives M structural candidates with low energies. These structures are further optimized with simulated annealing (SA) [42, 43]. During the SA procedure, we adopt the FRESS method (fragment regrowth via energy-guided sequential sampling) developed by Zhang et al. [44] to update the structures. In detail, for each among M structural candidates,
In both the SMC and SA procedures, we adopt the Ribonucleic Acids Statistical Potential (RASP) developed by Capriotti et al. [46]. In brief, RASP is an all-atom knowledge-based or statistical potential derived from a non-redundant set of 85 RNA structures. The statistical energy is calculated as a function of atom types, distance and sequence separation. It also includes a representation for local and non-local interactions in RNA structures. The threshold of sequence separation used to differentiate these two interactions is optimized with information theory. The base pairing and base stacking interactions are implicitly incorporated in the potential. More details of the RASP potential can be found in the reference [46]. The reason for choosing RASP over the other all-atom energy functions such as AMBER force field is multifold. First, it is a statistical potential and therefore more consistent with the statistical model developed here. Second, according to the authors RASP performs better than ROSETTA FARFAR force field in the selection of accurate models. Third, it is easier to be incorporated into our C++ code.
Since RASP is an all-atom statistical potential, the corresponding all-atom structures of the CG conformations generated by our approach need to be reconstructed. The procedure is as follows. Based on the atoms of the nucleobases, namely N1, C2, and Bc, the coordinates of the heavy atoms in the nucleobases are calculated with the pre-built templates taken from the RNA09 database. There are four such templates, built for A, U, G, and C, respectively. The coordinates of atoms in the backbone are calculated based on the coordinates of the atoms P, C4’, and N1 (in the bases of U and C) or N9 (in the bases of A and G) with a pre-built template. After the reconstruction we check if there are steric clashes between atoms. If any clash is found, the structure is discarded.
Here we test our approach and compare it with two other RNA tertiary structure predictors, i.e., RLooM [47] and iFoldRNA [13, 14], and also with RNAnbds [29], which was developed previously in our lab. RLooM is based on homology modeling and utilizes template structures extracted from a PDB database. iFoldRNA represents an ab initio way of prediction based on physical interactions and discrete molecular dynamics (MD) simulation. These two are selected to represent two very different strategies of structure prediction in the field. The testing set that appears in the original RLooM paper [47] is used for the comparison. This set contains 9 RNA loops with their length ranging from 6 to 17-nt. For RLooM, the RMSD values of the predicted loops with respect to the experimental ones were directly taken from the RLooM paper [47]. For iFoldRNA, they were obtained by feeding to the iFoldRNA-v2 server the whole sequences plus the corresponding base pairing information as constraints [14]. For RNApps, the input includes the whole sequence and the structure other than the loops. The output is the center of the largest cluster calculated from the trajectory of the SA simulation, as described in the Methods section. For a fair comparison, the RMSDs of the reduced backbones (P, O5’, C5’, C4’, C3’, and O3’ atoms) given optimal superposition with the experimental ones are calculated, following the same methodology as in RLooM and iFoldRNA. For the comparison with our previous RNAnbds, the RMSDs of all heavy atoms are used. The results are summarized in Table 1.
In the first part of the table, RNApps is compared with RLooM and iFoldRNA. It can be seen that for this specific testing set, RNApps performs comparably with RLooM or slightly worse in a few cases. This is normal since the homology modeling based predictors usually perform better than those based on energy rules. However, such predictors are sensitive to the testing set. For example, we tested RLooM with 28 fragments of length of 8-nt taken from the RNA pseudoknot 1E95 and found that RLooM failed for all of them. The reason is that there are no homological partners of this RNA in the RLooM database. In contrast, our method guarantees to give a reasonable prediction (data shown in S1B Fig). The comparison with iFoldRNA shows that RNApps gives smaller RMSDs in seven out of nine RNAs, and slightly larger RMSDs in two cases. Therefore, for this testing set taken from the third party (the RLooM paper), our method shows a better performance than iFoldRNA. However, it is also worth mentioning that iFoldRNA is not only a structure predictor, but also can be used for studying folding dynamics as a result of the employed molecular dynamics simulation.
We then compare RNApps with our previous RNAnbds. It can be seen that the new approach has a considerably better performance, giving smaller RMSDs for seven out of nine loops and two larger RMSDs. The improvement is mostly attributed to the feature that RNApps is able to sample the structural space in a continuous way, while RNAnbds is based on a discrete-state RNA model and thus unable to reach some structural regions. The different energy functions employed may also contribute. In Table 1 we also give the RMSDs of heavy atoms calculated after optimal superposition of the anchors—the predicted and the experimental loops themselves are not superimposed. These RMSDs characterize how well the relative positions of the loops with respect to their environment are reproduced. It can be seen that the RMSDs are only slightly larger than those calculated after optimal superposition of the loops, indicating that this performance is good.
In Fig 5 we superimpose the predicted structures with the experimental ones for the nine RNAs to make a visual comparison. Since the approach actually predicts a structural ensemble, for each sequence we also show the largest cluster to which the predicted structure belongs. It can be seen that for seven out of nine loops, the predicted structures are similar to their corresponding experimental ones. As for the structural feature of the cluster, the trends of the backbones correlate well with each other, while the bases are somehow dynamic if they do not form base pairs.
The loops in 1Q9A and 2CKY (Fig 5F and 5H) are exceptional. In 1Q9A the concerned loop is known as the bulged-G motif, which is a ubiquitous loop and provides specific recognition sites for proteins and RNAs [48]. More specifically, it is an essential part of the binding site for elongation factors in Escherichia coli 23S ribosomal RNA. In this motif, one strand forms a kink so that the bulged-G can form a base triple by non-canonical interactions, permitting interstrand, but disrupting intrastrand stacking. The kink in one strand, together with the S-turn in the opposite strand, appears to make the motif amphipathic, with one surface more polar and ready for interacting with positive patches on proteins or a divalent metal ion, and the other surface less polar and suitable for making RNA-RNA contacts [48]. The above features indicate that this motif is highly optimized for binding its partners, which should be considered when making structure predictions. The RNA 2CKY, which is a riboswitch specifically recognizing the thiamine pyrophosphate, has a similar situation [49]. The loop shown in Fig 5H is in a highly distorted region that binds its ligand. These results suggest that, the consideration of the binding partners, if they have, is necessary for a correct prediction of the tertiary structures of RNAs.
Our approach can also be used for reconstructing a missing RNA fragment, in addition to a single loop. The fragment can be a mixing of helical and loop regions. Here we show its performance with a 23S ribosomal RNA (PDB ID 1FFK) [50] consisting of about 3,000 nucleotides. In total we make 105 tests, and in each test we delete one fragment of length N and then reconstruct its tertiary structure, pretending that the fragment is missing in the RNA structure. The fragments are chosen every S nucleotides along the sequence, where S is set to 25 so that the fragments are uniformly distributed in the whole RNA. The length N can be 5, 8, and 10, and the percentage of nucleotides in the helical region in the 105 fragments is 36%, 39%, and 39%, respectively. According to Fig 6, the percentage of RMSDs smaller than 4 Å is 76.2%, 71.4%, and 56.2% for the fragments of length 5, 8, and 10, respectively. The average RMSDs in the three cases are 2.55 Å, 3.23 Å, and 3.97 Å, respectively.
We further test our approach with a much larger dataset—RNA 3D Hub [30]. It contains 876 non-redundant RNAs (release ID 1.89), including all types of base-pairing, base-stacking, and base-backbone interactions. Out of this dataset we select 1768 hairpin loops, 1649 internal loops, 230 three-way junction loops and 113 four-way junction loops. We then delete these loops from the molecules and reconstruct them following the procedure described in the Methods section. For each loop, the structure at the center of the largest cluster is taken as the predicted structure.
In Fig 7 we show the RMSDs of the predicted structures with respect to the experimental ones as a function of loop length. Overall, for hairpin loops, internal loops, and three-way junction loops, the RMSDs increase gradually as the loop length increases. The average RMSDs are around 4 Å for hairpin loops of length smaller than 10-nt and increase from 5 to 9 Å for loops of length from 10- to 20-nt. For even longer loops, the average RMSDs are around 11 Å. For internal and three-way junction loops, the performance is better than that for hairpin loops, as may be attributed to the increased number of anchor nucleotides in these cases. For four-way junction loops, the average RMSDs undergo large fluctuations, due to the relatively smaller number of such loops in the dataset. The average RMSDs generally range from 3 Å to 6 Å for loops of length smaller than 20-nt.
In Fig 8 we show two examples for each one of four cases for visual comparison—one example has a medium loop length (9-nt) and one has a long loop length (17-nt). The loops with their RMSDs close to the average (marked by the red dots near the red lines in Fig 7) are selected, so that they are most representative. For the eight cases shown here, six loops are predicted with a reasonable accuracy, with RMSDs around 4 Å. Visual inspection of the figures shows that, the backbones of the predicted structures match those of the experimental ones, while the bases are somehow dynamic (e.g., Fig 8A and 8C).
Two loops are less well predicted. The loop shown in Fig 8B is an essential part of the binding site for elongation factors in rat 28S rRNA. The RMSD between the predicted structure and the experimental one is 6.46 Å. This relatively large value is due to the failure of reconstructing the bulge-G motif, similar to the case discussed in Fig 5F.
The loop of length 17-nt shown in Fig 8H belongs to a four-way junction, locating at the surface of a large ribosomal RNA. The loop is formed by four sub-loops of length 2-, 3-, 3-, and 9-nt, respectively. The RMSD between the predicted structure and the experimental one is 5.17 Å. In the experimental structure, the longest sub-loop forms little interaction with the other three ones and extends out into the solvent. The prediction well reproduced a conformation with this kind of characteristics. The largest deviation between prediction and experiment is in the 2-nt sub-loop G612-A613 and 3-nt sub-loop G647-A648-G649, where A613 and A648 are both in a sharp turn and form a stacking interaction with A668 and A638, respectively.
In this work we develop an approach named RNApps specifically designed for loop structure prediction. The approach includes a probabilistic coarse-grained RNA model, a sequential Monte Carlo growth algorithm, a simulated annealing procedure and an all-atom statistical energy function.
We tested the approach with a set of nine RNA loops, a 23S ribosomal RNA, and a large RNA dataset containing 876 RNAs (RNA 3D Hub, release ID 1.89). For the testing set including nine RNA loops, six loops can be predicted with good accuracy (RMSD < 2.5 Å), one loop has an RMSD of 3.01 Å and two have RMSDs around 6-7 Å. We compared the results with a homology modeling based predictor RLooM and an ab initio predictor iFoldRNA. It was found that RNApps performs comparably with RLooM while considerably better than iFoldRNA. However, we also noted that RLooM cannot guarantee a return of valid structures for some targets, due to the lack of their homology information in the database. In contrast, RNApps and iFoldRNA guarantee a result, and iFoldRNA can also be used for studying the folding dynamics. The tests with a ribosomal RNA showed that the average RMSD is 2.55 Å, 3.23 Å, and 3.97 Å for the rebuilt fragments of length 5-, 8-, and 10-nt, respectively. The tests with RNA 3D Hub showed that the average RMSDs for hairpin loops are around 4 Å and increase slightly with the loop length. The performance for internal loops, three-way and four-way junction loops is even better than that for hairpin loops, mostly due to the increased number of anchor nucleotides. Further analysis showed that although most RNA loops are predicted with good accuracy, some ones with non-canonical base pairs, base triples, or rare torsional angles are reproduced with a lower quality.
We note that our approach in it current form does not perform considerably better than the other predictors and even slightly worse than the homology modeling based one. However, our approach has many unique and promising features. First, it is not only designed for predicting structures of hairpin loops, but also for internal loops, three-way and four-way junction loops and even complex cases. One motivation is that many predictors have been designed to predict the relative position and orientation of the helices of an RNA molecule, however, less attention was paid to the construction of the loops connecting the helices. This makes our work necessary. Second, the efficiency of our approach is high. With the parameters used in this work, a prediction of loop with medium length takes several minutes without SA optimization or several tens of minutes with SA optimization. For example, the computational time for the longest loop (26-nt) in three- and four-way junctions is approximately one hour. This efficiency is much higher than that of most MD based algorithms and the approach is free from the problems associated with homology modeling based methods. The efficiency will be further improved by optimizing the SMC and SA algorithms. Third, the probabilistic and continuous nature of the approach guarantees the sampling of all the relevant phase space in principle, and allows a dynamic adjustment between accuracy and efficiency, which can be determined by users based on their own computational capacity. Fourth, the SMC framework of the approach makes the incorporation of constraints very easy. The constraints may be the experimental information of atomic distance, base pairs or base stacking, or information from users’ experience. With the introducing of such information, the sampling space can be greatly reduced and both accuracy and efficiency will be significantly improved. The way of incorporation of constraints into the SMC framework can be found in previous work [41]. We believe our approach is useful for predicting the tertiary structure of RNA loops.
We also noticed that there are two recent works in protein loop predictions that are similar to ours. In the first one, Tang et al. developed an approach named DiSGro based on sequential Monte Carlo method [51], which is the same as ours. With this approach, they are able to efficiently generate high quality protein loop conformations. The average minimum global backbone RMSD for 1,000 conformations of 12-residue loops is 1.53 Å, with a lowest energy RMSD of 2.99 Å, and an average ensemble RMSD of 5.23 Å. In the second work, the same authors upgraded their approach for the prediction of conformations of multiple interacting loops in proteins [52]. For the most challenging target proteins with four loops, the average RMSD of the lowest energy conformations is 2.3 Å. One novel feature of the approach is the simultaneous construction of multiple loops, making it less likely to over-sample conformations in certain local energy minima. This idea is naturally compatible with the framework of our approach and can be easily incorporated into it. We believe their approaches and ours can borrow ideas from each other and then improve the performance of both.
|
10.1371/journal.pntd.0005504 | Knockdown resistance mutations predict DDT resistance and pyrethroid tolerance in the visceral leishmaniasis vector Phlebotomus argentipes | Indoor residual spraying (IRS) with DDT has been the primary strategy for control of the visceral leishmaniasis (VL) vector Phlebotomus argentipes in India but efficacy may be compromised by resistance. Synthetic pyrethroids are now being introduced for IRS, but with a shared target site, the para voltage-gated sodium channel (VGSC), mutations affecting both insecticide classes could provide cross-resistance and represent a threat to sustainable IRS-based disease control.
A region of the Vgsc gene was sequenced in P. argentipes from the VL hotspot of Bihar, India. Two knockdown resistance (kdr) mutations were detected at codon 1014 (L1014F and L1014S), each common in mosquitoes, but previously unknown in phlebotomines. Both kdr mutations appear largely recessive, but as homozygotes (especially 1014F/F) or as 1014F/S heterozygotes exert a strong effect on DDT resistance, and significantly predict survivorship to class II pyrethroids in short-duration bioassays. The mutations are present at high frequency in wild P. argentipes populations from Bihar, with 1014F significantly more common in higher VL areas.
The Vgsc mutations detected appear to be a primary mechanism underlying DDT resistance in P. argentipes and a contributory factor in reduced pyrethroid susceptibility, suggesting a potential impact if P. argentipes are subjected to suboptimal levels of pyrethroid exposure, or additional resistance mechanisms evolve. The assays to detect kdr frequency changes provide a sensitive, high-throughput monitoring tool to detecting spatial and temporal variation in resistance in P. argentipes.
| Visceral leishmaniasis is a fatal disease transmitted solely by the sandfly Phlebotomus argentipes in India. For decades, indoor residual spraying targeting sandflies with DDT has been the main control tool in the region. Emergence of DDT resistance has compromised this strategy and pyrethroids are now being implemented as alternative. Here, we describe the first molecular markers in sandflies for DDT and pyrethroid resistance. Two knockdown resistance (kdr) mutations in the para voltage-gated sodium channel target site gene are found at high frequency in natural populations in Bihar, the Indian focus of VL. The kdr variants strongly predict DDT resistance in P. argentipes and survival to lower-level pyrethroid exposure. Careful monitoring of resistance, aided by these markers, and of spraying efficacy is required to monitor changes and forewarn of emergent higher-level pyrethroid resistance in the region.
| Visceral leishmaniasis (VL) causes over 20,000 deaths annually [1]. In the Indian subcontinent VL, also known as Kala-azar, is caused by the obligate intracellular protozoan Leishmania donovani that exhibits a continuous cycle comprising of humans, as the only known vertebrate host in the region, and females of the sandfly Phlebotomus argentipes [2]. Currently VL control and elimination strategies involve two main activities: rapid VL detection and treatment for humans and vector control of P. argentipes by indoor residual spraying (IRS) of neurotoxic insecticide [3]. In India, biannual DDT-based IRS has been a common sand fly control strategy in north-eastern regions where VL is most prevalent [4]. However, this control strategy has become compromised by the emergence of DDT resistance in P. argentipes; and quality of spraying is also a concern [4]. Surveillance studies using WHO assays have identified widespread DDT resistance in P. argentipes (mortality rates < 90% with DDT 4%), including in populations from regions with active VL transmission, notably in the state of Bihar [4–7]. Bihar is the main epicentre of VL transmission with 50% of the Indian subcontinent burden [8]. Visceral leishmaniasis has a higher incidence in densely-populated rural areas, particularly along the northern margins of the Ganges River, such as in Vaishali and neighbouring districts [9–12]. The elimination campaign has concentrated IRS in regions with higher VL incidence, spraying two times per year in houses and animal shelters. In 2015, prompted in part by DDT resistance in Vaishali and other northern districts such as Muzaffarpur and Samastipur [6,13], a switch to the pyrethroid alpha-cypermethrin was made for IRS. This scenario of VL incidence and DDT resistance contrasts with that found in Patna district along the southern margin of the Ganges River. Patna exhibits a lower VL incidence, particularly in urban areas of the state capital city [10], with a mosaic of DDT susceptible and resistant P. argentipes populations across villages, and IRS with DDT remains in use for control [13]. Currently there is a lack of knowledge of mechanisms underlying DDT resistance in P. argentipes, and phlebotomines generally, and how these might impact other insecticides, especially pyrethroids.
The major mechanisms of DDT resistance in insects are knockdown resistance mutations (kdr) within the para voltage-gated sodium channel gene in nerve cells (Vgsc) [14,15] and increased metabolism by glutathione S-transferases (GSTs) or, less commonly, cytochrome P450s [16,17]. Both DDT and pyrethroids target the insect VGSC protein. DDT increases sensitivity for depolarization of the channel, while pyrethroids only inhibit inactivation/deactivation processes of the channel; nevertheless each leads to stabilization of the open state and repetitive nerve firing, paralysis and ultimately death [18,19]. Multiple mutations in the Vgsc gene are linked to DDT and pyrethroid resistance in insects, particularly at codon 1014 (using Musca domestica codon numbering) [15]. Wild-type Vgsc-1014 is normally a leucine residue, and is located in the middle of hydrophobic segment S6 of domain II (IIS6) outside of the predicted binding domain [20]. Apparently contrary to this prediction, studies in cockroaches have indicated that kdr mutations in IIS6 may reduce binding affinity to insecticides [21]. Resistance might also arise via alternative mechanisms, such as changes in channel conformation and kinetics, indirectly affecting affinity between insecticides and their binding pockets [20,22]. Expression of insect para-VGSC in Xenopus oocytes has confirmed reduced sensitivity of the VGSC to insecticides for three kdr-1014 mutations (L1014F; L1014H and L1014S) though with substitution-specific impacts on different insecticides [23]. L1014F is the most common kdr mutation in insects, with L1014H also occurring across insect orders, whereas L1014S has only been found in mosquitoes to date [15], as have two other uncommon substitutions (L1014C [24] and L1014W [25]).
In the present study, a region of Vgsc in P. argentipes was sequenced and high-throughput SNP assays designed in order to: i) identify kdr mutations; ii) determine any association with resistance/tolerance to DDT and pyrethroids; and iii) assess their frequencies in natural populations of Bihar contrasting in VL incidence. We report the first resistance markers in a phlebotomine sandfly; with significant predictive utility as diagnostics to support sustainable insecticide use against P. argentipes.
Weekly indoor resting collections of P. argentipes were carried out in central districts of Bihar state (India) between May and October 2015. Manual aspirators and torches were used to collect sand flies from inside human dwellings and cattle sheds during the day. Live sand flies were transported to the Rajendra Memorial Research Institute (RMRI) insectary, where blood fed and gravid P. argentipes females were placed into individual pots and maintained (at 28 ± 2°C; 70 ± 4% RH) until oviposition. Eggs were reared until the adult stage to obtain F1 adult samples. Further field collections were carried out using CDC light traps during the night in villages of two districts, Patna and Vaishali, between September and November 2015 (Table A in S1 Text). Sand flies were identified using morphological keys [26] and P. argentipes specimens stored at room temperature over silica gel. Householders provided informed consent for the collections, permission for which was granted by the ethical review boards of LSTM (protocol references 15.023, 15.036) and RMRI (reference 13/IEC/2015).
Bioassays followed the World Health Organization tube assay procedures [27] using papers impregnated with 4% DDT, 0.05% deltamethrin and 0.05% alpha-cypermethrin. Approximately 25 one day-old sugar-fed adult female P. argentipes were used in all bioassays. All sand flies were fed by cotton wool pad soaked with 10% sucrose. DDT and alpha-cypermethrin bioassays used the F1 progeny of wild caught females; bioassays with deltamethrin were carried out with adult females from a colony maintained at RMRI without previous insecticide exposure, though with occasional supplements from wild populations. No accepted standard exposure time exists for phlebotomines, but the mosquito standard of 60 min has previously proved appropriate for DDT [5] and was used for these bioassays. For the two pyrethroids, a 60 min exposure produced near ubiquitous mortality, therefore in order to produce some survivors for association-testing of mutations, pyrethroid exposure times were reduced to nominal times of 30 min for alpha-cypermethrin and 20 min for deltamethrin. Following exposure, sand flies were transferred to a holding tube and supplied with a cotton wool pad soaked with 10% sucrose. Mortality was recorded 24 hours after the bioassay. Each live female was classed as resistant to DDT or tolerant to pyrethroids (‘tolerant’ terminology is used to indicate survival, but to a shorter than standard exposure), or susceptible if dead after the recovery period. All sand flies were stored individually at room temperature over silica gel.
Sequence information of the voltage-gated sodium channel gene (Vgsc) was obtained from genomes of two phlebotomine species, Lutzomyia longipalpis (VectorBase: Lutzomyia longipalpis Jacobina, LlonJ1.2) and Phlebotomus papatasi (VectorBase: P. papatasi Israel, PpapI1.2). The Vgsc sequences from these two species were used to design the conserved primers Vssc8F (5’–AATGTGGGATTGCATGCTGG–3’) and Vssc1bR (5’–CGTATCATTGTCTGCAGTTGGT–3’), which amplify a genomic DNA fragment from VGSC domain II, segment 6. This fragment included codon 1014 and other codons associated with insecticide resistance in Aedes aegypti (codons 1011 and 1016) and in Lepidoptera and Blattodea (codon 1020).
DNA extraction from individual P. argentipes was performed using the ChargeSwitch Forensic DNA Purification Kit (Thermo Fisher Scientific, MA, USA). Each amplification was performed separately in a 50 μl PCR reaction that contained 10X DreamTaq Green reaction buffer (Thermo Fisher Scientific), 2 mM MgCl2, 0.20 mM of each dNTP, 0.20 μM of each primer and 1U of DreamTaq DNA polymerase (Thermo Fisher Scientific). Thermocycling conditions included an initial denaturation step of 5 min at 95°C followed by 30 cycles each of 96°C for 30 s, 56°C for 30 s, and 72°C for 30 s, and a final extension step of 72°C for 5 min. The PCR products were purified with the QIAquick PCR Purification kit (Qiagen) and sequenced (in forward and reverse directions) using the same primers. Sequences were aligned using Codon Code Aligner version 4.2.7 (CodonCode Corporation, MA, USA).
From the sequence data, two TaqMan SNP genotyping assays (Thermo Fisher Scientific) were designed to identify four alleles detected at codon 1014 (see Results). The novel primer and probe sequences are provided in Table B in S1 Text. TaqMan reactions were performed in 10 μl volumes containing 1X SensiMix (Bioline, UK), 800 nM each primer, and 200 nM each probe on an Mx3005P qPCR thermal cycler (Agilent Technologies, CA, USA) with initial denaturation of 10 min at 95°C followed by 40 cycles of 15 s at 92°C and 1 min at 60°C.
Abbott’s formula [28] was used to correct mortality rates of susceptibility bioassays with DDT, deltamethrin and alpha-cypermethrin. Chi-square tests or Fisher exact tests (where expected frequencies were low) were used to assess differences between resistant/tolerant and susceptible sand flies in allele and genotype frequencies at Vgsc codon 1014, with effect sizes measured using odds ratios. The proportion of each genotype group surviving was determined by dividing the number of live sand flies by the total number of sand flies per genotype. Lower and upper 95% confidence interval limits with correction for continuity were calculated for the survival proportions [29]. Multiple post-hoc pairwise comparisons used the Marascuilo and McSweeney method [30] (http://www.statstodo.com/MultiProp_Pgm.php) to identify significant differences between survival proportions of genotype groups. Sensitivity (ability of the test correctly identify resistant/tolerant individuals) and specificity (ability of the test correctly identify susceptible individuals) were calculated to verify the robustness of Vgsc-1014 assays to infer DDT resistance or pyrethroid tolerance in P. argentipes. Moreover, positive and negative predictive values for resistance/tolerance (probability that resistance/tolerance are presented for each kdr genotype group). Whenever multiple testing was performed, the nominal significance level of α = 0.05 was corrected by the sequential Bonferroni procedure [31].
Sequences were submitted to GenBank (Accession Numbers: KY114615-19).
Bioassay mortalities were 43% with DDT (60 min exposure), 84% with alpha-cypermethrin (30 min exposure), and 56% with deltamethrin (20 min exposure) (Table C in S1 Text). A domain IIS6 fragment of the Vgsc gene was sequenced in a subsample of 48 P. argentipes (24 DDT resistant and 24 DDT susceptible). The wild-type leucine (L) codon (TTA) and three knockdown resistant polymorphisms were detected at codon 1014, comprising of a replacement of leucine with serine, L1014S (TCA) and two alleles (TTC and TTT) which each replace leucine with phenylalanine, L1014F (Fig 1 and Fig A in S2 Text; GenBank Accession Numbers: KY114615-19). Only wild-type sequences were identified at the other three codons within the fragment with previous association with insecticide resistance (i.e. 1011I/I, 1016V/V, and 1020F/F). No further non-synonymous mutations were detected within the IIS6 fragment sequenced. These sequences were used to design a pair of TaqMan SNP Genotyping Assays. The first assay differentiates two bases at the 2nd nucleotide position of codon 1014 (TTA vs. TCA; Fig B in S2 Text), whereas the second assay differentiates the two alleles at the 3rd nucleotide position (TTC vs. TTT; Fig C in S2 Text). Results of the assays are combined to define the 1014 genotype.
Allele frequencies at codon 1014 in females surviving or killed by each insecticide are shown in Table 1. For all insecticides there was no significant difference between survivorship for females possessing the two alternate phenylalanine alleles (TTC vs. TTT; Fisher exact test, minimum P = 0.13) and these were pooled for subsequent analyses. All comparisons between the wt-leucine allele and kdr alleles yielded significant differences associated with lower survival odds for the wt-leucine allele. Survival odds were significantly lower for 1014S than for 1014F for DDT, though not for either pyrethroid (Table 1).
Based on a null hypothesis of recessivity for each kdr allele alone, but some degree of additivity when alternate kdr alleles are present as heterozygotes, the ten possible genotypes at codon 1014 were divided into four groups (Table 2). Low, and homogeneous survival among Leu homozygote and Leu/(Ser or Phe) heterozygote genotypes for DDT and alpha-cypermethrin exposure was concordant with the hypothesis of recessivity, although for deltamethrin Leu/Phe heterozygotes exhibited significantly lower mortality (Table 2). Genotype group frequencies differed strongly between surviving and dead sand flies for each insecticide (DDT: χ23 = 75.2, P = 3 x 10−16; delta: χ23 = 14.9, P = 0.002; alpha: χ23 = 62.6, P = 2 x 10−13; Table 2), and the frequency of leucine-containing genotypes was much higher among dead sand flies (58%-87%) than in survivors (11%-23%). All kdr-only-genotype frequencies, whether homozygote or Ser/Phe heterozygotes, were higher in survivors than dead groups, supporting the hypothesis of additivity of alternate kdr alleles (Table 2).
Survivorship for each genotype group is summarised in Fig D in S2 Text and Fig E in S2 Text. Leucine genotypes exhibited consistently lower survival rates than kdr-only genotypes for DDT and alpha-cypermethrin (Table D in S1 Text, Fig D in S2 Text). For deltamethrin, significant variation in survival was only found between leucine genotypes and phenylalanine homozygotes (Table D in S1 Text), as a result of significantly higher survival rates of Leu/Phe genotypes than other leucine genotypes (Marascuilo post-hoc analysis, P = 0.0014; Fig D in S2 Text). When these Leu/Phe genotypes are removed from the leucine genotype group, a survival pattern more similar to that for DDT and alpha-cypermethrin is observed (Table D in S1 Text, Fig E in S2 Text).
The Vgsc-1014 assay gives generally (though not significantly) higher sensitivity (correct identification as resistant/tolerant) than specificity (correct identification as susceptible) (Table 3). Highest sensitivity is for identification of DDT resistance, and the most consistent values of sensitivity and specificity are for alpha-cypermethrin tolerance. Significantly lower specificity was evident for deltamethrin exposure (Table 3), with the assay suffering from grouping of leucine genotypes differing in resistance association (Table 2; Fig D in S2 Text, Fig E in S2 Text). Predictive values for individual genotype groups were typically high (Fig 2), especially for prediction of susceptibility (negative predictive values). Positive predictive values for DDT resistance exceeded 80% for both phenylalanine genotypes (Ser/Phe, Phe/Phe), while predictive values for pyrethroid tolerance ranged between 53% and 78% for these genotypes. Predictive values for resistance/tolerance are the lowest in serine homozygotes (47%–65%) for all insecticides (Fig 2).
In order to determine whether allele frequencies at codon 1014 varied on scales relevant to local control we genotyped samples from across Patna and Vaishali districts (Fig 3). Variation within districts was relatively limited, with no differences between the two PHCs in Vaishali (χ22 = 0.80, P = 0.67); though a significant difference was present within Patna (χ22 = 10.8, P = 0.005) primarily attributable to heterogeneity in serine allele frequency (Fig 3 and Table E in S1 Text). Intra-district comparisons of genotypic frequencies did not show any significant differences (Patna: χ23 = 5.77, P = 0.12; Vaishali: χ23 = 0.53, P = 0.91; Table F in S1 Text).
Both allelic and genotypic comparisons between Patna and Vaishali districts were highly significant (allelic χ22 = 37.1, P = 9 x 10−9; genotypic: χ23 = 21.4, P = 9 x 10−5). Vaishali exhibited approximately 10% wild type leucine genotypes, with seven of the 14 villages sampled entirely lacking the susceptible allele, and approximately 70% of all samples comprised of phenylalanine genotypes (Table A in S1 Text, Table E in S1 Text, and Table F in S1 Text). In contrast, Patna exhibited a threefold higher frequency of leucine genotypes and a lower proportion of phenylalanine genotypes (44%; Table A in S1 Text and Table F in S1 Text). These frequency data suggest a higher resistance to DDT and higher tolerance to pyrethroids in the more VL-affected Vaishali district than in Patna.
To our knowledge, this is the first detection of resistance mutations in phlebotomine sandflies and adds significantly to the very limited body of knowledge on insecticide resistance mechanisms. Previous work has detected elevated esterase and acetylcholinesterase activity in Sri Lankan P. argentipes [32] and the latter in P. papatasi from Sudan [33], but without clear links to insecticide resistance. Sequencing of a fragment of the para gene in Lutzomyia longipalpis from Brazil detected non-synonymous variants, though the amino acid polymorphisms have not be linked with resistance in any previous studies [34,35]. In contrast, the three polymorphisms we identified in P. argentipes code for two of the most frequently detected kdr mutations, L1014F and L1014S. Whilst L1014F is found across several insect orders, L1014S has only been found in the culicidae (i.e. Anopheles and Culex), in which it is very common in mosquitoes with a TTA leucine codon as the wild-type allele [14,15]. L1014F is linked with lower voltage sensitivity, which requires a stronger depolarization to open the channel, whereas serine 1014 promotes a faster inactivation of the channel without changing the activation sensitivity [23]. Significant associations were found between sandfly survival following exposure to DDT and the two class II pyrethroids tested and both kdr-1014 variants. Moreover, the effect of L1014F on DDT survival was significantly greater than L1014S. For both DDT and alpha-cypermethrin, the effect sizes (measured by odds ratios) were large and at the higher end of the spectrum of estimates from Anopheles [36]. For deltamethrin the quantitative estimates of effect size are somewhat lower but should be interpreted more cautiously owing to the use of laboratory colony sand flies, rather than the F1 near-wild females used for the other insecticide bioassays.
The effects of each mutation seem to be partially or fully recessive. Heterozygotes with a wt-leucine exhibited similar tolerance to insecticide as wt-leucine homozygotes (survival rates << 0.4) with the exception of Leu/Phe heterozygotes for deltamethrin (survival rate ≈ 0.4). This largely recessive nature of the kdr mutations in P. argentipes is consistent with previous knowledge in mosquitoes and other dipterans [37–39]. For DDT we found significant differences between the moderate resistance of serine homozygotes (survival rate ≈ 0.65) and the strong resistance of phenylalanine homozygotes (100% survival rate). Heterozygotes possessing both kdr variants showed an intermediate survival rate suggesting some level of additivity of the alternate kdr alleles and a mechanism to overcome recessivity of the individual mutations. Such a stronger resistant phenotype in 1014S when paired with 1014F has also been reported (for the same two 1014 mutations) in African Anopheles mosquitoes [40]. Differences between the associations of the two kdr mutations were less clear, and not significant, for the two pyrethroids. For alpha-cypermethrin, kdr-only genotypes (excluding phenylalanine homozygotes due to low statistical power) were consistently different from the susceptible wt-leucine genotypes, but serine homozygotes exhibited a survival rate below 50% suggesting a weaker tolerance phenotype than for phenylalanine. For deltamethrin, only phenylalanine homozygotes were significantly different from wt-leucine genotypes, but heterogeneity among wt-leucine genotypes reduced power. Overall, results suggest that phenylalanine is likely to yield a stronger tolerance to class II pyrethroids than serine, which is consistent with expression assays of Drosophila VGSC in Xenopus oocytes [23] and data from Anopheles field populations [40]. Further studies will be necessary to confirm this hypothesis in P. argentipes under test conditions using established diagnostic doses.
Though L1014F seems to confer a stronger resistant/tolerant phenotype, L1014S may offer other advantages in the field by reducing potential fitness costs that are normally associated with L1014F “less excitable channels” [23,41]. Moreover, indirect costs may also emerge as a by-product of strong positive selection from insecticide resistance by reducing linked genetic diversity in neighbouring genomic regions of the para gene, which may interfere with other selection processes at nearby beneficial mutations [42]; for example in L. longipalpis, variation in the para gene is also associated with courtship song variation [34]. Fitness costs could reduce the viability of L1014F homozygotes and potentially increase the benefit of heterozygosity in natural populations, particularly in the absence of insecticide pressure. Nevertheless, our study regions have active control programs and significant deviations from Hardy-Weinberg equilibrium were absent from Vgsc-1014 (Bikram, Patna district; Vaishali district) or where detected, associated with a low proportion of leucine, rather than an excess of 1014F/S heterozygotes (Dhanarua, Patna district; Table G in S1 Text). These results suggest that any potential fitness costs associated with L1014F are currently being overcome by strong selective pressures, a hypothesis consistent with observation of significantly higher 1014F frequencies in Vaishali, which is subjected to greater IRS. Further studies are required to determine whether a reduction in DDT selection pressure may lead to an alteration in the balance of 1014F/S/L in P. argentipes and how kdr mutations in other regions of the sodium channel (e.g. IIS5, IIIS6) may be implicated in DDT/pyrethroid resistance of P. argentipes.
DDT resistance is now well described and established in P. argentipes from north-eastern India. This contrasts with an apparent lack of pyrethroid resistance in the region [4–7,13], though with the important caveat that accepted WHO diagnostic doses for definition of resistance have yet to be ascertained for Phlebotomus spp. Until very recently, DDT has been the main choice of Indian control programs since the 1950s, while a slightly earlier switch to pyrethroids was made by the VL control programs in Bangladesh and Nepal [43,44]. Long-term use of DDT presents a strong selective pressure for resistance and temporal analysis suggests an increase over time [4]. Capacity for populations to evolve resistance is also likely to be enhanced if sub-lethal concentrations are sprayed or maintained in the environment after spraying [4]. Given this long-term use of DDT it is perhaps surprising that kdr mutations provide such a strong prediction of resistance phenotype, because over time additional resistance mechanisms would be expected to develop [45]. Nevertheless, it is plausible that other forms of resistance mechanism may contribute to higher level resistance, a hypothesis of high importance for control that is currently under investigation.
Natural populations of P. argentipes in central Bihar have high frequencies of kdr-1014 mutations, particularly in Vaishali where in several villages wt-leucine alleles were absent from our samples suggesting possible fixation of kdr mutations; now, or in the near future. The high proportion of L1014F is consistent with insecticide bioassays that show high DDT resistance in most villages in Vaishali [13]. Bioassays for P. argentipes are very challenging to execute on a large scale because larvae are extremely difficult to collect and rearing of F1s is time-consuming and technically difficult. Monitoring of kdr-1014 mutations by molecular assays may thus provide a rapid, high throughput tool that does not require live sand flies to infer DDT resistance across P. argentipes populations when compared with bioassays (e.g. WHO cone and tube tests). DNA diagnostics provide greater flexibility in study methodology, allowing screening of stored samples (e.g. permitting a temporal gap between collection and analysis) and simplifying sampling for control programs that may potentiate wider geographic collections. The high proportion of L1014F in Vaishali could also bring challenges, linked to the higher tolerance to pyrethroids associated with kdr-1014 mutations. This may potentiate the emergence of problematic pyrethroid resistance by providing a foundation for resistance upon which other mechanisms can develop [36]. For this reason, it is important to evaluate insecticides with alternative target sites, such as acetylcholinesterase inhibitors (organophosphates and carbamates) that may be integrated into an IRS rotation system.
|
10.1371/journal.pgen.1005071 | Inhibition of Telomere Recombination by Inactivation of KEOPS Subunit Cgi121 Promotes Cell Longevity | DNA double strand break (DSB) is one of the major damages that cause genome instability and cellular aging. The homologous recombination (HR)-mediated repair of DSBs plays an essential role in assurance of genome stability and cell longevity. Telomeres resemble DSBs and are competent for HR. Here we show that in budding yeast Saccharomyces cerevisiae telomere recombination elicits genome instability and accelerates cellular aging. Inactivation of KEOPS subunit Cgi121 specifically inhibits telomere recombination, and significantly extends cell longevity in both telomerase-positive and pre-senescing telomerase-negative cells. Deletion of CGI121 in the short-lived yku80tel mutant restores lifespan to cgi121Δ level, supporting the function of Cgi121 in telomeric single-stranded DNA generation and thus in promotion of telomere recombination. Strikingly, inhibition of telomere recombination is able to further slow down the aging process in long-lived fob1Δ cells, in which rDNA recombination is restrained. Our study indicates that HR activity at telomeres interferes with telomerase to pose a negative impact on cellular longevity.
| Aging is a general biological process among the living organisms which is affected by environmental stimuli but also genetically controlled. Genome instability is one of the aging hallmarks and has long been implicated as one of the main causal factors in aging. DNA double strand breaks (DSBs) are the most deleterious DNA damages that cause genome instability. To counteract DNA damage of DSBs and maintain high level of genome integrity, cells have evolved powerful repair systems such as homologous recombination (HR). HR is crucial for DNA repair and genome integrity maintenance, and is generally believed to be essential for assurance of cell longevity. Telomeres, the physical ends of eukaryotic linear chromosomes, are preferentially elongated by telomerase, a specialized reverse transcriptase, in most cases. However, due to the resemblance of telomeres to DSBs, HR can not be eliminated but rather readily takes place on telomeres, even in the presence of telomerase. Here we show that HR at yeast telomeres elicits genome instability and accelerates cellular aging. Inactivation of the evolutionarily conserved KEOPS complex subunit Cgi121 specifically inhibits telomere HR and results in extremely long lifespan, indicating a dark side of HR in longevity regulation.
| Aging is generally defined as the time-dependent functional decline and increased mortality in most living organisms. Although aging appears to be a natural process, increasing evidence indicates that aging is genetically controlled. In order to elucidate how aging is influenced by intrinsic cellular traits, researchers have developed and employed various model organisms including yeast, worm, fly, fish, mouse and monkey to study the pathways that affect aging. The single-cell organism, budding yeast Saccharomyces cerevisiae represents a widely used tool for aging study [1,2,3]. A single yeast mother cell can only generate a limited number of daughter cells before its mitotic arrest [4]. This aging-associated phenotype is called replicative aging [5]. The organismal aging for multicellular species is likely (or at least partially) to be attributed to cellular aging in their corresponding organs and/or tissues.
The genome, which carries the genetic information of a cell, is continuously threatened by exogenous damages, as well as by endogenous threats such as DNA replication errors [6]. Genome instability is one of the aging hallmarks, and has long been implicated as one of the main causal factors in aging [7,8]. DNA damage (e.g. double strand break, DSB) is one of the major causes for genome instability. When the repair pathways are not efficient enough to cope with a given level of damage, cells may undergo cell cycle arrest, cellular senescence and cell death. For example, the Werner syndrome and Bloom syndrome, two typical progeroid syndromes, are respectively caused by defective helicases WRN and BLM, which are involved in DNA repair [9]. The cells from both syndromes show increased DNA damage accumulation [9]. Consistently, the deficiency in yeast Sgs1 helicase, the homologue of human WRN and BLM, also results in genome instability, such as enhancement of rDNA recombination and fragmentation of nucleolus, and leads to premature cellular aging [10]. To maintain genome stability, genome maintenance pathways have emerged during evolution, and function in longevity assurance. For example, homologous recombination (HR) and non-homologous end joining (NHEJ) pathways have been evolved to repair the most deleterious DNA damages, the DNA double strand breaks (DSBs). Accordingly, mutation of yeast DSB repair genes, such as RAD50, RAD51, RAD57 and RAD52, greatly reduces yeast replicative lifespan [11].
Telomeres are the physical ends of eukaryotic linear chromosomes, and are crucial for genome integrity and stability [12]. Although telomeres may look like DSBs as the chromosomal ends, they are distinguished by the specialized architecture, consisting of repetitive guanine-rich DNA bound by telomere-specific proteins. The yeast telomeric DNA consists of ∼350 bp of TG1–3/C1–3A repeats, and the G strand extends beyond its complementary strand to form a single-stranded overhang, called the G-overhang [12,13,14]. The telomerase complex, which consists of the catalytic subunit Est2, the template RNA moiety Tlc1, and two accessory subunits Est1 and Est3, is responsible for telomeric G-strand elongation, as well as telomere protection [15,16,17,18]. When telomerase is inactivated, telomeres keep shortening and most cells undergo critically short telomere-triggered cell cycle arrest, a process termed “replicative senescence” [16,19,20]. “Replicative senescence” is usually considered to be different from “replicative aging” as the former is largely attributed to critically short telomeres. Although telomeres are well protected and excluded from DSB repair at most of the time, yet there are several traits that make telomeres highly prone to be recombined. Firstly, all the telomeres are much alike in their repetitive sequences which could be favorable substrates for homologous recombination activities. Additionally, quite a few proteins (or protein complexes) involved in DNA repair pathways also bind and function at telomeres [14]. The yKu70/80 heterodimer, which is required for NHEJ [21,22], is indispensable for telomere protection, telomerase recruitment and telomeric heterochromatin maintenance [23,24,25,26,27,28]. Mre11/Rad50/Xrs2 heterotrimer, DNA helicase Sgs1 and endonuclease Sae2, which are critical for resection of the ends of DSBs in HR-mediated DSB repair, are involved in telomere 5’ end resection after DNA replication [29,30,31]. Moreover, the 3’ overhang generated by telomere end resection could be perceived as intermediates of DSB, as the strand invasion step requires the ssDNA [32]. Thus, considering all the traits of telomeres mentioned above, we propose that recombination activity in yeast might covet telomeres and interfere with telomerase to elicit genome instability under physiological conditions, and thereby affect cellular longevity.
The evolutionary conserved KEOPS complex which consists of five subunits, i.e. Cgi121, Bud32, Kae1, Gon7 and Pcc1 in yeast, was first identified as a telomere regulator [33,34]. Deletion of CGI121 or BUD32 reduces single-stranded telomeric DNA accumulated in cdc13-1 cells, and suppresses the temperature sensitivity of cdc13-1 mutant grown at 28°C [33], indicating that loss of Cgi121 or Bud32 limits the amount of ssDNA generated at uncapped telomeres. Moreover, deletion of any subunit of KEOPS complex results in defect in telomere recombination [35], suggesting that KEOPS complex promotes telomeric TG1–3 tracts recombination. In addition to telomere regulation, KEOPS complex also participates in tRNA modification (t6A) [36,37]. Interestingly, the Cgi121 subunit of the KEOPS complex is indispensable for both telomere length regulation and recombination, but not required for tRNA modification [33,37]. We therefore exploited the separation-of-function subunit Cgi121 to dissect the functions of KEOPS in telomere recombination from those in tRNA modification. Our data presented in the current work indicate that activation of telomere recombination accelerates cellular aging, and attenuation of telomere recombination, e.g. by inactivation of Cgi121, promotes cell longevity.
When telomerase is inactivated by deletion of a gene encoding telomerase subunit (Est1, Est2, Est3 or Tlc1), telomeres gradually shorten, and most cells undergo senescence after 75 to 100 generations due to the critical short telomeres [16,19,20]. Because critical short or deprotected telomeres are highly recombinogenic, a small percentage of the telomerase-null cell can overcome the crisis by using homologous recombination to maintain their telomeres. These so-called “survivors” either have telomeres with amplified subtelomeric Y’ sequence and short TG1–3 tracts (the Type I survivors) [38], or harbor long heterogeneous TG1–3 sequence (Type II survivors) [39].
In order to address the effect of telomere recombination on replicative aging, we examined lifespan of the survivor cells. After deletion of telomerase subunit TLC1, cells were serially passaged in solid or liquid medium to obtain Type I and Type II survivors respectively (Fig. 1A, left) [35]. Lifespan assay was performed with both types of survivors. The results showed that both the Type I and Type II survivors have much shorter lifespan than wild-type cells, and the lifespan of Type I survivors is extremely short (Fig. 1A, right). This result is consistent with our previous data that est2Δ Type II survivors have shorter lifespan [40].
The shorter lifespan in telomerase-null survivors suggests that telomere recombination elicits genome instability in telomerase-null survivors to accelerate cellular aging. To test this hypothesis, we re-introduced TLC1 gene back into the survivors that were derived from tlc1Δ cells, by integrating an intact copy of this gene into the genome. After serial passages, telomere structures were examined. The Southern blotting results showed that replenishment of telomerase activity leads to elongation of terminal telomeric TG1–3 tracts in Type I cells and the telomere pattern is stably maintained without further Y’ amplification (Fig. 1B). On the other hand, reduction of the heterogeneity of the long telomeres was observed in Type II cells after replenishment of TLC1, and the telomere structure was gradually restored to a wild-type pattern (Fig. 1B). These results indicate that the recombination activity on telomeres is inhibited by telomerase. Accordingly, reactivation of telomerase partially and completely restores the lifespan of Type I and Type II cells, respectively (Fig. 1C and 1D). These results support the notion that telomere recombination results in genome instability which causes shorter replicative lifespan. Notably, re-introduction of TLC1 only partially restored lifespan of Type I survivors (Fig. 1C). This phenotype is likely attributed to the abnormal karyotypes resulting from telomere end-to-end fusions during Type I survivor generation [38,41]. This explanation is supported by the observation that the severe growth defect of Type I survivors was partially recovered after re-introduction of TLC1.
In telomerase-proficient cells telomere recombination is largely inhibited. However, previous studies have indicated that telomerase seems not to be able to completely eliminate telomere recombination. For example, recombination-mediated telomere rapid deletion (TRD) has been observed in telomerase positive cells [42]. In mouse cells lacking the amino-terminal basic domain of TRF2, t-loop-sized telomeric circles can be excised from leading strand telomeres via homologous recombination [43]. Additionally, homologous recombination occurs with the same frequency in human telomerase-positive and telomerase-negative ALT (alternative lengthening of telomeres) cells [44]. It is conceivable that homologous recombination is intermittently competing with telomerase to contribute to telomere elongation. Due to the repetitive nature of telomeric DNA sequence, telomere recombination products cannot be readily distinguished from those generated by telomerase.
In order to detect telomere recombination event(s) in telomerase-proficient cells, we performed a chromosome healing (de novo telomere formation) assay (see Experimental procedures). The system is modified from that developed by Gottschling’s lab [45]. Briefly, 81 bp of TG1–3 telomeric “seed” (TG81) is inserted into the left arm of chromosome VII at the ADH4 locus, flanked by a TRP1 marker gene and an HO endonuclease cutting site (Fig. 2A). Upon HO endonuclease induction by galactose, the HO site is cut and the TG1–3 sequence is exposed to the very end. The newly formed telomere of 81 bp is critically short and has to be repaired through telomerase or recombination pathways. As a control, an isogenic strain with no TG1–3 seed imbedded (TG0) was included in the experiment. HO-cut will generate a none-telomeric DNA double strand break, which must be repaired to maintain cell viability. In order to examine the repair efficiency of the end generated by HO-cut, we cultured the TG81 or TG0 cells in galactose-containing solid medium, and colonies in which the short telomeres (in TG81 strains) or the DSBs (TG0) have presumably been repaired were counted. The repair efficiency was defined by the number of colonies formed on the galactose plate (cut) divided by that of the same strain on the glucose plate (uncut). The de novo generated short telomere can be efficiently repaired with an efficiency of ∼100% in the TG81 strain (Fig. 2B). Notably, the efficiency was reduced to ∼10% by deletion of TLC1 (Fig. 2B), suggesting a crucial role of telomerase in the elongation of this new short telomere. Deletion of RAD50, RAD51 or RAD52 resulted in no or slight reduction in the repair efficiency (Fig. 2B), supporting a predominant role of telomerase rather than recombination in the repairing process. In contrast, the none-telomeric control TG0 strain has very low repair efficiency (∼0.36%) (Fig. 2B). Deletion of TLC1 in the TG0 strain has little effect on the repair efficiency (Fig. 2B), consistent with the notion that the regular DSB generated by HO-cut can hardly be repaired by telomerase. These data indicate that HO-induced double strand break in TG81 strain can generate a bona fide new telomere, which can be readily elongated by telomerase while recombination could still make a minor contribution to its repair.
Next, we used this system to detect whether telomere recombination takes place in telomerase-positive cells. Although it is technically difficult to distinguish the telomeric tracts added by terminal TG1–3 recombination (Type II recombination) from those by telomerase, the addition of Y’ element to the end of this telomere (Type I recombination) can be detected by Southern blot following PCR amplification with primers specific for TRP1 region and Y’ consensus sequence (Fig. 2A). In this set of experiments, cells were harvested after induction with galactose for 24 h, and genomic DNA was extracted. PCR amplifications were performed using the genomic DNA as templates. The PCR products were subjected to Southern blot with a TG1–3 probe. Meanwhile, proportional amount of genomic DNA was hybridized to a POL1 probe as the internal control. DNA signals in the Southern blot results were quantified and the level of Y’ recombination was normalized to the corresponding internal control. The Y’ recombination efficiency of all the samples were compared with that of TG81 cut sample which was defined as “1”. We successfully detected the telomere recombination events in TG81 but not TG0 strain by Southern blot (Fig. 2C). To validate that the PCR-amplified fragments contained Y’-sequence, we cloned and sequenced some of the PCR products. The representative sequences of three clones are shown (Fig. 2D and S1 Fig.). As expected, the sequences of the PCR products contain part of the TRP1 promoter sequence (in green color), variable lengths (87 to 271 bp) of TG1–3 repeats (purple for TG seed and orange for telomere sequence from the donor chromosome) and the proximal parts of Y’ elements (in gray). Notably, sequences of three clones vary at both the length of internal TG1–3 tracts and the origins of Y’ elements. Two of the three clones captured the Y’ element from the left telomere of chromosome VI, and the third one copied the Y’ element from the right telomere of chromosome VIII. These data indicate that HO-induced short telomeres can be repaired by HR in the presence of telomerase, most likely through break-induced replication (BIR) [46].
We also checked the role of some recombination regulators in telomere recombination in TG81 strain in the presence of telomerase. Interestingly, deletion of RAD50 results in significantly reduced level of such recombination (Fig. 2C). RAD51-null cells have unaffected Y’ recombination, and deletion of RAD52 modestly reduces such recombination (Fig. 2C). Paradoxically, it is generally believed that Rad51 and Rad50 are required for the formation of Type I and Type II survivors respectively, and Rad52 is thought to be essential for virtually all homologous recombination activity [32,47]. It remains elusive why Rad51 is dispensable, or Rad52 plays a minor role for the Y’ telomere recombination in the presence of telomerase. This kind of recombination events may occur in a way similar to that of BIR, as it was reported that a rad51Δ strain still allows BIR to proceed [48], and the Rad51-independent BIR pathway is largely dependent on another set of recombination genes including RAD50 and RAD59 [46]. Nevertheless, these results support the argument that HR takes place in telomerase-positive cells.
Because telomere recombination affects cellular lifespan, we propose that inhibition of telomere recombination will be beneficial to cell longevity. Our previous genetic screenings have shown that the evolutionarily conserved KEOPS complex was required for telomeric TG1–3 tracts recombination [35]. Additionally KEOPS complex is likely to be involved in generation of telomeric ssDNA because deletion of either CGI121 or BUD32 reduces telomeric ssDNA level in cdc13-1 mutant and suppresses the temperature sensitivity [33]. We therefore wanted to establish the functional relevance of KEOPS complex between telomere recombination and cellular longevity regulation. We focused our efforts on Cgi121 due to several concerns. (1) Deletion of any of the other four subunits confers severe growth defect [33,34,35,49,50], while deletion of CGI121 only has minor effect on cell growth. (2) Structural studies indicate that lack of Cgi121 doesn’t affect the interactions between other subunits [49]. (3) More importantly, Cgi121 is not required for tRNA modification, but indispensable for both telomere length regulation and recombination [33,35,37], providing us a separation-of-function tool to conduct genetic analyses.
To validate that Cgi121 promotes telomere recombination, we firstly examined telomere recombination efficiency of cgi121Δ mutant in the presence of telomerase as in Fig. 2C. The result showed that the Y’ recombination efficiency was modestly reduced by deletion of CGI121 in telomerase-positive cells (Fig. 3A). Then we performed telomere sequencing to examine the role of Cgi121 in telomeric TG1–3 recombination in telomerase-negative tlc1Δ and tlc1Δ cgi121Δ cells. We constructed heterozygous diploid cells with one copy of telomerase RNA gene TLC1 and CGI121 deleted (BY4743 TLC1/ tlc1Δ CGI121/cgi121Δ). After sporulation and tetrad dissection, spores with different genotypes (tlc1Δ and tlc1Δ cgi121Δ) were identified. Spores from the same tetrad were used for further analysis as they had the same initial telomere lengths. Cells were cultured for 50 generations after tetrad dissection and genomic DNA was extracted. Telomere PCR was then performed as previously described to amplify the TG1–3 sequence of telomere IL [51,52]. The PCR products were then cloned to T vector for sequencing analyses. About 100 clones of each genotype were obtained and analyzed. The results showed that 20.59% of telomeres (21 out of 102) were elongated through recombination in tlc1Δ cells (Fig. 3B). In contrast, only 11.54% of telomeres (12 out of 104) were repaired through recombination in tlc1Δ cgi121Δ cells (Fig. 3B). These results confirmed that Cgi121 plays a positive regulatory role in telomere recombination in both telomere-positive and -negative cells [35].
Since Cgi121 promotes telomere recombination (Fig. 3A and 3B), and inhibition of telomere recombination restores cellular lifespan (Fig. 1C and 1D), we reasoned that deletion of CGI121 would suppress the recombination activity at telomeres, and thereby extend the replicative lifespan. Following this thought, we deleted CGI121 in a long lived yeast strain BY4742 (Mat α) which is commonly used in aging research and examined lifespan of the mutant. Deletion of CGI121 slowed down aging process strikingly, both the mean and maximum lifespan of cgi121Δ cells increased about 50% (Fig. 3C). The long live phenotype was also observed in the isogenic BY4741 cgi121Δ strain of Mat a mating type (S2A Fig.), indicating that Cgi121 affects lifespan independently of the mating type. Thus, we identified Cgi121 as a novel longevity regulator.
As we have mentioned above, in wild-type cells telomeres are maintained mainly by telomerase while telomere recombination occasionally occurs and brings the risk of genome instability. Deletion of CGI121 in these wild-type cells may inhibit telomere recombination and promote genome stability and cell longevity. If this is the case in the presence of telomerase, the activated telomere recombination in the absence of telomerase could also be compromised by deletion of CGI121, and extension of lifespan in telomerase–null cells would be expected.
To examine the role of Cgi121 on lifespan of telomerase-null pre-senescing cells, we obtained spores from heterozygous diploid cells (BY4743 TLC1/tlc1Δ CGI121/cgi121Δ) by tetrad dissection. Spores with Mat α (the same mating type as that of BY4742) were selected to perform the following lifespan assays. According to previously published data in our lab, cells immediately dissected have lifespan similar to that of wild-type cells as telomere recombination is not activated yet [40]. Thus, in our experiment, spores were grown for 50 generations after dissection so that telomeres are modestly shortened and telomere recombination level is elevated (Fig. 3D, left). These cells were then subjected to lifespan assay. As expected, tlc1Δ senescing cells show shortened lifespan, and deletion of CGI121 extends lifespan of tlc1Δ mutant (Fig. 3D, right). Consistently, this phenotype is also observed in est2Δ mutant (S2B Fig.). We noticed that lifespan of tlc1Δ cgi121Δ double mutant was not restored to the level seen in the cgi121Δ single mutant (Fig. 3D, right). That’s likely attributed to the continuous telomere shortening in the double mutant as the lifespan assay progressing. The emerging critically short telomeres can trigger cell cycle arrest and senescence. Therefore, the double mutant doesn’t have a full replicative capacity as the cgi121Δ single mutant.
To avoid the interference of critically short telomere(s) on lifespan, we generated heterozygous diploid cells with over-elongated telomeres by introducing a plasmid harboring a Cdc13-Est2 fusion protein [53]. Cells were serially passaged and telomeres were examined by telomeric Southern blot. Expression of the fusion gene conferred super-long telomeres of about 1 kb to the diploid cells (S2C Fig.). Then the plasmid encoding Cdc13-Est2 fusion protein was popped-out by negative selection and tetrad dissection was performed to obtain spores with different genotypes. The super-long telomeres in the dissected spores are about 800 bp (Fig. 3E, left), a length that prevents critically short telomeres from emerging during the lifespan assay. Over-elongating telomere has no effect on lifespan of wild-type cells (S2D Fig.). The tlc1Δ mutant with long telomeres shows similar lifespan to that of wild-type cells (Fig. 3E, right), probably because the long telomeres in this mutant result in similar telomere recombination state to that of wild-type strain. Deletion of CGI121 extends lifespan of tlc1Δ mutant significantly and the double mutant has lifespan similar to that of cgi121Δ single mutant (Fig. 3E, right). These data further support our hypothesis that inhibition of telomere recombination by deletion of CGI121 promotes cellular longevity.
It remains elusive how Cgi121 promotes telomere recombination to affect cell longevity. One possibility is through regulating generation of telomeric ssDNA which is essential for initiation of recombination events. Previous report suggests that Cgi121 functions in generation of telomeric ssDNA, as deletion of CGI121 inhibits accumulation of telomeric ssDNA in the temperature sensitive cdc13-1 mutant [33]. In wild-type cells, the level of telomeric ssDNA is relatively low. In yku80Δ cells, telomeres become deprotected and telomeric ssDNA is accumulated [23,25], and accordingly the replicative lifespan is shortened (S3 Fig.) [54]. However, the shortened lifespan of yku80Δ mutant is not restored to the length of wild-type cells by deletion of CGI121 (S3 Fig.). Considering that yKu80 plays multiple roles in DNA damage repair, as well as telomere maintenance, we then used a separate-of-function yku80tel allele, yku80-4, which retains the ability of DNA end-joining and telomerase activity regulation, but displays severe defects in telomere protection [55]. The yku80-4 mutant was constructed by integrating a plasmid bearing a yku80-4 allele into the genome of yku80Δ strain. In parallel, the vector plasmid or the plasmid harboring a wild-type copy of YKU80 was integrated to yku80Δ mutant respectively. As the yku80Δ null mutant, the yku80-4 mutant also shows significantly shortened lifespan (Fig. 4A), probably due to accumulated ssDNA [55]. Strikingly, the lifespan of yku80-4 cgi121Δ double mutant is fully restored and extended to a level equivalent to that of the cgi121Δ single mutant (Fig. 4B). These data support the conclusion that Cgi121 may facilitate ssDNA generation at telomeres [33], and therefore accelerate cellular aging.
In budding yeast, the rDNA consists of ∼150 copies of 9.1 kb rRNA genes, and is highly recombinogenic [56,57,58]. rDNA instability is promoted by Fob1-dependent DNA replication fork stalling which may cause DSBs within the rDNA [59,60,61]. Elimination of FOB1 gene reduces the rate of rDNA recombination [61], and significantly extends cellular lifespan [62]. In contrast, deletion of SIR2 gene disrupts the heterochromatin structure of rDNA loci and rDNA recombination level is elevated which confers shortened lifespan [54,63,64,65]. To investigate whether the effect of Cgi121 on cell longevity is attributed to rDNA recombination, we performed a marker loss assay to analyze the rDNA recombination rate [40]. As controls, sir2Δ cells show significantly elevated rDNA recombination level while fob1Δ cells have very low level of rDNA recombination (Fig. 5A). The rDNA recombination rate in cgi121Δ mutant was comparable to that in wild-type cells (Fig. 5A), indicating that Cgi121 is not involved in rDNA recombination. Consistently, the long lifespan of cgi121Δ cells could be further extended by deleting FOB1 (Fig. 5B). These results demonstrate that rDNA recombination and telomere recombination affect cellular lifespan in different pathways, and inhibition of both recombination activities have additive effect on cell longevity. Interestingly, deletion of Cgi121 affects neither homologous recombination activity at other genomic loci (S4A Fig.) nor the NHEJ efficiency (S4B Fig.), and the gross chromosomal rearrangement (GCR) rate is modestly elevated in cgi121Δ mutant (S4C Fig.). Thus, we conclude that the effect of Cgi121 in longevity regulation is attributed specifically to its role in telomere recombination.
Calorie restriction (CR) slows aging and increases life span in many organisms [66,67]. The life span extension by CR in yeast is mediated by the coordinated activity of three nutrient-responsive kinases: TOR (target of rapamycin), Sch9, and protein kinase A (PKA) [68,69,70,71]. To better understand the longevity regulation by Cgi121, we examined the lifespan of cgi121Δ cells under CR condition. CR treatment was achieved by reducing the glucose concentration in the growth medium from 2% to 0.05% [72]. In this assay, CR treatment extends lifespan of wild-type cells while the long lifespan of cgi121Δ cells could not be maintained under CR condition (Figs. 6A and S5A). Consistently, deletion of TOR1 which genetically mimics CR [69] shortens the mean and maximum life span of cgi121Δ cells (Fig. 6B). The phenotype of lifespan shortening by CR in cgi121Δ mutant is quite unexpected, but similar to those observed in W303AR cells, which has been commonly used in yeast aging research [73], as well as in long-lived Osh6 overexpression cells [74].
The observation that the longevity of cgi121Δ cells requires TOR activity leads us to speculate that Tor1 might be involved in telomere recombination. To test this possibility, we constructed heterozygous diploid cells with one copy of TOR1 and TLC1 deleted, and then obtained spores with different genotypes by tetrad dissection. These spores were cultured and serially passaged in solid or liquid medium to see whether deletion of TOR1 affects either type of telomere recombination. When cultured on solid medium, Type I survivors were readily obtained in tor1Δ tlc1Δ cells. 9 out of the 10 randomly selected clones were Type I survivors and the other clone was Type II (S5B Fig.). The emerging frequency of Type I survivors (90%) is highly consistent with previous reports [38,39], suggesting that Y’ recombination is unaffected by deletion of TOR1. On the other hand, when cultured in liquid medium, tlc1Δ cells showed a typical senescence phenotype, and deletion of TOR1 had no effect on senescence rate according to the growth curve (S5C Fig.). Southern blot analysis revealed that Type II survivors arised in both clones of either tlc1Δ or tor1Δ tlc1Δ cultures (S5D Fig.), arguing that TOR1 does not affect TG recombination. Taken together, these data suggest that Tor1 doesn’t affect telomere recombination.
Homologous recombination (HR) is generally a universal biological process across the living organisms. It not only serves to eliminate deleterious chromosome lesions (such as DSBs and interstrand crosslinks), but also is critical for the stabilization of stalled replication forks and chromosome segregation in meiosis. Therefore HR is indispensable for general maintenance of genome integrity and stability. Deletion of the genes in RAD52 epistasis group inhibits telomere recombination, but also results in inability of repairing deleterious lesions inevitably occurring at other chromosomal loci than telomeres. Thus the overall impact of inactivation of RAD52 epistasis genes on lifespan is negative, leading to genome instability and lifespan shortening [11]. In this work, we surprisingly found that HR activities at telomeres can elicit genome instability and pose a negative effect on cellular longevity. Deletion of the KEOPS subunit gene CGI121 specifically inhibits telomere recombination and significantly slows down replicative aging (Fig. 3). Cgi121 appears to be only required for HR at telomeres (Fig. 3A and 3B), but not for regular DSB repair by HR at other genomic loci including the rDNA (Figs. 5A and S4A), nor for other DNA repair pathways like NHEJ or GCR (S4B and S4C Fig.). Such a separation-of-function property of Cgi121 provides us a specific tool to assess the effect of telomere recombination in cellular longevity.
Although telomerase-mediated telomere replication is the major pathway that elongates telomeres in most eukaryotes, there are eukaryotes that do not have telomerase, but solely use recombination to maintain telomeres. From the evolutional point of view, HR may represent the earliest telomere maintenance mechanism, which precedes the evolution of telomerase-dependent maintenance of chromosomal termini [75]. In budding yeast, both telomerase and recombination can efficiently function to replicate telomeres, though the former is more preferred. In telomerase-proficient cells, the engagement of recombination in telomere elongation can have both beneficial and detrimental effects (Fig. 7A). On one hand, the recombination activity seems to be complementary to telomerase activity in telomere elongation (Fig. 2). This idea is also supported by the study in the Kluyveromyces lactis stn1-M1 mutant, in which recombination takes a dominant role in telomere elongation in spite of proficient telomerase activity [76]. Thus it is not surprising, but rather logical, to see that HR is able to function as a back-up system for telomere replication when telomerase pathway fails. On the other hand however, telomeres possess multiple characteristics of DSBs and recruit HR activity to undergo “repair” process. The HR-mediated repair activity tangles or competes with telomerase activity on telomeres (Fig. 2), leading to false-alarms which make the cells undergo aging. Thus, the balance of both activities results in metastable telomeres and normal lifespan (Fig. 7A).
In cells with no or less HR activities at telomeres (Fig. 7B), telomerase is not (or less) “harassed” by the HR activity, and the more stable telomeres result in longer lifespan (Fig. 7B). The KEOPS complex subunit Cgi121 is required for efficient telomere recombination (Fig. 3A), probably by functioning in telomeric ssDNA generation (Fig. 4B) [33]. Deletion of CGI121 specifically inhibits telomere recombination and confers extended lifespan (Fig. 3).
In telomerase-negative yeast cells, HR becomes the only means to maintain telomere length, and the unstable telomeres elicit genome instability. The trade-off of ensuring the viability of a population of cells through telomere recombination is to partially sacrifice the longevity of the mother cells (Fig. 7C). This notion is supported by several lines of evidence. The yeast cells that solely relied on HR (telomerase-null survivors) exhibited shorter lifespan than those that use telomerase to replicate telomeres (wild-type cells) (Fig. 1A) [40]. Additionally, replenishment of telomerase in telomerase-null yeast cells efficiently inhibits telomere recombination (Fig. 1B), and at least partially restores lifespan (Fig. 1C and 1D). Furthermore, attenuation of telomere recombination by deletion of CGI121 significantly increases lifespan in telomerase-negative cells (Figs. 3D, 3E and S2B). Therefore, we prefer the model that the illegitimate recombination activity involuntarily competes and interferes with telomerase activity to cause genome instability at telomeres, and results in acceleration of replicative aging (Fig. 7). It is possible that both genome stability and cell longevity are driving forces for the evolution of HR-to-telomerase in telomere maintenance mechanism.
HR is prone to take place at the genomic loci that share similar or the same DNA sequences. In yeast genome, both telomere regions and rDNA loci have repetitive sequence, and they are hot spots for intra and/or inter-chromosomal HR. At chromosomal ends, telomere recombination occurs in the presence of telomerase (Fig. 2). Deletion of CGI121 specifically inhibits telomere recombination and extends cell longevity both in telomerase-positive and -negative cells (Fig. 3). At rDNA loci, the Fob1-dependent replication fork stall causes replication stress and rDNA instability, and triggers recombination-mediated pop-out of rDNA circles [59,60,61]. Deletion of FOB1 reduces rDNA recombination [61], and extends cellular lifespan [62]. In spite of the differences between telomere and rDNA recombination, inhibition of recombination at both sites seen in the fob1Δ cgi121Δ double mutant cells has additive effect on slowing down of aging (Fig. 5B). This evidence further supports our model that the unregulated and/or illegitimate HR events occurring at certain genomic loci such as telomeres and rDNA elicit genome instability, and thereby pose a negative impact on cell longevity. It might be therapeutically significant to find a means to promote cell longevity by blocking recombination-mediated repair of telomeres.
In addition to Cgi121, the KEOPS complex contains four other subunits including a protein kinase (Bud32) and an ATPase (Kae1) [33,34]. It remains elusive whether the other subunits also regulate longevity. Due to the severe growth defect of the other four KEOPS mutants, lifespan assay could not be performed. When a second copy of the KEOPS genes respectively integrated into the genome, the mRNA levels of these genes were elevated respectively according to the qRT-PCR results, suggesting that the five subunits were overexpressed respectively (S6A Fig.). However, the lifespan of the cells overexpressing any of the five subunits was unchanged (S6B and S6C Fig.). We speculate that the subunits of KEOPS complex do not function individually, but rather as a whole complex to regulate aging. Considering the high conservation of KEOPS complex in evolution, it is also intriguing to investigate whether the counterpart of the KEOPS in higher organisms functions in the same way as in yeast in longevity regulation.
All the yeast strains used in this study were listed in S1 Table. Strains used in lifespan assay were all in BY4742 background unless stated otherwise. The de novo telomere addition system was modified from that reported by Gottschling’s lab [45]. The yku80-4 and related strains were constructed by integrating an MscI-linearized plasmid pRS303 bearing a copy of yku80-4 or YKU80 gene, or simply the vector plasmid into the his3Δ1 locus in the genome of yku80Δ mutant. Strains overexpressing KEOPS subunits were constructed by integrating an MscI-linearized plasmid pRS303 bearing sequences including ORF, endogenous promoter and terminator of the target genes into the his3Δ1 locus in the genome. Yeast strains were constructed either by transformation with a lithium acetate procedure or genetic cross (mating and tetrad dissection). The plasmids for gene deletion were constructed based on the pRS series [77].
Lifespan assay was performed as described previously [40]. Yeast strains were pre-grown overnight on solid YPD plates at 30°C. Cells were then streaked onto fresh YPD plates and grew for about 2 hours. Single cells were randomly selected and arrayed to the plates using a micromanipulator (Singer MSM). After 2 hours (about 1–2 divisions), virgin daughter cells were isolated as buds from mother cells and subjected to lifespan analysis. Daughter cells were then removed by gentle agitation with a dissecting needle and tabulated every 1–2 cell divisions until all the cells stopped dividing. Each experiment was performed with 50–60 cells for each strain. Statistical significance was determined by a Wilcoxon rank sum test using Stata 8 software and significant differences were stated for p < 0.05. The statistical data of the replicative lifespan experiments in this study were shown in S2 Table.
Genomic DNA was extracted and digested with XhoI and then subjected to telomere Southern blot as described previously using the TG1–3 probe [52].
Yeast cells were inoculated in yeast complete medium lacking both uracil and lysine (YCU-K-), containing 2.5% of raffinose (Sigma). The TG81 rad52Δ or TG81 rad50/rad51/rad52Δ cells were inoculated in YC medium lacking lysine (YCK-). Proportional cells were plated onto YC medium lacking uracil, containing glucose (2%) or galactose (3%) (Sigma). The repair efficiency was defined as the number of colonies on galactose plates (cut) divided by that on glucose plates (uncut). The data were summarized from four independent experimental duplicates and the error bars indicates the standard deviations. Statistic significances were calculated by Student t-test (*p < 0.05 and **p < 0.01).
Yeast cells were inoculated in YCU-K- medium plus 2.5% of raffinose, then diluted into YCU- plus 2.5% of raffinose and cultured to logarithmic phase. For TG81 rad52Δ or TG81 rad50/rad51/rad52Δ strain, cells were inoculated in YCK- medium and diluted into complete YC medium. Galactose was then added to a final concentration of 3%, and cells were cultured for an additional 24 h. Cells were harvested and genomic DNA was extracted following PCR amplification with primers specific for TRP1 promoter and consensus sequence of all the Y’ elements. The PCR products were subjected to Southern blot using a probe specific for TG1–3 repeats as used in other telomeric Southern blot in this study. Proportional genomic DNA was digested with EcoRI endonuclease to generate a POL1 DNA fragment of about 1 kb which was detected by a probe specific for POL1 gene and serves as internal control. Meanwhile, aliquots of the PCR products were cloned to pMD18-T vector (TAKARA) and subjected to sequencing.
Spores derived from diploid strain BY4743 TLC1/tlc1Δ CGI121/cgi121Δ were cultured for about 50 generations after dissection. Genomic DNA was extracted and subjected to telomere PCR as described previously [51,52]. PCR products were cloned to pMD18-T vector (TAKARA) and then subjected to sequencing.
rDNA recombination rate is assessed by the rate of loss of the URA3 reporter gene inserted at rDNA loci [40]. Cells grown in log phase were plated to YC medium with or without 0.15% 5-FOA. rDNA recombination rate is determined by dividing the number of colonies grown on 5-FOA-containing YC plates by that on YC plate without 5-FOA. The error bars indicates the standard deviations of data from three independent experiments. Significant differences were stated for *p < 0.05 and **p < 0.01.
|
10.1371/journal.pgen.1002742 | Genome-Wide Analysis of GLD-1–Mediated mRNA Regulation Suggests a Role in mRNA Storage | Translational repression is often accompanied by mRNA degradation. In contrast, many mRNAs in germ cells and neurons are “stored" in the cytoplasm in a repressed but stable form. Unlike repression, the stabilization of these mRNAs is surprisingly little understood. A key player in Caenorhabditis elegans germ cell development is the STAR domain protein GLD-1. By genome-wide analysis of mRNA regulation in the germ line, we observed that GLD-1 has a widespread role in repressing translation but, importantly, also in stabilizing a sub-population of its mRNA targets. Additionally, these mRNAs appear to be stabilized by the DDX6-like RNA helicase CGH-1, which is a conserved component of germ granules and processing bodies. Because many GLD-1 and CGH-1 stabilized mRNAs encode factors important for the oocyte-to-embryo transition (OET), our findings suggest that the regulation by GLD-1 and CGH-1 serves two purposes. Firstly, GLD-1–dependent repression prevents precocious translation of OET–promoting mRNAs. Secondly, GLD-1– and CGH-1–dependent stabilization ensures that these mRNAs are sufficiently abundant for robust translation when activated during OET. In the absence of this protective mechanism, the accumulation of OET–promoting mRNAs, and consequently the oocyte-to-embryo transition, might be compromised.
| One of the most striking developmental events is the oocyte-to-embryo transition that, in the absence of Pol II–dependent transcription, depends on regulated translation of maternal mRNAs. Prior to their activation, these maternal mRNAs need to be “stored" in the egg cytoplasm in a repressed but stable form. Surprisingly little is known about how the stored mRNAs are stabilized. The STAR family of RNA–binding proteins includes the C. elegans GLD-1, which controls many aspects of germ cell development. To obtain a comprehensive picture of GLD-1–dependent mRNA regulation, we performed a genome-wide survey of translational repression and mRNA stability of GLD-1 targets. This uncovered a potential role of GLD-1 in mRNA storage, as we found that GLD-1 both represses and stabilizes a subpopulation of its targets. The stabilization also involves a DDX6-like RNA helicase, CGH-1, which is a component of repressive germ granules and processing bodies. Remarkably, the GLD-1 and CGH-1 stabilized mRNAs encode regulators of the oocyte-to-embryo transition, providing an insight into how these functionally related mRNAs are specifically stabilized during germ cell formation. These findings have potential implications for oocyte quality and reproductive fitness, and for mRNA storage in other cell types such as neurons.
| The oocyte-to-embryo transition (OET), which encompasses oocyte maturation, ovulation, fertilization, and early embryogenesis, occurs while Pol II dependent transcription is globally repressed. This is why OET is largely driven by maternal mRNAs that are stored in the egg cytoplasm in a repressed but, importantly, also stable form [1]. In contrast to translational repression, stabilization of repressed mRNAs remains little understood. In Xenopus oocytes, mRNA stability is attributed to a global inhibition of decapping activity [2], [3]. On the other hand, in developing Drosophila oocytes, stabilization of the bicoid mRNA depends on the binding of a specific protein, BSF [4]. This suggests that, at least in some species, global inhibition of mRNA decay is not a general feature of oogenesis and thus mechanisms stabilizing specific germline messages might exist.
In C. elegans, the DDX6-like RNA helicase, CGH-1, associates with a large number of germline mRNAs [5]. Some of these mRNAs are less abundant in CGH-1 (-) germ cells, suggesting that this helicase plays a role in mRNA stabilization [5]. The DDX6-like helicases are present in various cytoplasmic ribonucleoprotein (RNP) particles such as processing (P) bodies [5]–[14]. In the C. elegans germ line, CGH-1 localizes to P granules, which are associated with the nuclear envelope, and to P body-like cytoplasmic granules [15]. In contrast to P bodies, the latter granules seem to be largely devoid of RNA decay enzymes and have thus been proposed to serve as vehicles of mRNA storage, which is consistent with a role of CGH-1 in mRNA stabilization, [5], [16]–[22]. However, because somatic P body formation is thought to be the consequence, not the cause, of mRNA repression [23], [24], the functional significance of these RNA granules for mRNA stabilization remains to be demonstrated.
Here, we report the C. elegans STAR-protein GLD-1 as a potential player in maternal mRNA storage. GLD-1 is expressed in the medial gonad (Figure 1A), where it promotes meiosis, oogenesis, and maintenance of germ cell identity by repressing the translation of diverse mRNAs [25]–[30]. Recently, we have shown that GLD-1 associates with hundreds of germline transcripts, and that this association is determined by the number and strength of 7-mer GLD-1 binding motifs (GBMs) within untranslated regions (UTRs) [31]. To understand how GLD-1 regulates its mRNA targets, we undertook a functional genomics approach. By transcriptome-wide polysome profiling, we found that GLD-1 has a widespread role in repressing translation. Our results also suggest that GLD-1 stabilizes many targets, which additionally involves the DDX6-like RNA helicase CGH-1. Because the stabilized mRNAs encode proteins critical for OET, and their stability appears to be important for efficient accumulation in oocytes, GLD-1 dependent mRNA storage might be important for a successful oocyte-to-embryo transition.
It is currently unknown how GLD-1 represses translation. By ‘polysome profiling’, in which poly-ribosomes (polysomes) are separated from single ribosomes and ribosomal subunits by sucrose density gradient ultracentrifugation, two of the GLD-1 targets, tra-2 and pal-1, have been suggested to be repressed at the initiation or elongation stage of translation, respectively [30], [32]. To globally examine the effect of GLD-1 on the translation of its targets, we performed polysome profiling on a transcriptome-wide scale. In general, while polysomal fractions contain translated mRNAs (as well as mRNAs repressed at the elongation or termination stage of translation), sub-polysomal fractions contain poorly translated mRNAs and transcripts repressed at the initiation stage of translation (Figure S1A–S1C). To determine the distribution of GLD-1 between fractions, we used monoclonal antibodies raised against GLD-1, and, as a control, against the translational activator polyA-binding protein, PAB-1. As expected, PAB-1 was enriched in polysomal fractions (Figure 1B). In contrast, the majority of GLD-1 was present in the sub-polysomal fractions (Figure 1B). By comparing polysomal and total mRNA levels by microarray analysis, we found that also most GLD-1 targets (Table S1 and Figure 1C and Figure S1D; mRNAs more than 3-fold enriched in GLD-1 IPs; also [31]) were enriched in sub-polysomal fractions (Figure 1C; GLD-1 targets are in red; transcripts more than 2-fold depleted from polysomes, including 64% of GLD-1 targets, are below the black line). To examine GLD-1 dependent repression, we tested whether GLD-1 targets shift to polysomal fractions in gld-1(q485) null mutant worms (hereafter called gld-1 mutants). Because gld-1 mutants develop germline tumors, we only examined young adults, in which the gonads contained large numbers of pachytene cells and only few ectopically proliferating cells [26], [33]. To collect sufficient quantities of mutant animals, gld-1 homozygous mutants were separated from heterozygous animals, carrying a GFP-tagged balancer, by fluorescence-activated sorting. Expectedly, we found that the loss of GLD-1 had little effect on the polysomal/total mRNA ratio of non-GLD-1 targets and of targets of an unrelated RBP, FBF [34] (Figure 1D and Figure S1E). In contrast, the loss of GLD-1 caused GLD-1 targets to shift to polysomes (Figure 1D; p<7.3e−16; p values were calculated with a t test). While the polysomal/total mRNA ratio of GLD-1 targets remained relatively low in gld-1 mutants compared to non-targets (possibly due to residual repression by other RBPs as has been observed for several GLD-1 targets [30], [35], [36]), these results collectively suggest that, although additional mechanisms may exist and contribute to repression, GLD-1 binding inhibits translational initiation.
Several GLD-1 targets have been observed by others to be less abundant in gld-1 mutants [26]–[29], [37], [38]. To globally determine a potential function of GLD-1 in mRNA stabilization, we analyzed the abundance of mRNAs in wild-type and gld-1 mutant gonads by microarray analysis. We then compared changes in the mRNA abundance in gld-1 mutants to GLD-1 binding (Table S1 and Figure 2A; the vertical dotted line separates non-targets on the left from presumed GLD-1 targets on the right and the horizontal lines demarcate a 2-fold change of mRNA abundance). We observed that a subset of GLD-1 targets (14%) were less abundant in gld-1 mutants (Figure 2A, transcripts marked in red, those encircled in blue were confirmed by RT-qPCR in 2B). Because these mRNAs also tend to shift to polysomes in gld-1 mutants (Figure S2), our results suggest that GLD-1 may control both their repression and stability.
To investigate potential partners of GLD-1 in mRNA stabilization, we immunopurified GLD-1 and analyzed co-purifed proteins by mass spectrometry. Top proteins most enriched in GLD-1 immunoprecipitates (IPs), together with their counterparts in other animals, are listed in Figure 3A (also see Figure S3A). These include the DDX6-like RNA helicase CGH-1, the Y-box proteins CEY-1-4, the Sm-like domain protein CAR-1, and the cytoplasmic polyA binding protein PAB-1, all of which are conserved components of repressive RNPs in germ cells and somatic cells, and which have been previously shown to interact with each other [5], [8]–[14], [22]. Using available antibodies, we confirmed by western blot analysis that the interactions between GLD-1 and CGH-1, CAR-1, and PAB-1, were specific (Figure 3B). CGH-1 has previously been implicated in the stabilization of at least some maternal mRNA [5], which is why we pursued its interaction with GLD-1 further. We observed that the GLD-1/CGH-1 interaction was dependent on RNA (Figure 3C) and confocal microscopy revealed that only a minor fraction of GLD-1 co-localized with CGH-1 in the germline cytoplasm (Figure S3B).
Despite the indirect (RNA-mediated) interaction between GLD-1 and CGH-1, we tested if the two proteins may be functionally related. Because we were unable to create a strain containing both gld-1(q485) and cgh-1(ok492) null mutations, the temperature-sensitive cgh-1(tn691) allele, hereafter called cgh-1ts, was used in many experiments. This is an antimorphic allele (Ikuko Yamamoto and David Greenstein, personal communication; Figure S3E), which nevertheless induces oocyte defects and sheet-like CAR-1 containing structures also observed in cgh-1 null or cgh-1(RNAi) gonads (data not shown; [5], [39]). We initially confirmed that the loss of CGH-1 activity had no obvious effect on the levels and distribution of GLD-1 ([39] and Figures S3C and S4B), nor did the loss of GLD-1 affect CGH-1 ([15] and Figure S3D). By microarrays, we examined the abundance of mRNAs in gonads dissected from cgh-1ts mutants grown at the restrictive temperature, and compared the changes in mRNA levels between gld-1 and cgh-1ts gonads (Table S1 and Figure 4; encircled transcripts were confirmed by RT-qPCR in subsequent figures). Importantly, we observed that similar transcripts were reduced in each mutant (Figure 4; Pearson correlation coefficient r = 0.426) and that 47% of the transcripts reduced in both gld-1 and cgh-1ts mutants were also GLD-1 targets (Figure 4, GLD-1 targets are in red); about four-fold more than expected by chance (p<2.2e-39, t test; only 12% of all germline mRNAs are bound by GLD-1). To make sure that the observed changes in mRNA levels were not unique to the cgh-1ts allele, we additionally analyzed mRNA changes in animals subjected to cgh-1 RNAi and in cgh-1 null mutants, and obtained similar results (Table S1 and Figure S4A–S4D). Combined, these results suggest that the mRNAs stabilized by GLD-1 are also stabilized by CGH-1. For the purpose of this study, we refer to those transcripts simply as ‘co-regulated mRNAs’ (Table S2; mRNAs that are GLD-1 targets, and which are less abundant in both gld-1 and cgh-1ts mutants).
GLD-1 binds its mRNA targets via specific GLD-1 binding motifs (GBMs), which are mostly present in 3′ UTRs [31]. This enabled us to test a direct versus indirect role of GLD-1 in both mRNA repression and stabilization in wild-type animals, by creating a series of reporters containing either wild-type or mutated GBMs. Specifically, a constitutive germline promoter (mex-5) was used to drive transcription of GFP fused to histone H2B (which concentrates GFP in the nucleus to facilitate detection) [40]. The GFP reporter was fused to various 3′ UTRs of co-regulated mRNAs that either contained wild-type GBMs (GBMwt), allowing GLD-1 binding and regulation, or mutated GBMs (GBMmut), preventing GLD-1 binding and regulation (Figure 5A). We examined the effect of GLD-1 binding on mRNA stability by analyzing the levels of GBMwt/mut reporter pairs by RT-qPCR and found that, in each case, the GBMmut mRNA was less abundant than the corresponding GBMwt mRNA (Figure 5B and Figure S5A). Because these reporters were expressed and analyzed in wild-type animals, and mutated GBMs do not cause destabilization when introduced into the 3′ UTR of a non-target mRNA [31], these results suggest that GLD-1 stabilizes at least some of the co-regulated mRNAs by directly associating with their 3′ UTRs.
Expectedly, we observed that the GBMmut reporters were de-repressed in the medial germ line (Figure 5C and Figure S5B). To test the effect of CGH-1 on translational repression, we crossed the GBMwt and mut reporters into the cgh-1ts mutant, and additionally subjected reporter strains to cgh-1 RNAi. Consistently with the observation that endogenous glp-1 and rme-2 mRNAs are not de-repressed in the medial gonad of cgh-1(RNAi) animals [39], we found that the GBMwt reporters were de-repressed in neither cgh-1ts nor cgh-1(RNAi) animals, nor were the GBMmut variants additionally de-repressed (Figure 5D and Figure S5C). These results suggest that, while CGH-1 contributes to the stabilization of GLD-1 targets, it does not appear to have a general role in their repression.
GLD-1 and CGH-1 may depend on each other for mRNA stabilization or function independently. To test whether GLD-1 and CGH-1 affect mRNA stability in an additive fashion, we determined the levels of endogenous co-regulated mRNAs in gld-1 and cgh-1ts single, and gld-1; cgh-1ts double mutants. We found that mRNA levels, which were reduced in both single mutants, were even further reduced in the gld-1; cgh-1ts double mutant (Figure 6A), suggesting that GLD-1 and CGH-1 may stabilize mRNAs by acting in parallel pathways. Furthermore, to investigate if GLD-1 and CGH-1 depend on each other for mRNA binding, endogenous co-regulated mRNAs were co-precipitated with GLD-1 and CGH-1, from extracts of wild-type and mutant animals. One caveat of this analysis is that the RNA levels in mutant animals are reduced and obtained values need to be normalized to the corresponding input levels. By this approach, we observed that GLD-1 could still bind mRNAs in cgh-1ts and cgh-1(RNAi) animals, and likewise CGH-1 could bind mRNAs in the absence of GLD-1 (Figure 6B–6C and Figure S6). To test this further, we IP-ed GBMwt and GBMmut variants of the co-regulated oma-2 reporter with GLD-1 and CGH-1 from worm lysates. As expected, we found that mutating GBMs in the oma-2 3′UTR dramatically decreased GLD-1 binding (Figure 6D). In contrast, we observed no reduction in CGH-1 binding (Figure 6D). Together, these results suggest that GLD-1 and CGH-1 are recruited to an mRNA independently of each other and that, although GLD-1 and CGH-1 stabilize largely the same mRNAs, their contributions appear to be distinct.
Many of the co-regulated mRNAs encode proteins that have been studied in at least some detail. Remarkably, most of them (34/38) are important during the oocyte-to-embryo transition (Table S2). Some of these proteins function specifically during oogenesis (for example PUF-5; [35]), fertilization (EGG-1; [41]), or early embryogenesis (POS-1; [42]). Others, such as OMA-2, function at multiple times during OET [43]–[45]. Thus, GLD-1 and CGH-1 appear to stabilize messages related by their function in promoting OET. This was unexpected, because GLD-1 is expressed in the medial germ line but is absent from oocytes. To examine this seeming discrepancy, we tested by in situ hybridization whether the reporters of co-regulated mRNAs accumulate in oocytes in a GBM-dependent manner. Indeed, we found that the GBMmut reporters were less abundant not only in the medial, GLD-1 expressing part of the gonad, but also in the proximal gonad, suggesting that GLD-1 mediated stabilization is important for OET mRNA accumulation in oocytes (Figure 7A; position of oocytes is indicated by red brackets; see discussion).
Previously, we identified hundreds of germline transcripts associated with GLD-1 [31]. Although the precise mechanism(s) remain unknown, here we present evidence that in general these messages are repressed by GLD-1, consistently with a recent report describing global protein changes in GLD-1 depleted animals [46]. We found that GLD-1 interacts with components of repressive germline mRNA complexes, including the DDX6 helicase CGH-1. In Xenopus and Drosophila, similar complexes also contain eIF4E-binding proteins (4E-BPs), suggesting that they repress translation by interfering with the assembly of the basic translation initiation factor eIF4F [12], [14], [47]–[50]. Interestingly, in Drosophila, the same proteins have also been implicated in oskar mRNA repression by a cap-independent mechanism, presumably by sequestering mRNAs away from the translation machinery [51]. Because we found neither basic initiation factors nor 4E-BPs among GLD-1 interacting proteins, one possibility is that GLD-1 represses its targets via a similar ‘sequestering’ mechanism, which might also protect them from the decay machinery. However, since GLD-1 appears to stabilize only a subset of its targets, and CGH-1 seems to protect but not repress them, translational repression and mRNA stabilization of co-regulated mRNAs are not necessarily coupled.
Our findings suggest that, in addition to repressing translation, GLD-1 stabilizes a subpopulation of its targets. The most compelling evidence comes from the GBM+/− reporter studies, which directly demonstrate that GLD-1 binding can stabilize a target mRNA. However, we noticed that the changes in mRNA levels induced by GBM mutations were smaller than the changes in the endogenous mRNAs observed between wild type and gld-1 mutants. This could be due to a number of differences between the synthetic and endogenous mRNAs (such as expression from different promoters, splicing, potential additional regulatory motifs, etc.) or reflect a stronger decrease in mRNA levels in the mutant due to indirect effects. Thus, the precise magnitude of GLD-1 mediated stabilization remains to be determined. Besides GLD-1, our findings additionally implicate CGH-1 in the stabilization of some GLD-1 targets, but suggest that the two proteins regulate mRNA stability independently of each other. Possible interpretations of this data are that these proteins largely associate with separate cytoplasmic pools of mRNAs and/or protect mRNAs in different parts of the gonad. Yet, inconsistently with the later scenario, we noticed that the levels of several mRNAs tested by in situ hybridization were reduced also in the medial, GLD-1 expressing parts of cgh-1ts gonads (our unpublished observation). Interestingly, while CGH-1 appears to affect the stability of specific GLD-1 targets, it may be dispensable for their repression. This contrasts with the function of DDX6 helicases in translational repression in some models [12], [48], [49] but agrees with the role of the protist DDX6-like helicase, DOZI, which stabilizes repressed transcripts in female gametocytes [52]. Thus, either DDX6 helicases have distinct roles in different organisms, or their function in mRNA stabilization and/or mRNA repression depends on a specific mRNA.
Our finding, that GLD-1 and CGH-1-dependent stabilization may be important for efficient accumulation of OET transcripts, implies that the decay machinery responsible for the degradation of unprotected mRNAs is active in the germ line. Two GLD-1 targets containing upstream open reading frames are thought to be protected by GLD-1 from nonsense-mediated decay (NMD) [37]. However, we found no evidence that the degradation of unprotected GLD-1 targets described here depends on NMD (Figure S7). Interestingly, many OET mRNAs protected by GLD-1 and CGH-1 are degraded in early embryos (our unpublished observation). Thus, one possibility is that the machinery degrading maternal transcripts in the embryo degrades also unprotected OET mRNAs in the germ line. In D. melanogaster, the embryonic degradation of maternal mRNAs requires the protein Smaug and miRNAs [53]–[55]. Whether related factors degrade maternal mRNAs in the C. elegans embryo and/or unprotected OET mRNAs in the gonad remains to be tested.
Intriguingly, our findings suggest that GLD-1 dependent stabilization of mRNAs is important for their accumulation in oocytes, i.e. in cells in which GLD-1 is no longer present. In C. elegans, oocyte growth depends on an influx of cytoplasmic material originating in undifferentiated, GLD-1 expressing cells, which may be analogous to the cytoplasmic transport from nurse cells into oocytes in the Drosophila ovary [56]. Thus, one explanation for GLD-1 dependent accumulation of mRNAs in oocytes is that GLD-1 binding protects mRNAs before and/or during their transport into growing oocytes (Figure 7B). Once in oocytes, these mRNAs might be stable due to a general suppression of mRNA decay, as described in Xenopus oocytes [2], [3]. Alternatively, GLD-1 might only be required for the initiation but not the maintenance of mRNA protection, which in oocytes may depend on CGH-1 and/or other RBPs.
Animals were typically maintained at 25°C using standard procedures, unless indicated otherwise. The temperature sensitive strain cgh-1(tn691) was maintained at 15°C and shifted to 25°C as L4 larvae for subsequent analysis of adult animals. Synchronous cultures were obtained by collecting eggs from bleached adults and synchronizing larvae by starvation before feeding. In all experiments young adults that produced oocytes but not yet embryos were analyzed. For RNAi experiments, we used the Open Biosystems cgh-1 and smg-2 bacterial strains and, as a control, bacteria harboring an ‘empty’ vector. Larvae were transferred to RNAi feeding plates directly after synchronization and animals were cultured at 25°C.
The following mutant and transgenic strains have been described previously: cgh-1(ok492)/hT2[qIs48]; gld-1(q485)/hT2[qIs48]; rrrSi 38/39/40 [mex-5 pro::PEST:GFP-H2B::oma-2 3′UTR; unc-119(+)]II; and rrrSi 53/54/56[mex-5 pro::PEST:GFP-H2B::oma-2 GBMmut 3′UTR; unc-119(+)]II [11], [31], [57]. The cgh-1(tn691) strain was obtained from CGC (DG1701); the cgh-1(tn691) mutation induces 100% sterility at the restrictive temperature (25°C).
To minimize variation between ‘GBMwt’ and ‘GBMmut’ pairs of reporters, transgenic strains were created by Mos1 transposase mediated Single Copy gene Insertion (MosSCI) into a single genomic locus as previously described [31], [58]. GBM mutations introduced are shown in Table S3. Oligos used to amplify 3′ UTR sequences (from the STOP codon to 50 bp downstream of the polyA site) are described in Table S4. All strains were outcrossed at least twice against wild-type animals before being analyzed. Table S5 shows all reporter strains utilized in this study. The 7 nt substitution that was used to mutate GBMs (in GBMmut reporters) does not by itself destabilize mRNA [31].
We used the COPAS Biosort from Union Biometrica to separate homozygous GFP (−) gld-1 mutants from heterozygous GFP (+) gld-1(q485)/hT2[qIs48] animals.
The assay was performed as previously described [59], with the following changes. Synchronized worms were harvested as young adults, frozen in 100 µl ‘worm pellet’ aliquots. Subsequently, each aliquot was re-suspended in 500 µl lysis buffer. An initial centrifugation step was included (5 min at 5000 g, 4°C) and worm lysates were layered on 5% (w/v) to 45% (w/v) sucrose gradients. To correct for variations in RNA isolation and reverse transcription efficiency between sucrose fractions, we added 2 µg of total RNA from mouse brain (Stratagene) to each fraction. RNA from fractions was extracted using TRIzol (Invitrogen) according to the manufacturer's recommendations. RNA integrity was confirmed on ethidium bromide-stained agarose gels before proceeding to RT. Proteins from fractions were isolated by chloroform/methanol precipitation and investigated by western blotting. To analyze mRNAs by tiling arrays, we extracted RNA from pooled fractions 8 to 12 (polysomal) and fractions 1–12 (total), in four biological replicates.
50 gonads from wild-type, gld-1(q485), and cgh-1(tn691) worms were dissected in triplicates in M9 buffer for tiling array analysis. The PicoPure RNA Isolation Kit was used according to the manufacturer's recommendations to extract RNA from gonads (Figure 2A and Figure 4). To analyze the mRNA abundance of various strains, RNA from 30 animals was extracted with the PicoPure RNA Isolation Kit (Figure 2B, Figure 5B, Figure 6A; Figures S4A, S5A, S7). To determine mRNA levels in mock and cgh-1 RNAi treated animals, RNA was Trizol extracted from Input IP samples according to the manufacturer's recommendations.
Peptides (Bachem) were used to generate mouse monoclonal antibodies according to standard procedures (PAB-1 = aa 542–560; GLD-1 = aa 65–79). PAB-1 antibody was diluted 1∶50 for western blot analysis. The mouse monoclonal GLD-1 antibody was used for immunoprecipitation (100 µl per reaction). Rabbit polyclonal GLD-1 antibody was used for western blot analysis (1∶50 dilution) and immunostaining (1∶500 dilution) [38]. Additional antibodies used: ACT-1 (MAB1501, Chemicon), CAR-1, CGH-1 [11], FLAG M2 (Sigma), GLH-1 [60], Myc (9E10), PGL-1 [61], GLH-1 [62].
GLD-1 and CGH-1 immunoprecipitations were performed as previously described [5], [26], [31]. To globally identify GLD-1 targets by tiling arrays we compared anti-GLD-1 IPs with anti-Myc IPs. To determine GLD-1 mRNA binding in cgh-1ts and cgh-1(RNAi) animals, and CGH-1 mRNA binding in gld-1 animals we compared anti-GLD-1 IPs with anti-FLAG IPs, and anti-CGH-1 IPs with anti-IgG IPs. RNA was eluted from beads with TRIzol. Precipitation efficiency was enhanced by adding 5 µg total RNA from mouse brain (Stratagene) to each IP sample.
GLD-1-associated proteins were identified by comparing anti-GLD-1 IPs with anti-FLAG IPs. RNase treated IP samples were incubated with 0.1 mg/ml RNase A (Qiagen) for 15 minutes at 37°C. Proteins were separated by SDS-PAGE and Coomassie stained. Bands were cut, washed and in-gel digested with trypsin overnight at 37°C. Tryptic peptides were separated on an Agilent 1100 nanoLC system (Agilent Technologies) coupled to an LTQ Orbitrap Velos hybrid mass spectrometer (Thermo Scientific). The LC system was equipped with a Peptide CapTrap column (Michrom BioResources, Inc.) and a capillary column with integrated nanospray tip (75 mm i.d.×100 mm, Swiss BioAnalytics AG) filled with MagicC18 (Michrom Bioresources, Inc.). Elution was performed with a gradient of 0–45% solvent B in 30 min at a flow rate of 400 nl/min. Solvent A consisted of 0.1% formic acid/2% acetonitrile, solvent B was composed of 0.1% formic acid/80% acetonitrile. The mass spectrometer operated in positive mode using the top 20 DDA method. Peptides were identified searching UniProt 15.14 using Mascot Distiller 2.3 and Mascot 2.2 (Matrix Science). Results were compiled in Scaffold 2.06. (Proteome Software).
Reverse transcription reactions were performed using the ImProm-II Reverse Transcription System (Promega). To ensure that we are detecting full-length, polyadenylated transcripts we used oligo dT(15) primers for RT reactions on RNA from polysome profile fractions. Identical results were obtained using random hexamer oligonucleotides. To compare total mRNA levels and analyze co-immunoprecipitated RNA, cDNA was generated using random hexamer primers. qPCR reactions were performed as described previously [26]. At least one primer in each pair is specific for an exon-exon junction (Table S4). Mouse RNA (Cyt-c) was added to polysomal fractions before RNA isolation and RT, allowing us to normalize all obtained qPCR results to Cyt-c, thereby correcting for variations in RNA isolation and RT. To compare mRNA levels between different mutants and analyze co-immunoprecipitated RNA, qPCR results were normalized as indicated.
300 ng of RNA (pooled gradient fractions, IP, RNAi treated animal extracts) or 5 µl of RNA (corresponding to 25 dissected gonads and isolated with the PicoPure Kit) were amplified once into dsDNA. 7.5 µg of cDNA was subsequently fragmented and labeled according to the “GeneChip Expression Analysis Technical Manual" (Affymetrix). 6 µg of fragmented and labeled DNA were hybridized to the Affymetrix C. elegans tiling array chip according to the Affymetrix Expression Analysis Technical Manual. Microarray sample preparation, hybridization and scanning were performed in the FMI genomics facility.
Tiling arrays were processed in R (www.r-project.org; [63]) using bioconductor [64], and the packages tilingArray [65] and preprocessCore. The arrays were RMA background corrected and log2 transformed on the oligo level using the command:We mapped the oligos from the tiling array (bpmap file from www.affymetrix.com) to the C. elegans genome assembly ce6 (www.genome.ucsc.edu) using bowtie [66] allowing no error and unique mapping position. Expression of individual transcripts was calculated by intersecting the genomic positions of oligos with transcript annotation (WormBase WS190) and averaging the intensity of the respective oligos. Quantile normalization: each of the datasets was processed with an individual quantile normalization scheme. For IP experiments, no quantile normalization was performed as the distribution between GLD-1 IPs and control IPs differs substantially. In the case of the polysome dataset (containing polysome and total RNA samples) quantile normalization was performed twice. Once containing all the polysome samples and once for all the total RNA samples. The dataset containing either total RNA from purified gonads or total RNA from mock and cgh-1(RNAi) treated animals were each quantile normalized in one single step.
Averages, fold changes (including polysomal shifts) and standard deviations of all analyses are shown in Table S6. The p values in Figure 1D, Figure 4, Figure 5B; Figures S1E, S2 and S5A were calculated with a t test. To further investigate co-regulated mRNAs, we used the following cut-offs: gld-1/wild-type [log2]<−1; cgh-1ts/wild-type [log2]<−0.5. Data discussed in this publication has been deposited in NCBI's Gene Expression Omnibus and is accessible through GEO Series accession number GSE33084 (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE33084).
RNA in situ hybridization was performed and analyzed as previously described [26]. The probes generated from cDNA correspond to 1–714 (gfp) (Table S4). gfp antisense control on wild-type animals was included and negative. More than 60 gonads were scored and we obtained the following numbers (n = number of scored gonads):
oma-2 GBMwt (n = 85; 27% strong; 65% medium; 6% weak); shown is a medium stained gonad
oma-2 GBMmut (n = 65; 12% medium; 88% weak); shown is a weakly stained gonad
egg-1 GBMwt (n = 70; 16% strong; 81% medium; 3% weak); shown is a medium stained gonad
egg-1 GBMmut (n = 68; 35% medium; 65% weak); shown is a weakly stained gonad
Unless indicated otherwise, images were captured with a Zeiss AxioImager Z1 microscope, equipped with an Axiocam MRm REV2 CCD camera. Images were acquired in the linear mode of the Axiovision software (Zeiss) and processed with Adobe Photoshop CS4 in an identical manner.
An LSM700 confocal microscope equipped with a Plan-Apochromat 63×/1.40 Oil DIC M27 objective was used to capture images with a voxel size of 0.052 µm×0.052 µm×0.2 µm (x, y, z). Used lasers: track 1: 405 nm (2%) and 555 nm (10%); track 2: 488 nm (4%). Beam splitters: MBS 405/488/555/639; DBS1: 531 nm (track1) and 578 nm (track 2). Filters: SP490 (Track 1,Channel 1); LP 560 (Track 1, Channel 2); 0–587 (Track 2, Channel 1). Pinhole: 40 µm (Track 1); 41 µm (Track 2). Pictures were deconvolved with the Huygens software, using Remote Manager v1.2.3, a SNR of 8, 8, 8, 100 iterations, and the cmle deconvolution algorithm (quality change stopping criterion: 0.1). Deconvolved images were processed in Imaris XP 7.1.1 using the coloc function. In Figure S3C, only 5.41% of the total data set voxels co-localize.
|
10.1371/journal.ppat.1006476 | Innate immunity restricts Citrobacter rodentium A/E pathogenesis initiation to an early window of opportunity | Citrobacter rodentium infection is a mouse model for the important human diarrheal infection caused by enteropathogenic E. coli (EPEC). The pathogenesis of both species is very similar and depends on their unique ability to form intimately epithelium-adherent microcolonies, also known as “attachment/effacement” (A/E) lesions. These microcolonies must be dynamic and able to self-renew by continuous re-infection of the rapidly regenerating epithelium. It is unknown whether sustained epithelial A/E lesion pathogenesis is achieved through re-infection by planktonic bacteria from the luminal compartment or local spread of sessile bacteria without a planktonic phase. Focusing on the earliest events as C. rodentium becomes established, we show here that all colonic epithelial A/E microcolonies are clonal bacterial populations, and thus depend on local clonal growth to persist. In wild-type mice, microcolonies are established exclusively within the first 18 hours of infection. These early events shape the ongoing intestinal geography and severity of infection despite the continuous presence of phenotypically virulent luminal bacteria. Mechanistically, induced resistance to A/E lesion de-novo formation is mediated by TLR-MyD88/Trif-dependent signaling and is induced specifically by virulent C. rodentium in a virulence gene-dependent manner. Our data demonstrate that the establishment phase of C. rodentium pathogenesis in vivo is restricted to a very short window of opportunity that determines both disease geography and severity.
| The so-called “attaching and effacing” (A/E) pathogens enteropathogenic E. coli (EPEC) and enterohemorrhagic E. coli (EHEC) cause serious human diarrheal infections by adhering to and damaging the intestinal epithelium. Previous work on the mouse A/E pathogen Citrobacter rodentium has established that host adaptive immune response and intestinal microbiota cooperate to control the epithelial infection and colonization with this pathogen. We found that this is complemented by a rapid pathogen-induced mucosal innate immune response that is essential to prevent excessive pathogenesis before adaptive immunity takes effect. Its effectiveness is demonstrated by the fact that it normally limits the duration during which the bacteria can induce epithelial lesions to the first 18 hours of infection. Later, luminal virulent bacteria can no longer induce new A/E lesions, but those induced already in the first 18 hours of infection persist through localized epithelial re-infection. Severity of the disease at the peak of infection is consequently shaped by the early events of the first 18 hours. This information may be important for the development of effective therapies and vaccines.
| Enteropathogenic Escherichia coli (EPEC) and enterohemorrhagic E. coli (EHEC) serotype O157:H7 remain important causes of human diarrhea and mortality worldwide [1,2]. Citrobacter rodentium is a natural pathogen of mice which is used to model human EPEC and EHEC infection, as it shares the pathogenic mechanism of epithelial “attachment and effacement” (A/E). These pathogens attach to and colonize the intestinal mucosal surface in the form of epithelium-adherent microcolonies, also known as A/E lesions [3], comprised of bacteria that intimately attach to the apical surface of epithelial cells and induce the local destruction (“effacement”) of the epithelial brush border. This triggers reactive colonic epithelial hyperplasia, inflammation and the induction of IL-17-producing CD4 helper T (Th17) cells and IL-22-producing type 3 innate lymphoid cells (ILC3) [4,5]. C. rodentium has thus become a valued A/E infection mouse model.
Disease severity and mucosal pathology elicited by C. rodentium correlate with the extent of A/E microcolony formation. The intestinal epithelium protects its structural integrity through rapid cellular renewal driven by crypt stem cell proliferation and crypt-to-villous migration of differentiated epithelial cells, with a complete crypt turnover time of typically 3–5 days in the healthy mouse [6]. Epithelial A/E infection with C. rodentium lasts for 2–3 weeks. Thus, to compensate for the associated continuous exfoliation of infected cells, A/E microcolonies must also be dynamic and able to renew by continuous epithelial re-infection. The route of this re-infection, required for sustained epithelial A/E pathogenesis, and the dynamics of A/E lesion development remain largely unknown. Re-infection by planktonic luminal colonizers or bacteria shed from A/E microcolonies, or local spread of sessile A/E lesion-associated bacteria without planktonic phase may all contribute (Fig 1A).
All A/E pathogens share the intestinal virulence determinant “locus of enterocyte effacement” (LEE), a genomic island that is essential for A/E lesion formation [3]. LEE encodes a type 3 secretion system that translocates a cocktail of type 3 effector proteins into enterocytes, including Tir (Translocated intimin receptor) that inserts into the enterocyte plasma membrane and functions as adhesion receptor for the LEE-encoded outer membrane protein intimin. The Tir-intimin interaction triggers actin rearrangements that cause the epithelial microvilli to vanish (“effacement”) and pedestal-shaped membrane structures to emerge underneath the attached bacteria [3,7]. A/E virulence expression is regulated at the level of LEE transcription through master regulator Ler (encoded in LEE) that is in turn subject to a complex regulatory network that responds to intestinal environmental factors [3,8]. Ler-deficient bacteria are consequently avirulent [7].
Resolution of mucosal C. rodentium A/E pathology within approximately 21 days of infection depends on functional B and CD4+ T cells [9,10]. Furthermore, IL-17-producing Th17 cells, IL-22 production by type 3 innate lymphoid cells (ILC3) and Th22 CD4+ effector T cells have all been shown to contribute to C. rodentium infection control [4,5]. A complex gut microbiota acts as a biological barrier against intestinal C. rodentium colonization [11,12] and is necessary to clear C. rodentium from the intestinal lumen once mucosal A/E lesions and inflammation have resolved [13]. However, even though germ-free animals cannot clear C. rodentium from the gut lumen, they can survive and resolve C. rodentium mucosal pathology with normal kinetics (within approximately 21 days), by virulence-neutralizing adaptive humoral immune responses [13,14].
In wild-type animals, C. rodentium-induced A/E pathology progresses slowly with marked intestinal pathology not developing until after 4–5 days of infection. This delay has been assumed to result from microbiota-dependent colonization resistance decelerating intestinal colonization. Disease onset and severity are also strongly determined by innate immunity. MyD88 knockout [15,16] and MyD88-Trif double knockout mice in particular [17], which are deficient in Toll-like receptor (TLR) and IL-1 receptor (IL-1R) superfamily signaling, develop exacerbated and accelerated A/E pathology.
Previous studies have predominantly focused on the acute and resolution phases of infection. Little is known about the nature and importance of host-bacteria interactions occurring during the first hours of infection. In the present study we focused on the early dynamics and control of A/E lesion development and show that the initiation of colonic C. rodentium microcolonies is limited to a very early stage of infection. These early-established, clonally-renewing A/E lesions trigger a local innate immune response that protects the epithelium from de novo A/E lesion formation by luminal bacteria and thus critically determine disease severity and outcome.
To track and quantify C. rodentium microcolony development in vivo over the course of infection, we first inoculated SPF wild type mice with GFP (green fluorescent protein)- and mCherry (red fluorescent protein)-expressing C. rodentium wild-type bacteria. Microscopic analysis of animals that had been inoculated with 1010 CFU of a 1:1 mixture of GFP- and mCherry-tagged bacteria at day 7 post infection revealed that only single-colored microcolonies formed in the colon (Fig 1E), showing that they develop by local clonal expansion (that is, they are derived from the local proliferation of single bacteria; Fig 1E). The fact that many of these clonal microcolonies span several neighboring epithelial cells shows that luminally colonizing bacteria do not contribute to the epithelial re-infection required for the maintenance of established epithelial A/E lesions; otherwise an amalgamation between green and red bacteria over time or a random mosaic of green and red fluorescent bacteria-infected epithelial cells would be expected, neither of which could be observed over the entire course of infection. Thus, colonic A/E lesions trace back to single bacteria and appear to depend on local cell-to-cell spread and not continuous re-seeding from planktonic luminal colonizers to compensate for epithelial shedding. These data further reveal that the epithelium-adherent population of C. rodentium is derived from a very small founder population (one bacterium per microcolony) and therefore subject to a population bottleneck.
We next addressed whether this population bottleneck is caused by the colonization barrier conferred by the gut microbiota. Murine C. rodentium infection is subject to intestinal colonization resistance conferred by a complex gut microbiota through metabolic competition for simple sugars [12,13]. In agreement with this previous work, we found that SPF animals were completely protected against colonization with an inoculum of 104 CFU (Fig 1B), whereas mice inoculated with a very large inoculum of 1010 CFU became consistently, albeit still variably, colonized with mean fecal densities that gradually increased to reach a plateau by around day 5 post infection (Fig 1B).
We hypothesized that microbiota-mediated metabolic competition not only impacts on intestinal luminal colonization density but also restricts epithelial colonization and A/E pathogenesis. Colonization resistance may be responsible for the observed population bottleneck in the A/E lesion-associated bacteria through restriction of bacterial access to the epithelium or limitation of growth and survival of attached bacteria, either through direct metabolic competition or by maturing the epithelium and inducing protective non-specific immunity. To test this hypothesis, we studied germ-free mice infected with different doses of C. rodentium. As expected, in germ-free mice fecal bacterial populations rapidly and invariably reached high densities of between 109 and 1010 CFU/g, regardless of inoculum dose (Fig 1C and 1D; S1A Fig). However, the striking clonality of the A/E lesions under SPF conditions was precisely phenocopied in germ-free mice, and was hence independent of microbiota-associated luminal colonization resistance or the effects of microbiota on mucosal immunity (Fig 1F and 1H). Taken together, these experiments revealed that the population bottleneck of early C. rodentium infection is mediated independently of microbiota-mediated effects.
We next quantitated the numbers of colonic epithelial A/E microcolonies in germ-free and SPF animals over time by fluorescence microscopy (Fig 1I and 1J). Microcolony numbers correlated well with the fraction of infected epithelial area measured by computational image analysis [18] (supplementary S1C and S1D Fig). Unlike in colonization resistant SPF mice (Fig 1E, 1G and 1I), the infection of germ-free mice with a wide range of inoculum sizes (ranging from 102 to 108 CFU) led to peak numbers of colonic A/E lesions that were similar to those observed in SPF mice infected with 1010 CFU (Fig 1J and S1B Fig). Thus, in absence of colonization resistance, also lower inocula led to efficient epithelial A/E infection.
Remarkably however, the inoculation of germ-free mice with the extremely high dose of 1010 CFU was associated with significantly reduced virulence, leading to significantly lowered numbers of colonic A/E lesions (Fig 1F; compare with Fig 1H). Although the intestinal colonization kinetics in lower-dose versus high-dose infections only differed during the first <16 hours (Fig 1D and S2B Fig), A/E lesion numbers remained significantly decreased over the entire course of infection (Fig 1J). This was associated with significantly improved clinical disease parameters between day 8 and 21, including weight loss (Fig 2A and 2B), systemic bacteremia (manifesting in live recoverability of C. rodentium from the spleen; Fig 2C), colonic mucosal histopathology (Fig 2D and 2E), and fecal concentrations of lipocalin-2 (a fecal marker for colitis severity; Fig 2F).
We considered two alternative explanations for the mechanism by which the size of the inoculum in germ-free animals could have such a persistent effect on disease severity: First, ongoing bacteria-intrinsic downregulation of virulence in luminal C. rodentium after reaching high intestinal luminal density, or second, inhibition of ongoing A/E lesion induction by an early C. rodentium-induced host immune response.
Bacteria-intrinsic virulence regulation is considered important for the in vivo fitness of this pathogen [19]. We found that colonic luminal C. rodentium ler transcription was reduced at 16 hours after inoculation in mice that had been inoculated with 1010 CFU compared to mice inoculated with only 104 CFU (S2A and S2B Fig). The main difference between the infection with an inoculum of 1010 CFU and that with lower inocula is in the early population dynamics in vivo. Upon inoculation, bacterial populations enter an early logarithmic growth phase followed by a steady-state stationary phase (see Fig 1D). In germ-free mice inoculated with 104 CFU, the logarithmic growth phase lasts for approximately 15.5 h (Fig 1D). In germ-free mice inoculated with 1010 CFU, the bacteria lack this growth phase (Fig 1D). The associated colonic virulence gene expression differences we measured at 16 hours after inoculation could explain a delayed disease onset, but are an unlikely sole explanation for the maintained reduction in A/E lesion burden, as it is known that fully colonized germ-free animals remain luminally colonized with phenotypically virulent (ler-expressing) bacteria for at least 7 days [13]. Moreover, population density-dependent virulence regulation by the croIR-encoded acyl homoserine lactone (AHL) type quorum sensing system, reported previously to regulate bacterial adhesion and virulence negatively (in an ler/LEE-independent manner) at high population density [20], did not explain our observations. Infection with 1010 CFU of a quorum sensing-deficient croI mutant reproduced the reduced A/E lesion density phenotype (S2D Fig).
In support of a host-mediated mechanism, however, we found that grossly innate immunocompromised germ-free Myd88/Trif double-deficient mice [21], unlike wild-type animals (Figs 1F, 1H, 1J and 2), did not show a persistent disease attenuation following infection with 1010 CFU of C. rodentium. Even though the reduction of ler expression following infection with 1010 CFU reproduced in MyD88/Trif-deficient mice (Fig 3A) and was associated with a transiently reduced extent of colonic A/E microcolony formation on day 2 post infection, microcolony numbers were no longer different on day 3 post infection (Fig 3B and 3D). In an independent experiment, the weight loss of infected MyD88/Trif double-deficient germ-free mice was followed until animals reached a defined ethical endpoint (≤ 80% initial body weight). Mice that were infected with 1010 CFU continued to lose weight at a rate similar to mice infected with 104 CFU, but with a delay of approximately 24 hours (Fig 3C). These data suggest that an early MyD88/Trif-dependent immune response restricts the ongoing induction of A/E lesions, rather than the maintained downregulation of bacterial virulence.
MyD88-/Trif double-knockout mice are deficient for TLR as well as IL-1R family receptor signaling. To investigate the role of TLR- and inflammasome-related pro-inflammatory signaling, we analyzed germ-free Caspase-1/11-deficient mice. Although these mice have been described as hyper-susceptible to C. rodentium infection [22], we found that they developed a persistently attenuated A/E pathogenesis in response to an infection with 1010 CFU of C. rodentium, similar to wild-type mice (S3 Fig). Thus, mucosal innate immune activation through MyD88/Trif-TLR-signaling determines severity of ongoing A/E pathogenesis in an early phase of infection.
Taken together, our data indicated that an early host innate immune response is induced at an early stage of infection and determines ongoing disease severity by limiting the induction of A/E lesions to an early window of opportunity. The remaining course of infection would then be driven by the early induced and persisting A/E lesions. A testable prediction of this hypothesis is that C. rodentium-infected wild-type animals should become insensitive to superinfection with a congenic strain of C. rodentium after the postulated window of opportunity.
To test this hypothesis, we developed a strain combination infection-superinfection model, in which two differentially tagged C. rodentium strains could be administered sequentially. This allowed us to elucidate the effect of the infection with the first strain on the pathogenesis of the second and how this depended on bacterial genotype, host genotype, and time interval between primary and superinfection (see schematic in Fig 4A). The aim of this approach was (i) to demonstrate and map the duration of the proposed infection window of opportunity directly, (ii) to show definitively the inhibition of A/E lesion initiation by luminal bacteria outside the window of opportunity, and (iii) to differentiate between bacterial and host contributions to its induction.
Germ-free mice were pre-infected with a kanamycin-resistant (KanR) mCherry-tagged C. rodentium strain (wild-type or ler mutant), followed 18 hours later by super-infection (or in the negative-control group co-infected) with 104 CFU of an isogenic tetracycline-resistant (TetR) wild-type strain (schematized and data shown in Fig 4A). All groups were treated orally with the bacteriostatic antibiotic tetracycline starting 3 h prior to the superinfection to shift colonization from the Tet-sensitive primary strain to the Tet-resistant super-infecting strain. The luminal colonization kinetics of each strain were determined by selective bacterial plating from feces (Fig 4A). On day 7 after superinfection the animals were sacrificed and colonic A/E lesions (all of which were derived from super-infection, and therefore mCherry-negative) were quantified after labeling of C. rodentium with an anti-C. rodentium LPS antibody (Fig 4B–4G).
Super-infection was influenced by bacterial-bacterial competition between pre-and super-infecting bacteria, since there was overlap between pre- and super-infection. The co-infection control in which pre- and super-infection inocula were administered simultaneously (Fig 4, “1010 WT 0h”) consequently presented with 2.2±3.7 -fold (mean±SD) reduced numbers of colonic A/E lesions compared to the 104-CFU-super-infection-only control (Fig 4B). This effect reduced A/E lesion density to values similar to those observed in the super-infection-only control inoculated with 1010 C. rodentium (Fig 4, “super-infection only 1010 WT). However, 18-hour pre-infection with 1010 wild-type C. rodentium significantly reduced the A/E lesion formation a further 3.3±0.3 -fold relative to the co-infection control (Fig 4B and 4D). Remarkably, 18-hour pre-infection with an isogenic avirulent ler mutant had no such additional protective effect (Fig 4B and 4E), showing that LEE-dependent (ler regulated) pathogenesis in the first 18 hours of pre-infection induced epithelial resistance to A/E lesion induction. Moreover, the effect was even more pronounced when pre-infecting with 104 (near-full protection, more than 25-fold reduced A/E lesion relative to co-infection control; Fig 4B and 4C) instead of 1010 wild-type C. rodentium, indicating that the epithelial resistance induction by pre-infection correlated with infectiousness rather than bacterial load of the pre-infection inoculum.
This early mucosal resistance phenotype was associated with an early increase of mucosal protein levels of IL-1β, IL-22, KC/CXCL1 and MCP-1 (Fig 5D–5G) and elevated transcription rates of Cxcl1 and Nos2/iNos (Fig 5B and 5C) measured at 18 hours post infection. At the same time point we observed the increased recruitment of neutrophils and monocytes into blood and lamina propria, associated with an early increase of lipocalin-2 (mainly produced by neutrophils) detectable in the feces (Fig 5A and 5H–5K). Colonic histological analysis confirmed that these changes occurred prior to the onset of acute colonic histopathology (S4 Fig), although the response markers measured may overlap with those associated with the acute A/E lesion-associated inflammation developing later. These responses were induced specifically by virulent C. rodentium, but not an avirulent ler mutant or the commensal E. coli strain HS. Thus, the virulent C. rodentium-induced resistance to epithelial A/E superinfection with luminal bacteria was associated with an early mucosal innate immune response. It was more potently induced by a low-dose (104 CFU) than a high-dose (1010 CFU) infection with wild-type C. rodentium. Hence, LEE/ler-dependent C. rodentium virulence, rather than unspecific bacterial innate immune receptor agonists like LPS (expected to be increased upon administration of 1010 bacteria and similar between C. rodentium and E. coli), appears to be the main driver of this response. Finally, the response was abrogated in MyD88-/Trif-double-deficient mice (Fig 5, open symbols). MyD88/Trif double-deficient mice consequently lacked protection against epithelial superinfection, with no protective effect of wild-type C. rodentium pre-infection over the control pre-treatment with a Δler mutant (Fig 6A–6C). The specific cellular and molecular innate immune components necessary and sufficient for this protective mechanism remain to be identified.
Overall, our data show that mucosal innate immunity, specifically activated by virulent C. rodentium in an ler/LEE-driven and TLR-signaling dependent manner, very effectively limits the establishment of A/E pathogenesis of C. rodentium to a window of opportunity of <18 hours duration that determines ongoing geography and severity of infection.
We show in this paper that the development of colonic C. rodentium A/E pathogenesis is critically dependent on very early bacteria-host interactions. The characteristic A/E lesions that determine the severity of infection originate from a limiting number of single-bacterial epithelial infection events that need to occur within the first 18 hours of infection. The ongoing infection is driven by the early induced and locally growing epithelial A/E lesions that counteract epithelial shedding by continuous localized epithelial reinfection. Whilst this localized epithelial re-infection allows A/E lesions to persist on the epithelium for more than 10 days, until an adaptive immune response clears the infection, the window of opportunity for their induction by planktonic bacteria present in the lumen is limited to the first 18 hours. This is remarkable because the colonic lumen is known to remain colonized by phenotypically virulent bacteria for at least a week [13] and these bacteria are assumed to be more infectious than in-vitro-grown bacteria, which is important for effective fecal-to-oral transmission [23]. Restriction of A/E lesion induction to such an early time window is due to an early host response specifically induced by virulent C. rodentium, but not by an avirulent C. rodentium mutant. Similarly, mucosal stimulation with the related Gram-negative commensal E. coli could not phenocopy the observed phenomenon. Also the physiologic mucosal conditioning with a complex microbiota does not protect against epithelial A/E pathogenesis, but only impedes intestinal colonization with this pathogen. The abolition of the pathogen-induced mucosal resistance by a deficiency in MyD88-/Trif-signaling and superinfection experiments clearly show that an innate host response, rather than bacteria-intrinsic virulence gene regulation or bacteria-bacteria interaction, is mechanistically responsible. The specific cellular and molecular innate immune components required and sufficient for this protective effect remain to be identified. Whilst highly effective against induction of A/E lesions, the MyD88-/Trif-dependent innate immunity is not sufficient to abolish the localized re-infection occurring around persisting A/E lesions as well, although it may limit the rate of local spread of A/E lesions across the epithelial surface.
Notwithstanding the fact that early host immunity is the main driver of the phenomenon described, the bacteria-intrinsic regulation of C. rodentium virulence gene expression is known to be central to the pathogenesis of this species. The altered virulence of C. rodentium following an inoculation in unphysiologically high numbers of 1010/dose into germ-free animals provided us with an experimental tool to demonstrate the importance of bacterial virulence during the critical early phase of infection. In this context individual A/E lesions developed normally but in persistently reduced number, with extended areas of surface epithelium remaining uninfected. The ongoing epithelial re-infection that must occur to compensate for the effect of epithelial shedding thus appears to have a remarkably short range and may depend on epithelial adherence that is already established. Exact elucidation of this specialized re-infection mechanism is beyond the scope of the present study but warrants future efforts.
Our work in gnotobiotic models demonstrates that colonic A/E lesions induced by C. rodentium in an early phase of the infection—in immunocompetent mice—limit the ongoing induction of new A/E lesions. As a consequence the survival and localized renewal of early-induced A/E lesions, rather than the continued exposure to virulent bacteria in the environment or intestinal lumen is the main determinant of infection severity. This was unexpected, as the virulent bacteria shed into the colonic content and consequently the feces have been shown previously to have a “hyper-infectious” phenotype [23]. However, the situation in conventional, colonization resistant, mice may be more complex: It has been shown that, in SPF mice, in-vitro-grown C. rodentium initially infects and accumulates mainly at the cecal patch and only approximately two days later spreads to the colon, strongly suggesting that longitudinal cecal-to-colonic re-infection can occur [23]. This has been ascribed to an “hyper-infectiousness” phenotype of C. rodentium acquired during intestinal passage that is associated with the increased expression of LEE and related virulence determinants [24]. It has further been shown by others that virulence is required to efficiently overcome intestinal colonization resistance in SPF mice [7,13]. In contrast to the delayed colonic infection observed with in-vitro-grown C. rodentium, the inoculation of naïve SPF animals with feces from infected mice led to the more direct infection of the colon [23]. It is readily conceivable that in mice receiving an in-vitro-grown inoculum, such a “hyper-infectious” bacterial population can build up not only during the initial cecal patch infection observed in SPF mice, but also more rapidly in the non-colonization-resistant germ-free intestine. Whilst our experiments in SPF mice (also using in-vitro-grown bacteria) recapitulated the previously described 2-day delay in colonic infection (see Fig 1), the colonic infection kinetics in the germ-free mice was equivalent to that described for SPF mice infected via the fecal-to-oral route. Thus, in absence of microbiota-mediated colonization resistance, cecal patch colonization may not be a prerequisite for colonic infection, but we cannot rule out that cecal infection plays an important role in all models, also independent of colonization resistance.
Our observation that very high inocula of (stationary phase) C. rodentium in germ-free mice are hypovirulent is relevant for the correct interpretation of earlier work in germ-free animals. For instance, the surprising finding that germ-free mice survive and can clear C. rodentium infection with similar mucosal immune activation kinetics as SPF mice, described recently [13], was reproduced in our experiments but depended on the infectious dose. We found a very large dose of 1010 CFU (most labs use between 107 and 109 CFU) was less virulent in germ-free than in SPF mice. Conversely, we show that an “optimal” dose of 104 CFU is highly pathogenic in germ-free mice, yet below the minimum infectious dose in SPF mice due to microbiota-related colonization resistance. Thus, it is not straightforward to compare C. rodentium disease kinetics between animals with different microbiota statuses.
Our findings may also be relevant for other intestinal pathogens. Clonal intraepithelial microcolony formation has recently been described for the invasive epithelial pathogenesis of Salmonella typhimurium [25,26]. Sellin et al. described an early epithelial-intrinsic protective mechanism dependent on activation of epithelial Nod-like receptor NLRC4 and Caspase-1 that leads to elimination of infected enterocytes and curbs epithelial infection within the first hours of infection [25]. C. rodentium A/E infection has been shown to be more severe in NLRC4- [27] and Caspase-1-deficient mice [22]. Although we found that Caspase-1/11-deficient animals did not phenocopy MyD88-/Trif-deficiency in our analyses, the very early mucosal innate immune response appears to be similarly important in the control of invasive intestinal pathogens.
Further studies are needed to analyze how the new information obtained with the murine C. rodentium model translates into human A/E infection. EPEC-induced acute diarrhea is primarily a neonatal and infantile infection, whereas adults are relatively protected from natural infection by intestinal colonization resistance and develop disease only after intake of very large inocula (108−1010 CFU; [28]). A recent study using a novel neonatal murine EPEC infection model revealed strikingly similar A/E microcolony formation in the neonatal small intestine, and revealed an age-dependent but microbiota-independent resistance of the adult epithelium against A/E lesion formation [29].
Our findings may have several medical implications. First, since the function of LEE is highly conserved between murine and human A/E pathogens, the observed LEE-mediated induction of epithelial resistance to super-infection may also be relevant to human infection. We have identified the initial 18 hours of murine A/E infection as a critical time window for controlling and surviving, or potentially preventing the infection. Even incompletely or transiently neutralizing antibodies may thus be able to substantially reduce intestinal morbidity. As human A/E infection is most severe in newborns and infants, the potential benefit of maternal immunity, protective lactational antibodies and mucosal vaccination will be an important aspect of future work. Remarkably, to the authors’ knowledge it is still unclear to what extent human EPEC infection generates long lasting immunity or neonate-protective maternal antibodies [30], and there is no EPEC vaccine currently available. However, we also need to consider the possibility that specific intestinal antibodies, through dampening the observed rapid innate immune activation, may alter disease progression and even worsen disease outcome.
Second, antibiotic therapy is not a recommended therapy for human A/E infection, and contraindicated in EHEC infection where antibiotic treatment is associated with an increased risk of hemolytic-uremic syndrome. Due to worldwide emergence of multi-resistant E. coli strains and the negative effect on microbiota-conferred colonization resistance, empiric antibiotic therapy is limited to severe cases. However, our findings in the murine model demonstrate that A/E microcolony formation is sensitive to antibiotic treatment, whilst epithelial re-infection by the incompletely eradicated luminal bacteria or superinfection by resistant bacteria is efficiently inhibited by mucosal innate immunity. Hence, optimized antibiotic regimens or novel selective antimicrobials, such as type 3 secretion system inhibitors [31], may have therapeutic potential in the treatment of EPEC infection.
All animal experiments were performed according to protocols approved by the Bernese cantonal ethical committee for animal experiments and carried out in accordance with Swiss Federal law for animal experimentation (license numbers BE94/11 and BE91/14).
All mice had a C57BL/6 background. Germ-free animals were derived and maintained germ-free in flexible film isolators in the Genaxen Foundation Clean Mouse Facility (CMF) of the University of Bern as described [32]. Myd88-/- Ticam1/Triflps/lps mice were kindly provided by Prof B. A. Beutler, The Scripps Research Institute (la Jolla, CA, USA) [33] and derived germ-free as described [21]. Germ-free caspase-1/11-deficient mice [34] were generated by axenic embryo transfer of cryopreserved 2-cell-stage embryos (The Jackson Laboratory, Bar Harbor ME, USA). Experimental germ-free mice were aseptically transferred to autoclaved sealsafe-plus individually ventilated cages (IVCs) under positive pressure (Tecniplast, Italy) in a gnotobiotic barrier unit. Cage changes were carried out under strictly aseptic conditions. In all experiments animals were provided with sterile mouse chow (Kliba 3437; autoclaved) and autoclaved water ad libitum. SPF mice were acquired from Envigo (former Harlan, the Netherlands) and housed in IVCs in the Zentrale Tierställe of the University of Bern.
To generate contamination-free bacterial inocula, LB medium supplemented with appropriate antibiotics (see section bacterial culture for details) in sterile-filter-sealed flasks was aseptically inoculated from single colonies of the test bacterium and incubated, with shaking at 150 rpm at 37°C for 16 h. Bacteria were harvested by centrifugation (7 min, 4700 × g, 4°C) in a sterile aerosol-proof assembly, washed in autoclaved sterile PBS and concentrated to a density of 5 × 1010 CFU/mL in sterile PBS, performed aseptically under a sterile laminar airflow. To generate lower densities cultures were diluted appropriately in PBS. The bacterial suspensions were aseptically aliquoted in autoclaved plastic tubes and sealed in a sterilized secondary containment. The sterile tubes containing the inocula and germ-free mice were aseptically imported into a sterilized laminar flow hood, and each animal inoculated with 200 μL of bacterial suspension by gavage, carried out wearing sterile surgical gowns and sterile surgical gloves. Fresh fecal pellets were collected aseptically, suspended in sterile PBS, and plated in serial dilutions on LB plates containing the appropriate antibiotics and incubated at 37°C for 24 h. Weight of mice was measured using a hand held balance and calculated as a percentage of the body weight in the beginning of the experiment. For tetracycline treatment, mice were gavaged with 200 μL of a sterile tetracycline solution (50 mg/L) containing 2% sucrose, and the same solution was used as drinking water. The tetracycline-water was protected from light and replaced every 3 days.
LB medium (Sigma-Aldrich, Germany) was used as the standard growth media. Where required, the following supplements were added to the media: chloramphenicol (Sigma-Aldrich, China, 6 μg/mL), tetracycline (Sigma-Aldrich, USA, 12.5 μg/mL), kanamycin (Calbiochem, USA, 50 μg/mL). Tetracycline counterselective plates were prepared as described [35,36].
All bacterial strains and plasmids used or generated in this study are specified in S1 Table. All Citrobacter rodentium strains used in this study are derivatives of the wild type strain ATCC51459 (ATCC, Manassas VA). All deletions were carried out by Lambda Red recombineering. Mutagenesis primer sequences are specified in S2 Table. (i) Strain HA526 (Δler::tetRA) was generated by deletion of ler using recombineering plasmid pSIM5 as described [36–38]. A tetRA recombineering amplicon was amplified from genomic tetRA template DNA (isolated from a Tn10-containing bacterial strain) with primers CR-ler-mutF and CR-ler-mutR. (ii) The tetRA cassette was removed with the same protocol using primer CR-ler-rmvl. Tetracycline sensitive clones were selected by growth on Tetracycline counterselective plates [35,36] at 32°C, for 2 days, yielding HA539 (Δler). (iii) Following the same protocol as (i), primers CR-croI-mutF and CR-croI-mutR were used to amplify the tetRA amplicon. The resulting amplicon was used to delete croI in C. rodentium resulting in strain HA528 (croI::tetRA). A tetracycline-resistant strain of C. rodentium was constructed by inserting the tetRA cassette into the locus of dadX, using the primers CR-dadX-mut-F and CR-dadX-mut-R, yielding in strain HA538 (Δdadx::tetRA). Plasmid pHA500 was derived from pM965 [39], by exchanging the ampicillin resistance with a kanamycin resistance cassette. This was done using pSIM9 and lambda red recombineering. A kan amplicon was amplified using pKD4 [40] as a template with primers pM965-kan-F and pM965-kan-R. The recombineering was carried out in an E. coli K12 strain. Plasmid pHA501 was derived from pM2120 [41], by exchanging the ampicillin resistance with a kanamycin resistance. This was done using pSIM9 and lambda red recombineering. A kan amplicon was amplified using pKD4 [40] as a template with primers pM2120-kan-F and pM2120-kan-R. The recombineering was carried out in an E. coli K12 strain. All deletions were verified phenotypically and by control PCR (control primers specified in S2 Table). Plasmids (see S1 Table for a complete list) were introduced by electroporation following standard protocols.
The distal 2 cm of the colon were fixed in 4% PFA in PBS for 24 h, then transferred to a 20% sucrose solution for 24–48 h. The fixed distal colon was then cut into three equally large parts and embedded in OTC media, and flash frozen with liquid nitrogen. Sections of 7 μm thickness were cut and rinsed twice in PBS and once in PBS/2%BSA. Sections were stained with a PBS/2%BSA solution containing DAPI (Sigma-Aldrich, USA, 0.01 mg/mL final concentration) and Phalloidin-Atto 647N (Sigma-Aldrich, USA, 0.04 nmol/mL final concentration) for 30min. Sections were rinsed again twice in PBS and one time in PBS/2%BSA and mounted under Vectashield (Vector laboratories, USA). Images were acquired using a Zeiss LSM710 Laser scanning microscope. In experiment with C. rodentium without a fluorescent plasmid, the bacteria were stained with an anti-E. coli O152 antiserum. Cryosections were washed twice in PBS and once in PBS/2%BSA and stained with a primary antibody (Rabbit anti E. coli O152 Antibody, Abcam, 1:400) for 1h. After washing with PBS the sections were stained with a secondary antibody (Alexa Fluor 488 AffiniPure goat anti-rabbit IgG, LucernaChem, USA, 1:400), DAPI (Sigma-Aldrich, USA, 0.01 mg/mL final concentration), and Phalloidin-Atto 647N (Sigma-Aldrich, USA, 0.04 nmol/mL final concentration) for 1h. Sections were rinsed again twice in PBS and one time in PBS/2%BSA and mounted under Vectashield (Vector laboratories, USA). Images were acquired using a Zeiss LSM710 Laser scanning microscope.
Fluorescently labelled distal colon sections were used to quantify microcolonies (see section “fluorescent microscopy”). For each mouse, three different horizontal sections of the distal colon were analyzed. Profiles of microcolonies were counted manually, a microcolony being defined as a cluster of bacteria on the epithelial surface interspaced by either uninfected epithelium from the next microcolony or by a microcolony of a different color. In order to analyze the percentage of infected surface, the whole horizontal section was imaged as a tile scan using either a Zeiss LSM710 Laser scanning microscope or a Zeiss Observer standard fluorescence microscope. The image was then sliced to acquire the original individual images that composed the tile (approx. 100 images per horizontal section of colon). Sections containing parts of the epithelial surface were selected. Of this selection every fifth image was chosen for analysis, choosing the first picture to be analyzed randomly (random number between 1 and 5 was drawn). This selection yielded in approx. 20 images per mouse, originating from all three original sections. Images were then analyzed using the STEPanalyzer stereology tool [18]. As horizontal sections were chosen, and no vertical design was applied, the stereological analysis is only valid in relative numbers, not absolute.
The same colon tissue as used for fluorescent imaging was processed for histopathology evaluation. The distal 2 cm of the colon were fixed in 4% PFA in PBS for 24 h, then transferred to a 20% sucrose solution for 24–48 h. The fixed distal colon was then cut into three equally large parts and embedded in OTC media, and flash frozen with liquid nitrogen. Sections of 7 μm thickness were cut and stained with hematoxylin and eosin. Sections were scored by an expert pathologist in a blinded manner: Epithelial hyperplasia (scored based on percentage above the height of the control, where 0 = no change (up to 0.2 mm); 1 = 1–50%; 2 = 51–100%; 3 = > 100%); epithelial integrity (0 = no pathological changes detectable, 1 = epithelial desquamation, 2 = erosion of the epithelial surface (gaps of 1 to 10 epithelial cells/lesion), 3 = epithelial ulceration (gaps of > 10 epithelial cells/ lesion); inflammatory cell infiltrate (0 = <5 polymorphonuclear granulocytes (PMN)/ high-power field, 1 = 5 to 20 PMN/high-power field, 2 = 21 to 60 PMN/high-power field, 3 = > 60 PMN/high-power field); loss of goblet cells (0 = >28 goblet cells/high-power field, 1 = 11–28 goblet cells/high-power field, 2 = 1 to 10 goblet cells/high-power field, 3 = < 1 goblet cells/high-power field); and submucosal edema (0 = no pathological changes, 1 = mild edema, 2 = moderate edema, 3 = profound edema); The reported combined score of maximal 15 is the sum of the individual scores. This scoring system was adapted from references [42] and [43].
Fecal lipocalin-2 was measured using the solid phase Sandwich ELISA Mouse Lipocalin-2/NGAL DuoSet (R&D Systems, USA). The assay was performed according to the manufacturer’s instructions, except for the horseradish peroxidase, HRP-SA which was used from Biolegend (USA). Fecal pellets from mice were dissolved in 0.5 mL sterile PBS and debris removed by centrifugation (microfuge 7000 rpm, 5min). The supernatant was serially diluted in triplicates (1:3 dilution steps, 6 or 12 dilutions). The standard was serially diluted in duplicates (1:3 dilution steps, 12 dilutions). The OD405 was measured using a Thermomax microplate reader (Molecular Devices, USA) Data was analyzed using graph pad Prism program. A four parameter dose-response curve was fitted, using equal hill slopes for all samples on the same plate. Samples with low OD405 values resulting in a curve fit with R2 values below 0.95 were defined below detection limit. By comparison with the standard, the EC50 was used to calculate the absolute amount of Lipocalin-2 per sample.
For validation of the method on in-vitro grown bacteria (S2C Fig), an overnight culture was prepared in LB and grown for 16 h at 37°C, 150 rpm. The culture was then diluted 1:100 in either LB and grown at 37°C, 150 rpm or in DMEM (Gibco, Netherlands) and grown at 37°C, 5% CO2 without shaking. To determine density, at the time of sampling the OD600 was determined. Samples were removed at different time points, volumes adjusted to yield approx. 1x109 bacteria (OD600 = 1). Samples were spun down (4500 g, 5 min), and resuspended in RNAprotect reagent (Qiagen, Germany). After 5 min incubation the bacterial suspension was spun again and the pellet resuspended in 200 μL lysis buffer (TE buffer containing 10 mg/mL Lysozyme, Sigma-Aldrich, Canada). The bacteria were incubated at room temperature for 10 min. RNA was isolated using the RNeasy mini kit (Qiagen, Germany), including the on column DNase digestion step. Concentration of RNA was measured using Nanodrop (ThermoScientific). Quality of the samples was tested with the RNA 6000 Nano kit (Agilent, USA). Colon contents were collected from mice and suspended in 5 mL sterile PBS. Large debris were sedimented for 10 sec and the supernatant transferred to a new tube. This solution containing the bacteria was spun down (4700 g, 7 min) and the pellet resuspended in 0.5 mL RNAprotect reagent prior to RNA isolation. Quantitative PCR (qPCR) was performed in duplicates or triplicates in a volume of 20 μL containing 600 ng RNA using the TaqMan RNA-to-CT 1-Step kit (Applied Biosystems, USA) and the Quant Studio 7 Flex System (ThermoFisher, USA). Following primers and probes were used: Ler-F-primer: GAG CAG GAG ATT CAA ACT GTA A; Ler-R-primer: CGT CTT CAT TAC GGT AGT ATA CC; ler-probe: CGG CGA GCA AGA GCA CCA TCA (modifications: 6-carboxyfluorescein (FAM) and Black hole quencher (BHQ-1)); rpoD-F-primer: AAG CGA AAG TCC TGC GTA TG, rpoD-R-primer: GCT TCG ATC TGA CGG ATA CG rpoD-probe: CGA TAT GAA CAC CGA CCA CAC GCT G (modifications: Yakima yellow (YYE) and BHQ-1). All probes and primers were synthesized and tested by Microsynth (Switzerland). The following program was run: 15 min 48°C, 10 min 95°C followed by 50 cycles of 15 s 95°C and 1 min 60°C. Fluorescence was measured for each cycle. Ct values for both genes were averaged and the mean Ct of ler was subtracted from the mean Ct value of rpoD to normalize.
Colon tissue was emptied from all fecal matter, washed in sterile PBS and immediately stored in RNAlater (Qiagen, Germany). RNA was extracted from the tissue using the RNeasy kit (Qiagen, Germany) according to the manufacturer’s instructions. RNA quality was verified with the Agilent Bioanalyzer system using RNA micro chips (Agilent Technologies, USA). Using 400 ng RNA, cDNA was prepared using the RT2 Easy First Strand Kit (Qiagen) following manufacture`s protocol.
To measure expression level of 12 genes (Reg3g, S100a8, Cxcl9, Ncr1 (NKp46), Cxcl1, Ifng, Cx3cl1, Il17a, Ccl22, Nos2 (iNOS), Bcl6, Foxp3) a custom made plate was used (Qiagen). RT2 SYBR Green Master Mix (87.5 μL, 2x, Qiagen), cDNA (13.7 μL) and nuclease-free water (73.8 μL, Qiagen) were mixed and 10 μl added to each well. Plates were run on QuantStudio 7 Flex Real-Time PCR System (ThermoFisher, USA) according to manufacturer’s protocol. Fluorescence was measured for each cycle. Ct values for each gene were subtracted from the mean Ct value of the housekeeping genes to normalize (Actb). The upper CT limit was fixed to 35 cycles.
Lamina propria lymphocytes: Cecum and colon was removed, longitudinally opened and incubated in HBSS medium supplemented with DTT (308 mg/L), EDTA (0.0005 M), glucose (1 g/L), horse serum (2%), HEPES (0.01 M), and NaHCO3 (2.1 g/L) at 37°C for 15 min on a shaker for 2–3 rounds. Tissue was washed twice with PBS and digested in 0.5 mg/mL collagenase at 37°C on a shaker for 40 min. Cells were filtered and loaded onto a 80%-40% Percoll-gradient. Interphase was collected and spun down (8 min, 800 X g, 4°C). Cells were resuspended in 300 μL PBS/ 2%BSA and Fc-blocked (Anti CD16/32 AB, clone 93, Biolegend) for 10 min. Staining with antibodies (Ly6G (FITC labeled, clone 1A8), MHCII (PE labeled, clone M5/114.15.2), CD45 (PerCP labeled, clone 30-F11), CD11b (Pac. Blue labeled, clone M1/70), Ly6C (Alexa Fluor 700 labeled, clone HK1.4), CD64 (BV711 labeled, clone X54-5/7.1) all from Biolegend) was performed for 1 h on ice in a volume of 100 μL. Cells were washed and fixed with fixation buffer (Lysing Solution BDBioscience) for 10 min at RT. Cells were washed and resuspended for acquisition on a SORP LSRII flow cytometer (Becton Dickinson). Data was analyzed using FlowJo software (Treestar, USA). Population frequencies were calculated as percentage of CD45+ population. The gating strategy is shown in S4 Fig.
Blood was collected from the vena cava using a 18G syringe precoated with 7.5% EDTA (Sigma). Blood (40 μL) was distributed in 96-well U-bottom shaped microplates. Cells were counted with a VetABC animal blood counter (Medical Solution) and stained in 20 μl antibody mix diluted in FACS buffer directly added to blood for 1h at 4°C (Ly6G (FITC labeled, clone 1A8), CD45 (PerCP labeled, clone 30-F11), CD115 (APC labeled, clone AFS98), CD11b (Pac. Blue labeled, clone M1/70), Ly6C (Alexa Fluor 700 labeled, clone HK1.4) all from Biolegend). After washing two times, blood cell lysis was performed using FACS Lysing Solution (BDBioscience) for 10 min at room temperature. Cells were washed and acquired on a SORP LSRII flow cytometer (Becton Dickinson). Data was analyzed using FlowJo software (Treestar, USA). The gating strategy is shown in S5 Fig.
Distal colon tissue was washed in cold PBS and directly frozen in 500 μL of PBS containing 0.5% BSA, 0.5% Tergitol and protease inhibitor (Sigma-Aldrich). Tissue was lysed using a steel bead and a Qiagen tissue lyser (10 min, full speed) and centrifuged (20 min, 5000 g, 4°C). Supernatant was used undiluted for bioplex assay (Biorad). Concentration of the following cytokines was analyzed using the standard manufacturer’s protocol: GM-CSF, KC, IFN-g, IL-1b, IL-6, IL-10, IL-7a, IL-22, IL-18, IL-23p19, IL-33, MCP-1 (CCL2), and MIG (CXCL9). Samples were acquired using the Bio-Plex 3D instrument (Biorad) and the xPONENT Software (Luminex). Data was analyzed using the Bioplex Manager 6.1 (Biorad).
Statistical analysis was performed using the Prism software (GraphPad, LaJolla, CA, USA). Details for statistical tests used are given in the figure legends.
|
10.1371/journal.ppat.1002583 | Preferential Entry of Botulinum Neurotoxin A Hc Domain through Intestinal Crypt Cells and Targeting to Cholinergic Neurons of the Mouse Intestine | Botulism, characterized by flaccid paralysis, commonly results from botulinum neurotoxin (BoNT) absorption across the epithelial barrier from the digestive tract and then dissemination through the blood circulation to target autonomic and motor nerve terminals. The trafficking pathway of BoNT/A passage through the intestinal barrier is not yet fully understood. We report that intralumenal administration of purified BoNT/A into mouse ileum segment impaired spontaneous muscle contractions and abolished the smooth muscle contractions evoked by electric field stimulation. Entry of BoNT/A into the mouse upper small intestine was monitored with fluorescent HcA (half C-terminal domain of heavy chain) which interacts with cell surface receptor(s). We show that HcA preferentially recognizes a subset of neuroendocrine intestinal crypt cells, which probably represent the entry site of the toxin through the intestinal barrier, then targets specific neurons in the submucosa and later (90–120 min) in the musculosa. HcA mainly binds to certain cholinergic neurons of both submucosal and myenteric plexuses, but also recognizes, although to a lower extent, other neuronal cells including glutamatergic and serotoninergic neurons in the submucosa. Intestinal cholinergic neuron targeting by HcA could account for the inhibition of intestinal peristaltism and secretion observed in botulism, but the consequences of the targeting to non-cholinergic neurons remains to be determined.
| Botulism is a severe and often fatal disease in man and animals characterized by flaccid paralysis. Clostridium botulinum produces a potent neurotoxin (botulinum neurotoxin) responsible for all the symptoms of botulism. Botulism is most often acquired by ingesting preformed botulinum neurotoxin in contaminated food or after intestinal colonization by C. botulinum under certain circumstances, such as in infant botulism, and toxin production in the intestine. The first step of the disease consists in the passage of the botulinum neurotoxin through the intestinal barrier, which is still poorly understood. We investigated the trafficking of the botulinum neurotoxin in a mouse intestinal loop model, using fluorescent HcA (half C-terminal domain of the heavy chain). We observed that HcA preferentially recognizes neuroendocrine intestinal crypt cells, which likely represent the entry site of the toxin through the intestinal barrier, then targets specific neurons, mainly cholinergic neurons, in the submucosa, and later (90–120 min) in the musculosa leading to local paralytic effects such as inhibition of intestinal peristaltism. These results represent an important advance in the understanding of the initial steps of botulism intoxication and can be the basis for the development of new specific countermeasures against botulism.
| Botulinum neurotoxins (BoNTs) are responsible for a severe nervous disease in man and animals known as botulism, characterized by skeletal muscle flaccid paralysis and respiratory arrest, resulting from inhibition of acetylcholine (ACh) release in peripheral cholinergic nerve terminals. BoNTs are produced by Clostridium botulinum as single chain proteins (ap. 150 kDa), which are divided into 7 toxinotypes (A to G) according to their immunogenic properties. The toxins are exported outside the bacteria and are proteolytically cleaved into a heavy chain (H; ap. 100 kDa) and a light chain (L; ap. 50 kDa), which remain linked by a disulfide bridge. The di-chain molecule constitutes the active neurotoxin. The half C-terminal domain of the H-chain (Hc) is involved in binding to specific receptors on target neuronal cells and in driving the toxin entry pathway into cells, whereas the N-terminal part permits the translocation of the L chain into the cytosol. The L chain catalyzes a zinc-dependent proteolysis of one or two of the three proteins of the SNARE complex, which play an essential role in evoking neurotransmitter exocytosis. The BoNT/A L-chain cleaves the synaptosomal associated protein SNAP25 at the neuromuscular junction [1]–[4]. The highly specific binding of BoNTs to target nerve endings involves protein and ganglioside receptors that localize at the neuronal plasma membrane [5]. Gangliosides of GD1b and GT1b series are involved in binding and functional entry into cells of BoNT/A and BoNT/B [6]–[9]. The protein receptors on neuronal cells have been identified as synaptotagmin I and II for both BoNT/B and BoNT/G, and synaptic vesicle protein SV2 (isoforms A, B and C) for BoNT/A [8], [10]–[14], BoNT/E [15], and BoNT/F [16], [17]. SV2C is the preferred BoNT/A neuronal receptor [14], whereas BoNT/E recognizes glycosylated SV2A and SV2B [15]. BoNT/D also uses SV2 proteins as receptor in association with gangliosides for its entry into neuronal cells, but binds to SV2 via a distinct mechanism than BoNT/A and BoNT/E [18]. In addition, SV2A and SV2B have also been evidenced to mediate the entry of tetanus toxin (TeNT) into the central target neurons including hippocampal and spinal cord neurons [19].
Botulism usually results from the ingestion of preformed neurotoxin in contaminated food, or ingestion of spores or bacteria, which under certain circumstances, may colonize the gut and produce the neurotoxin in situ [20]. In either case, BoNT escapes the gastro-intestinal tract to reach the target cholinergic nerve endings, possibly through the blood and lymph circulation [21]. Indeed, previous observations have shown that after oral administration of BoNT in experimental animals, the toxin enters the blood and lymph circulation. The upper small intestine was found to be the primary site of absorption [22]–[25], but BoNT can also be absorbed from the stomach [21]. Penetration of BoNT through an epithelial cell barrier and its subsequent migration to cholinergic nerve endings are the essential first steps of botulinum intoxication. In in vitro models, BoNTs have been found to bind to polarized epithelial cells and to undergo receptor-mediated endocytosis and transcytosis from apical to basolateral sides [24], [26]–[30]. However, little is known about the precise pathway of BoNT migration from the intestinal lumen to the target nerve endings.
The digestive tract contains its own independent nervous system, the enteric nervous system (ENS), which is as complex as the central nervous system, and it is also referred as the “brain of the gut”. ENS controls and coordinates motility, exocrine and endocrine secretions, and blood microcirculation of the gastrointestinal tract. Nerve cell bodies of ENS are clustered into small ganglia which are organized in two major plexuses: the myenteric plexus between the longitudinal and circular muscle layers, and the submucosal plexus associated with the mucosal epithelium between the circular muscles and the muscularis mucosa. Ganglia also contain glial cells and their extensions. ENS neurons can be classified as afferent sensory neurons, interneurons, and motor neurons, which are connected to the central autonomic nervous system through both sensory and motor pathways. More than 20 types of neurotransmitters have been identified in ENS, and most enteric neurons may produce and release several of them. However, neurotransmitter functions have not been fully identified. Secretory and motor neurons are cholinergic, these latter also contain substance P. Myenteric neurons are connected to the cholinergic parasympathetic neurons through nicotinic, and in some areas, muscarinic receptors [31], [32]. Vasoactive intestinal protein (VIP) and serotonin are also major neurotransmitters in the regulation of normal gut function and interconnection with the central nervous system [33].
In this study, we used fluorescent Hc fragment from BoNT/A to monitor the trafficking of the toxin into the mouse intestinal mucosa. It has been previously shown that the Hc domain from TeNT, which shares similar structural organization and catalytic activity with BoNTs, is a useful tool to investigate the intracellular trafficking of the neurotoxin [34]–[36]. Although, it cannot be ruled out that a cross interplay between the BoNT domains may modify the toxin routing driven by Hc [37], recombinant HcA has been reported to retain the same structure than that of the receptor binding domain of the BoNT/A holotoxin and to enter hippocampal neurons similarly to the whole neurotoxin [16], [38]. In addition, HcA has been found to bind and transcytose through intestinal cells as well as the holotoxin [26], [39], validating its use to investigate the intestinal trafficking of the toxin.
In striated muscle, BoNT/A is known to inhibit ACh release from motor nerve terminals by cleaving the synaptosomal associated protein SNAP-25, leading to the inability of synaptic vesicles containing ACh to undergo transmitter release [for a review, see [3]]. In the gastrointestinal smooth muscle, BoNT/A also impairs cholinergic transmission by inhibiting ACh release from postganglionic cholinergic nerve endings in vitro and in vivo [40], [41]. The first aim of this study was to determine whether BoNT/A affected smooth muscle contractility when applied intra-luminally on isolated mouse ileum segments.
In preparations that were equilibrated in the standard oxygenated solution for about 30 min, spontaneous contractile responses were usually observed at a frequency rate of about 6–10 min−1 (n = 4). These spontaneous contractions had a peak force that was variable from preparation to preparation, but comprised between 0.5 and 1.2 g (n = 4). BoNT/A reduced their frequency after 2, 3 and 4 h of the intralumenal injection (Figure 1A). It is worth noting that for control preparations maintained for 4 h in the same conditions as the ones treated with BoNT/A, not only there was no reduction of the spontaneous contractions, but a small increase in their frequency (10% after 2 and 3 h) (Figure 1A). The contraction pattern changed also after exposure to BoNT/A, from very regular oscillations during the first hour to very irregular contractions after more than 2 h (Figure 1B and 1C).
Electric field stimulation evoked contractile responses that attained a peak force comprised between 2 and 7 mN (n = 3) under control conditions (Figure 1E), with little rundown of the responses when stimulations were applied every 50–60 min (Figure 1D). As shown in Figure 1D, BoNT/A reduced the electrically-evoked contractile response in a time-dependent manner. The time to decrease to 50% the evoked tension-time integral response was about 110 min, and the toxin reduced to about 90% electrically-evoked contractions within 240 min. Although it is difficult to exclude the possibility of direct muscle stimulation by the applied field-stimuli, several lines of evidence indicate that this would represent no more than 10–15% of the evoked tension-time integral response in our experimental conditions. Most of this evidence comes from: (i) Data obtained with BoNT/A showing that blockade of the evoked contraction attains a maximum around 86–90% of control values, and was never complete. The remaining tension can be suspected to be due to direct muscle stimulation unaffected by BoNT/A. (ii) If direct stimulation of the muscle would occur, one would expect that tension levels would be sustained and maintained during the field stimulation, which is not the case (Figure 1E). (iii) Spontaneous contractions occurring during the falling phase of the evoked-contractile response were not enhanced in amplitude, which is consistent with the low influence of direct muscle stimulation under the experimental conditions used. Interestingly, after BoNT/A has blocked the evoked response by field stimulation, the addition of carbachol (20 µM) or ACh (data not shown) to the standard solution evoked a contractile response (Figure 1F). These results suggest that under the conditions used BoNT/A is able to exert an action on cholinergic terminals that leads to a blockade of the contractile responses evoked by electric field stimulation, while spontaneous myogenic contractions were reduced in frequency, but not completely abolished. The fact that carbachol could induce contractile activity after BoNT/A-induced blockade of contraction evoked by electric field stimulation, strongly suggests that the sensitivity of ACh receptors (muscarinic and nicotinic) is not affected by the toxin (Figure 1F).
In a previous report, we have found that BoNT/A transcytosis through intestinal cell monolayers grown on filters is mediated by SV2C or, at least, an immunologically related protein [30]. To address whether SV2C might be a functional receptor in the ex vivo intestinal tract model, BoNT/A was preincubated with the intravesicular domain segment L4 of SV2C prior to injection into ligated intestinal loop. As shown in Figure 1A, preincubation with SV2C/L4 significantly prevented the BoNT/A inhibitory effects on spontaneous intestinal smooth muscle contractions, suggesting a partially SV2C-dependent BoNT/A uptake through the intestinal mucosa. However, these results do not rule out that SV2C/L4 also passed, independently or associated with BoNT/A, through the intestinal barrier and impaired BoNT/A uptake by nerve terminals.
We first investigated the potential binding sites for HcA in mouse intestine mucosa and submucosa, as well as in the musculosa. For that, cryosections of mouse small intestine, fixed on glass slides, were incubated with fluorescent HcA and analyzed by confocal microscopy. Since BoNT has been reported to be preferentially absorbed from the upper small intestine [23], [24], sections from ileum were analyzed. Only a faint staining was observed in brush border of enterocytes along intestinal villi (Figure 2A). However, a strong HcA binding was observed on intestinal crypt localized at the bottom of villi. Paneth cells, which are characterized by their numerous secretory-granule content, are spatially restricted to intestinal crypts and were used as a marker of these regions. Staining of Paneth cells with the TRITC-labeled lectin Urex europaeus agglutinin type 1 (UEA1) [42] did not significantly colocalize with HcA (Figure 2B). This may indicate that some cells from intestinal crypts, distinct from Paneth cells, exhibit preferential binding sites for HcA. Moreover, small cells scattered along the villi were stained with HcA (Figure 2A) and some of them co-stained with UEA1 (Figure 2B). These UEA1 positive cells likely correspond to goblet cells [43].
It is worth noting that HcA stained neuronal cell bodies and neuronal structures in the submucosa and musculosa. However, only a low proportion of neuronal structures were recognized and labeled by HcA, as revealed by immunolabeling neurofilaments and co-staining with the fluorescent toxin fragment (Figure 2C). Hence, the binding domain of BoNT/A potentially targets epithelial cells in intestinal crypts, and neuronal structures of intestinal plexuses. Note that anti-neurofilament antibodies are not specific of the nerve endings, where BoNT/A is assumed to bind, but recognize neuronal structures all along the neuronal cells. This probably accounts for the irregular co-staining between HcA and anti-neurofilament antibodies. In addition, the irregular pattern of HcA staining might also be related to the variability in orientation and size of the cryosections. Moreover, cryosections might artificially expose certain antigens which are buried in intact tissues. Thus, immunostaining pattern in cryosections has to be considered with caution and confirmed in ex vivo experiments with intact tissues as shown in the following figures.
To analyze HcA entry into the intestinal mucosa and submucosa, ex vivo experiments were performed, as previously described with the whole toxin. For this, excised small intestine loops were washed, ligated at both extremities, and incubated in oxygenated Krebs-Ringer solution at 37°C. Fluorescent HcA was inoculated into the intestinal lumen, and at various time intervals, intestinal loops were washed, fixed and processed for dissection, and immunostaining. A competition assay between HcA-Cy3 and native BoNT/A injected into an ileum loop and monitored by fluorescence analysis of the intestinal mucosa, supported that fluorescent HcA follows the same entry pathway than native BoNT/A (Figure 3A). Fluorescent HcA entered similarly ileum, duodenum or jejunum segments, as tested by mucosal fluorescence analysis (Figure 3A). After 30–60 min incubation, labeled HcA was detected inside the lumen of intestinal crypts, and in some crypt cells, but not or with a low intensity in enterocytes or other cells in the villi (Figure 3 B and C). HcA also labeled long cell extensions in the submucosa, which correspond to nerve fibers or neuronal extensions (Figure 3D; arrow head) since they were co-stained with antibodies against neurofilaments (not shown). Longer incubation periods (90–120 min) permitted to visualize HcA staining of long filaments in the musculosa, (Figure 3E), which were identified as nerve fibers from the myenteric plexus (see below), but with a weaker intensity. This suggests a progressive entry of HcA from the intestinal lumen through the mucosa, preferentially through intestinal crypts, to certain neuronal cell and extensions in the submucosa, and then in the musculosa.
To identify the intestinal crypt cells targeted by HcA, fluorescent HcA was injected into the lumen of an intestinal loop. After an incubation of 15–30 min, the intestinal mucosa was prepared for microscopy observation. Only few numbers (1 or 2) of small cells from each intestinal crypt were stained with HcA (Figure 4). Cells stained with HcA were distinct from Paneth cells, which were easily detectable by their numerous granules (Figure 4), and by their staining with UEA1 (not shown). Chromogranin-A antibodies, a common marker of neuroendocrine cells in the gastrointestinal tract [44], colocalized with HcA (Figure 4A). All cells stained with HcA were also stained with chromogranin-A antibodies, indicating that HcA specifically entered neuroendocrine cells from intestinal crypts. However, not all chromogranin-A positive cells were stained with HcA, but only about 80%. Serotonin-producing cells, which are abundant in the ENS, were investigated for their colocalization with HcA. In the intestinal crypts, all the cells stained with HcA were also immunolabeled with serotonin antibodies (Figure 4B and C). Interestingly, HcA accumulated in the basal pole of neuroendocrine cells, which is wider than the apical pole exposed to the intestinal crypt lumen. This strongly supports that HcA uses neuroendocrine cells, mostly serotonin-producing cells, from intestinal crypts for its transport through the intestinal mucosa.
In addition, we checked whether BoNT/A can be transcytosed through the mouse neuroendocrine intestinal cell line STC-1 [45]. As shown in Figure 5A, the passage of biologically active BoNT/A was monitored from apical to basolateral side of STC-1 cell monolayers. The transcytotic passage of BoNT/A through STC-1 cells was not statistically different from that through Caco-2 enterocytes, but it was lower than through the mouse intestinal crypt cell line m-ICcl2 as shown in Figure 5A (p<0.05) and [30]. It is noteworthy that the passage yield through STC-1 cells was more difficult to assess (high standard deviation values), since these cells do not form tight junctions as epithelial cells. Since epithelial cells such as Caco-2 and HT29 cells express at the cell surface and secrete from apical and basolateral sides several types of proteases [46]–[48], we tested whether these proteases degrade BoNT/A, thus impairing or decreasing the transcytosis level. As shown in Figure 5A, a 2 to 4 fold higher level of BoNT/A transcytosis was observed in Caco-2 and m-ICcl2 cells incubated with a cocktail of anti-proteases. However, even in the presence of anti-proteases, BoNT/A transport was more efficient (20-fold) in m-ICcl2 than in Caco-2 cell monolayers. Thus, the decreased BoNT/A transcytosis through Caco-2 cells is not likely due to a higher protease degradation of BoNT/A before and/or after transport. STC-1 cells possibly also secrete proteases, and a higher level of BoNT/A transcytosis through this cell type might be expected. However, since STC-1 cells do not form tightly organized cell monolayers, the results of experiments with anti-proteases were inconclusive.
As we have previously found that m-ICcl2 express SV2C or an imunologically related protein as a putative BoNT/A receptor [30], we investigated the presence of SV2 proteins and chromogranin A in STC-1 and intestinal cells by Western blotting (Figure 5B). Chromogranin A was strongly expressed in STC-1 confirming its neuroendocrine type, and to a lower extent in m-ICcl2 and even less in Caco-2 cells, but not in Vero cells used as negative control. Antibodies against SV2A and SV2B showed no protein related to the expected size of SV2 (82 kDa) in intestinal and STC-1 cells. In contrast, specific SV2C antibodies directed against the N-terminal part or the intravesicular loop L4, which is assumed to be the receptor binding domain of BoNT/A [11], [14], recognized a protein with the expected size in the intestinal cells and STC-1, but not in Vero cells (Figure 5B). The specificity of SV2 antibodies is shown in Figure S1. Next, we investigated whether SV2C was involved in the entry of HcA into m-ICcl2 and STC-1 cells by a competition assay between fluorescent HcA and SV2C/L4. Cells grown on glass cover slides were exposed to HcA-Cy3 or a combination of fluorescent Hc with a 10-fold higher molar concentration of SV2C/L4 for 10 min at 37°C and then processed for microscopic observation. As shown in Figure 5C, HcA entered into m-ICcl2 and STC-1 cells, and SV2C-L4 greatly impaired the entry of HcA into both cell types by 97 and 94%, respectively, as determined by counting the number of fluorescent HcA patches per µm2 of cell area (Figure 5D), supporting the view that SV2C participates in the entry mechanism of HcA into cells.
After 30–60 min incubation of fluorescent HcA into an intestinal loop lumen, HcA was detected in the intestinal submucosa, where it stained certain neuronal structures (Figure 6A and data not shown). Antibodies against neurofilaments allowed visualizing a complex and abundant network of neuronal cell bodies and neuronal extensions in the submucosal plexus immediately underneath intestinal villi and crypts. Only some of these neuronal structures were stained with HcA (Figure 6A).
BoNTs are well known to interact with cholinergic neurons and to specifically block spontaneous and evoked quantal acetylcholine release [4], [49]. However, it has been shown that BoNT/A and BoNT/E are also able to enter other neuronal cell types such as glutamatergic and gamma-aminobutyric acid (GABA)-ergic neurons, as well as astrocytes [50], [51]. Since ENS contains a large variety of neuronal cell types, we investigated the most representative types as putative targets of HcA in the mouse small intestine.
Cholinergic neurons were monitored by immunostaining with antibodies for choline acetyltransferase (ChAT). ChAT-immunoreactive neurons are abundant (about 55%) in the submucosal plexus, where they are involved in various gut functions including the control of evoked anion secretion by the jejunal and ileal epithelium, and they also interact with Peyer's patch follicles [52], [53]. Most of ChAT-immunoreactive nerve terminals from the intestinal submucosa were labeled with fluorescent HcA (Figure 6B and G).
Vasoactive intestinal peptide (VIP)-immunoreactive neurons are known to be located in jejunum and ileum submucosal plexus, as well as in other organs. VIP modulates several basic functions including blood flow, smooth muscle relaxation, and exocrine secretion [54]. VIP-immunoreactive neurons are estimated to represent about 45% of neurons from the submucosal plexus [53], [55]. Numerous cells were immunostained with anti-VIP antibodies in mouse intestinal submucosa, but a colocalization between VIP immunoreactivity and HcA staining was only observed in a few of them (less than 3%) (Figure 6C and G). Note that some of the filaments and cell bodies stained with anti-VIP antibodies did not colocalize with neurofilament staining, and may probably represent glial structures.
Glutamatergic neurons, which are the major neurons from the central nervous system involved in excitatory responses, are also present in ENS. Glutamate receptors have been detected in enteric neurons and glutamatergic enteric neurons where they have been found to mediate excitatory synaptic transmission, whereas only a subset of them are involved in sensory responses [56]. In mouse intestinal submucosa, only a low number of glutamatergic neurons (less than 1%), as evidenced by anti-glutamate antibodies, colocalized with HcA (Figure 6D and G).
Serotonin is also an important neurotransmitter in ENS, where it is involved in the control of motility, secretion and sensory functions. Serotonin is produced by 2 to 20% of all enteric neurons [33]. In our analysis only a few neuronal cell extensions were stained with anti-serotonin antibodies in the submucosal plexus, and some of them were also labeled with HcA, as shown in Figure 6E.
Glial cells were neither immunostained with anti-neurofilament antibodies nor with HcA, but exhibited a clear immunolabelling with GFAP antibodies (data not shown). These results support the view that HcA binding is mostly specific of nerve endings in the intestinal submucosa.
Also, we investigated the intestine musculosa, which is the predicted target tissue of BoNT/A for its inhibitory activity on intestinal motility. No significant HcA staining was observed in the musculosa 30 or 60 min after incubation with the fluorescent probe in the intestinal loop, and only a few cell bodies or extensions were stained after a longer incubation period (90–120 min) in our experimental conditions. Cell extensions stained with HcA colocalized with neurofilament staining (Figure 7). However, only some of the nerve endings were labeled with HcA. Almost all cell bodies labeled with HcA exhibited ChAT-immunoreactivity. Similar results of colocalization with ChAT-immunoreactive neurons were obtained using full length BoNT/A injected into the intestinal loop lumen and detected with anti-HcA antibodies (Figure S2). This is consistent with the fact that a large majority of neurons in the myenteric plexus are immunoreactive for ChAT, albeit many of them produce additional neurotransmitters [55]. No significant colocalization was observed between HcA staining and glutamate- or serotonin-producing neurons in the myenteric plexus (data not shown).
The protein receptor of BoNT/A on neuronal cells has been identified as SV2. Among the three SV2 isoforms, SV2C shows the highest affinity to BoNT/A in vitro, whereas BoNT/A binds to SV2A and SV2B with a lower strength [11], [14], [57]. SV2A is present in almost all neurons whatever their neurotransmitter type is, while SV2B shows a more restricted distribution. SV2C is reported to be present in a subset of neurons [58]. However, SV2 proteins are expressed not only in neuronal cells, but also in other cell types such as neuroendocrine cells, in particular in the gastrointestinal tract [59]. We investigated the distribution of SV2 proteins in mouse intestinal mucosa with our in toto tissue model. Numerous cell extensions in the submucosal plexus were stained with anti-SV2C antibodies (Figure 6F), whereas no specific staining was observed with anti-SV2B antibodies, and only a diffuse staining in certain crypt cells and cell extensions in the submucosa was evidenced with anti-SV2A antibodies (data not shown). The mouse intestinal crypt cell line m-ICcl2 was found to express SV2C at a higher level than in enterocyte-type cell lines (Figure 5B and [30]). However, in our ex vivo conditions intestinal crypt cells were not, or only weakly immunoreactive with anti SV2C antibodies. This does not preclude that a subset of crypt cells express a significant level of SV2C that was not detected in our conditions. Most of SV2C-immunoreative filaments in the submucosa, were also labeled with anti-neurofilament antibodies, indicating that SV2C is widely distributed in neuronal cells and neuronal extensions from the small intestine. However, HcA signal was observed in only some neuronal endings bearing SV2C, but not along the neuronal extensions stained with anti-SV2C antibodies (Figure 6F). Only a low proportion of SV2C-immunoreactive cells were labeled with HcA (Figure 6G). It is worth noting that a distinct network of thin filaments around small blood vessels was stained with anti-SV2C, and only weakly with anti-neurofilament antibodies. No specific binding of HcA was observed on these structures (Figure S3A and B). In the intestinal submucosa and musculosa, large cells with wide cellular bodies and short extensions were stained with anti-SV2C antibodies, but not with anti-neurofilament antibodies. Co-labelling with anti-GFAP indicated that they were glial cells (Figure S3C). However, HcA, as reported above, was not found to label glial cells.
The standard scheme of botulism intoxication includes BoNT transit through the digestive tract, passage across the intestinal epithelial barrier and subsequent delivery to the blood circulation and dissemination to the target motor nerve endings. Indeed, BoNT has been found to be absorbed preferentially from the upper small intestine, but also from the stomach in experimental rodents, and to be delivered in the blood and lymph circulation [21], [23], [24], [60]. The aim of this study was to identify the entry pathway and target cells of BoNT/A in the mouse intestinal wall. For that, we checked the activity of BoNT/A in mouse intestine following intralumenal administration and we used the fluorescent Hc domain, which is the functional binding domain of BoNT, to monitor the trafficking of BoNT/A.
First, we tested whether BoNT/A injected into intestinal lumen was able to enter intestinal mucosa and to induce local effects on the intestine. In vitro studies have already shown that BoNT/A reduces cholinergic transmission in gastrointestinal smooth muscles as well as pylori and Oddi sphincter muscles by inhibiting ACh release [40], [41], [61]–[63]. In our experimental conditions, BoNT/A passed through the epithelial intestinal barrier and by diffusion through the extracellular space, locally targeted intestinal neurons independently of the blood circulation. BoNT/A reduced the frequency of spontaneous contractions of small intestine and inhibited the contractile response evoked by electric field stimulation within 2–4 h after intralumenal administration. Since carbachol was still able to stimulate muscle contraction after BoNT/A treatment, this supports a toxin-dependent inhibition of ACh release. To the best of our knowledge, this is the first report demonstrating a local intestinal effect of botulism after intralumenal administration of purified BoNT/A. Constipation is often (about 70%), but not always, associated with food-borne botulism and its participation to the progression of the disease is unknown [64], [65]. However, constipation is a major and early symptom of botulism resulting from an intestinal colonization by C. botulinum such as during infant botulism [66], [67]. This digestive symptom might result from a local effect of BoNT after crossing the intestinal barrier instead of toxin dissemination through the general circulation. BoNT locally synthesized in the intestine is possibly absorbed in a higher local concentration able to induce an intestinal muscle paralysis, than toxin orally ingested which disseminates more broadly through the digestive tract. Interestingly, BoNT/A-dependent inhibition of evoked smooth muscle contraction was significantly prevented by preincubation with the intravesicular domain of SV2C. This indicates that a protein related to SV2C/L4 might impair the BoNT/A passage through the intestinal barrier and/or toxin uptake by the underlying nerve terminals. We have previously found that SV2C, or a related protein, is part of BoNT/A receptor mediating toxin transcytosis through cultured intestinal cell monolayers [68]. In addition, SV2C/L4, which is expressed by intestinal and neuroendocrine STC-1 cells (Figure 5B), significantly prevented HcA entry into the intestinal crypt m-ICcl2 and STC-1 cells (Figure 5C, D). Taken together, these data suggest that SV2C, or a related protein, facilitates BoNT/A uptake through the mouse intestinal barrier.
BoNT/A trafficking in mouse intestinal wall was investigated with fluorescent HcA as already used [38]. First, we investigated the potential binding sites of HcA by using mouse small intestine cryosections overlaid with fluorescent probes. Thereby, HcA was found to bind only to certain cell types of the intestinal mucosa, preferentially from intestinal crypts, whereas enterocytes showed only a weak staining of the brush border. In contrast, it was reported that the botulinum complexes type A or type C (BoNT and associated non-toxin proteins, ANTPs), strongly bind to the epithelia cell surface and goblet cells of guinea pig small intestine. Moreover, this binding was shown to be mediated by the hemagglutinins HA1 and HA3b, which interact with distinct gangliosides and/or glycoproteins [24], [29], [60], [69] from those recognized by the neurotoxin alone [11], [14]. This certainly accounts for the differential binding between progenitor toxin and BoNT to intestinal epithelial cells.
The functions of ANTPs are still controversial. Ancillary proteins probably participate in BoNT protection from degradation inside the digestive tract in a dose-dependent manner, particularly in the stomach [21]. In addition, it is assumed that HAs are involved in the internalization of progenitor type C toxin into intestinal cells and subsequently in the small intestine [24], [29], [70], and that they disrupt the intestinal epithelial barrier facilitating toxin absorption via the paracellular route [71]–[74]. However, since BoNT/A absorption from mouse stomach or small intestine was found to occur independently of the presence of ANTPs, HAs might have not an essential role but an additional facilitating effect on BoNT/A passage through the intestinal barrier. At pH around neutrality, as found in the intestine, BoNT dissociates from ANTPs [75], and thus might be absorbed through the epithelial intestinal barrier independently of ancillary proteins. In our model of ligated mouse intestinal loops and recording of muscle contraction, BoNT/A entered the intestinal mucosa in the absence of ANTPs supporting that ANTPs are not absolutely required in the intestinal uptake of BoNT/A.
The main finding of the ex vivo intestinal loop experiments was that fluorescent HcA preferentially recognized certain crypt cells. Intestinal crypts contain stem cells which proliferate and differentiate in four main cell types including enterocytes, mucus, endocrine, and Paneth cells, which are the most abundant cell type at the bottom of intestinal crypts [76]. HcA labeling was observed in chromogranin- and serotonin-immunoreactive cells, but not in Paneth cells. It is noteworthy that HcA was detected in only some, but not all, chromogranin-immunreactive cells. This strongly suggests that HcA specifically binds to a subset of neuroendocrine cells from intestinal crypts, which possibly represent the preferential site of toxin uptake.
BoNT/A has been found to pass through intestinal cell monolayers grown on filter by a transcytotic mechanism [26], [27]. We have previously reported that the passage rate of biologically active BoNT/A was higher (about 10-fold more) through the mouse crypt cell line m-ICcl2 than through the colonic enterocyte lines Caco-2 and T84 [30]. Here, we found that the neuroendocrine intestinal cell line STC-1 also permits a transcytotic passage of BoNT/A not statistically different from that through Caco-2 cells and to a lower extent than in m-ICcl2 cells. This further supports that a subset of intestinal crypt cells constitute a preferential entry site of BoNT/A within the intestinal mucosa. However, STC-1 cells, which derive from an intestinal endocrine tumor, secrete several hormones and neuropeptides, but not serotonin [45]. Thereby, STC1 cells are not the most representative cells of the serotonin-reactive cells co-labeled with HcA, which have been identified in intestinal loop tests. To the best of our knowledge, no intestinal serotonin-secreting endocrine cell line is known at present. m-ICcl2 cells, which have been isolated from fetal mouse intestine, express various markers of intestinal cells and thus are a totipotent intestinal crypt cell line in a less differentiated state than the tumor cell line STC-1 [77]. m-ICcl2 cells show a neuroendocrine-like phenotype since they synthesize chromogranin A albeit to a lower extent than STC-1, but more significantly than Caco-2 cells (Figure 5B). In addition, m-ICcl2 cells retain transport pathways similar to those of native crypt cells. Indeed, in contrast to colonocytes such as Caco-2 and HT29 cells, m-ICcl2 cells express polymeric Ig receptors (pIgRs), which mediate the transepithelial transport of polymeric IgA and IgM [77]. It is noteworthy that only a low number of cells (1 to 2) from each intestinal crypt were labeled by HcA and are likely susceptible to transport the toxin through the intestinal barrier. This might account for the low rate of BoNT absorption from the digestive tract to the general circulation. Indeed, the transport rate of BoNT/B from rat duodenum to the lymphatic circulation has been estimated from 0.01 to 0.1% [22]. In addition, chromogranin-immunoreactive cells are distributed throughout the gastrointestinal tract, but are more predominant in the pylorus and duodenum [78]. This might support the observation that the upper small intestine is the preferential site of BoNT absorption [22]–[25].
Using the intestinal loop model, we found that neuronal cells in the submucosa and more lately in the musculosa were stained with HcA. However, not all the neuronal cells were recognized by HcA as tested by co-labeling with antibodies against neurofilament, indicating that HcA targeted specific neurons in the intestine. About half of neurons in the submucosa are ChAT-immunoreactive, and most of them produce other neurotransmitter types [53], [55]. As previously found [79], HcA stained most of the cholinergic neurons (more than 90%) of the submucosa. However, in contrast to previous findings [79], BoNT/A recognized other neuronal cells in the intestinal mucosa, albeit to a lower extent. Indeed, a low proportion of VIP-immunoreactive neurons, which are also largely present in the submucosa (about 45%) [53] as well as a low proportion of glutamatergic and serotoninergic neurons were targeted by HcA. It is noteworthy that the BoNT/A receptor, assigned to ganglioside GD1b/GT1b and SV2 protein [6], [7], [10], [14], is part of a synaptic vesicle complex, which contains additional membrane proteins including vesicular glutamate transporters (cGLY-1 and vGLUT-2) [80], supporting that BoNT/A may target glutamatergic neurons. BoNTs are known to block the release of ACh, but also that of other neurotransmitters and neuropeptides the effects of which are still poorly known [61], [81]–[86]. The BoNT/A-dependent inhibitory activity on non-ChAT neurons in the intestine may contribute to the local effects of the toxin.
In the myenteric plexus, HcA stained essentially cholinergic neurons, which are the predominant neuronal cells in this tissue (about 80%) [53]. But, only a part of ChAT-immunoreactive neurons were labeled with HcA. This might be due either to a low number of Hc molecules reaching the myenteric plexus, or to the fact that HcA recognized only a subset of ChAT-reactive neurons, which remains to be determined. Targeting of cholinergic neurons of the myenteric plexus by HcA is consistent with constipation which frequently occurs during botulism, and with the BoNT/A-dependent inhibition of evoked intestinal contractions in the ex vivo experiments.
SV2C, the protein receptor with the highest affinity to BoNT/A [10], [14], was the main isoform detected in the mouse intestine, whereas no, or a weak staining was obtained with anti-SV2B or -SV2A antibodies. Immunostaining with anti-SV2C antibodies was observed in cells with thin arborescent extensions, which were also labeled with anti-neurofilament antibodies. However, SV2C staining did not match all the cells labeled with anti-neurofilament antibodies, indicating a SV2C distribution in a broader range of cell types than neuronal cells in the intestinal mucosa. HcA did not bind all along the neuronal extensions or cell structures which are recognized by anti-SV2C antibodies, but only in discrete zones. One possibility is that HcA staining was too weak to be visualized. Alternatively, the BoNT/A high-affinity receptors, which consist of a ganglioside part and a protein part such as SV2C [11], [14]–[17], are distributed in only some restricted areas on the neuronal cell surface. This is supported by the observed colocalization between HcA and SV2C, which is restricted to only some cell areas (Figure 6F). Moreover, SV2, which is an integral protein of the synaptic vesicle membrane, has to be exposed to the extracellular compartment to be accessible to BoNT/A, as during synaptic vesicle fusion with the presynaptic membrane. Another type of thin cell extensions stained with anti-SV2C antibodies, formed a dense network all around microvessels (Figure S2). These neuronal extensions, which were only faintly stained with anti-neurofilament antibodies, are probably involved in vessel innervation [87], but were not observed to be targeted by HcA under our experimental conditions. In addition, glial cells identified by GFAP immunofluorescence were also stained with anti-SV2C antibodies but not with HcA (Figure S3C). Thereby, SV2C seems to have a broad distribution and to form BoNT/A high-affinity receptor only when associated with gangliosides in certain neuronal membrane domains.
When passed across the intestinal barrier, BoNT is delivered to the connective tissue of the submucosa, where the toxin can have access to microvessel endothelial cells. BoNT passage into the blood and lymph circulation, supposes a transcytotic transport of the toxin through the endothelial cell barrier of vessels. In ligated intestinal loop experiments, no staining of endothelial cells forming small vessels or microvessel structure was observed in the submucosa and musculosa. Under natural conditions, BoNT binding to endothelial cells is possibly low and/or transient permitting the passage of the toxin into the vessel lumen. In our experimental conditions, the blood circulation was interrupted, thus preventing HcA passage into vessels. But, the diffusion through the matrix of connective tissue can mediate the toxin trafficking until the target cells in the submucosa and musculosa. Indeed, HcA staining of neuronal cells in the submucosa after 30–60 min incubation of the probe into the intestinal lumen and later (60–120 min) in the musculosa, as well as the decrease in fluorescence intensity of HcA from the mucosa to musculosa likely reflects the progressive diffusion of the recombinant protein in the tissue extracellular matrix. It cannot be ruled out that a shorter time period may be required in natural conditions for the passage of low amounts of HcA across the intestinal epithelial cell barrier and subsequent targeting to neuronal cells. Alternatively, another mode of HcA transport might be involved. Indeed, like TeNT, which undergoes a retrograde transport into motor neurons, a transcytotic mechanism might also be supported by neuronal cells from the ENS to disseminate BoNT/A to target neuronal cells both locally and even at distance from the intestine. One possibility is that BoNT/A uses non-cholinergic neurons such as those identified with fluorescent HcA to be transported to other target neurons. Indeed, neurons from ENS are highly interconnected between them and with neurons of the central nervous system [31], [32]. Moreover, it has been recently found that BoNT/A can use a retrograde transport and transcytosis to migrate from peripheral neurons to central circuits [88], [89]. This opens novel ways of investigation to unravel the mechanism of BoNT transport from the intestinal barrier to target motoneurons, which is still poorly known.
In conclusion, in this work we have shown that BoNT/A enters the intestinal mucosa, possibly via an uptake process involving SV2C or a related protein, and impairs spontaneous and electrically-evoked muscle contractions. Following intralumenal administration, HcA preferentially recognized a subset of neuroendocrine crypt cells in the mouse upper small intestine, which likely represents the main pathway for toxin entry and passage across the epithelial cell barrier. HcA diffused progressively within the submucosa, and later reached the musculosa, where it targeted specifically some cholinergic neurons and to a lower extent glutamatergic and serotoninergic neurons. The implication of these non-cholinergic neurons in the botulinum intoxication process remains to be determined.
All experiments were performed in accordance with French and European Community guidelines for laboratory animal handling. The protocols of experiments were approved by the Pasteur Institute (Agreement of laboratory animal use n° 75-279).
The primary antibodies used recognized neurofilament (Sigma; mouse, 1∶500 dilution), Glial Fibrillary Acidic Protein (GFAP) (Sigma; rabbit, 1∶200 dilution), Choline Acetyltransferase (ChAT) (Chemicon International, Temecula, CA, USA; goat, 1∶100 dilution), VIP (Abcam; rabbit, diluted 1∶200), Serotonin (Sigma; rabbit, diluted 1∶100; and Abcam ; goat, 1/200 dilution), Glutamate (Chemicon; rabbit, diluted 1∶200), SV2A, SV2B (Synaptic systems; rabbit, diluted 1∶200), SV2C (Santa Cruz; goat, diluted 1∶200), chromogranin A (Abcam; rabbit, diluted 1∶200). For Western blotting rabbit anti-SV2C (Abcam 33892) and rabbit anti-SV2C intravesicular domain were used. SV2C/L4 domain (amino acid 454 to 579) was produced as a GST-fusion protein [30] and was used to immunize rabbits. Specific antibodies against SV2C/L4 were purified by immunoaffinity with SV2C/L4 produced as a histidine-tagged protein in pET28 vector (Novagen) and immobilized on cyanogen bromide activated Sepharose-4B (GE Healthcare).
The secondary antibodies used were: Cy5 coupled donkey anti-mouse IgG (US Biological), Cy3 coupled donkey anti-goat IgG (US Biological), Alexa488 donkey coupled anti-goat IgG (Invitrogen), Alexa488 coupled anti-rabbit IgG (Invitrogen). 4′-6′-Diamidino-2-phenylindole (DAPI, SIGMA) was incubated with secondary antibodies to stain cell nuclei.
Adult male IOPS mice (20–25 g body weight) purchased from Charles River Laboratories (L'Arbresle, France) were anesthetized by inhalation with Isoflurane (AErrane, Baxter S.A., Lessines, Belgium), and euthanized by dislocation of the cervical vertebrae, as specified by the CNRS Animal Ethics User's Committee.
BoNT/A was produced and purified as previously described [90].
Recombinant His-tag Hc fragment of BoNT/A was produced and purified from pET28b vector containing DNA encoding for HcA cloned into BamHI and SalI sites, as previously described [91]. A recombinant derivative plasmid encoding HcA with a N-terminal tag containing 4 cystein residues, as described for TeNT-Hc [92], was performed inserting the two complementary oligonucleotides P1132 (5′-TATGGCAGAGGCAGCAGCACGAGAGGCTTGTTGTCGAGAGTGTTGTGCACGAG-3′) and P1131 (5′-GATCCCTCGTGCACAACACTCTCGACAACAAGCCTCTCGTGCTGC TGCCTCTGCCA-3′) into NdeI and BamHI sites of pET28b vector. HcA with 4 Cys tag was then labeled with either maleimide Alexa 488 reagent (InVitrogen, Cergy-Pontoise, France) or maleimide Cy3 (Amersham, Les Ulis, France) according to the manufacturer's recommendations.
cDNA coding the intraluminal SV2C fragment L4 (amino acid 454 to 579) was PCR-amplified as previously described [68] and cloned into pGEX-2T. The fusion protein, GST-SV2C/L4, was produced in E. coli BL21 and purified on glutathione agarose matrix (SIGMA) equilibrated with 50 mM Tris-HCl, 300 mM NaCl, pH 7.5 and eluted with 40 mM reduced glutathione in the same buffer. Purified fusion protein was then dialyzed against phosphate balanced solution (PBS) before use.
The mouse ileum was removed and placed in an oxygenated standard Krebs-Ringer solution composed of 154 mM NaCl, 5 mM KCl, 2 mM CaCl2, 1 mM MgCl2, 11 mM glucose and 5 mM HEPES (pH 7.4). An ileum segment of about 2 cm length was extensively washed to flush out the intestinal content, and mounted in a silicone-lined chamber (4 ml volume), bathed in the standard medium. For tension measurements, one of the ends of the ileum segment was tied with silk thread, via an adjustable stainless-steel hook, to an FT03 isometric transducer (Grass Instruments, AstroMed, W. Warwick, RI, USA), and the other end was ligatured and pinned onto the silicone-coated bath via stainless-steel micro pins. BoNT/A (5 µg/ml) diluted in the Krebs-Ringer solution was injected into the ileum segment, between the two ligatures, with a micro syringe. Electric field stimulation was performed with an electrode assembly, placed along and on both sides of the length of the ileum segment, and connected to S-48 Grass stimulator. Preparations were usually stimulated with pulses of 0.15 ms duration at 30 Hz for 15 or 25 s, every hour. The resting tension was adjusted for each preparation investigated with a mobile micrometer stage (to allow incremental adjustments of ileum length) in order to obtain maximal spontaneous or evoked contractile responses, and was monitored during the whole duration of the experiment. Carbachol (Sigma, 20 µM) was added to the bath solution of ileal segment exposed to BoNT/A for 3–4 hours to check that ACh receptors were not affected by intoxication. Tension signals from the isometric transducer were amplified, collected, and digitized with the aid of a computer equipped with a DT2821 analogue to digital interface board (Data Translation, Marlboro, USA) and expressed in g or N. Data acquisition and analysis were performed with a program kindly provided by Dr. John Dempster (University of Strathclyde, Scotland). All experiments were performed at 22±0.5°C.
Segments of mouse ileum (2 cm length) were removed and immersed in oxygenated Krebs-Ringer solution. After extensive wash to flush out the luminal contents, segments were ligated on both extremities and injected with a micro syringe with fluorescent HcA (0.5 µg) diluted in 150 µl of Krebs-Ringer solution (5.5 10−8 M). The injection site was isolated with another ligature and ileal segments were incubated in oxygenated Krebs-Ringer solution for different times (30 to 120 min) before washing. Tissues were cut along the mesenteric border and pinned out with the mucosal surface facing down in a silicone-lined Petri dish. After fixation with 4% paraformaldehyde (PFA; 1 h, 22°C), tissues were washed in phosphate buffer saline (PBS) and autofluorescence quenched with 50 mM NH4Cl (30 min, 22°C).
Specimens were dissected by carefully separating the mucosa and submucosa from the muscle layers under a microscope. Whole-mount preparations of the myenteric and submucosal plexuses of the ileum were permeabilized and blocked with PBS/bovine serum albumin (BSA) 2%/Donkey Serum 10%/Triton 2% for 1 h at room temperature (RT) and incubated with primary antibody for 16 h at RT. After 3 washes of 10 min in PBS, tissue sections were incubated for 4 h at RT with the appropriate secondary antibodies diluted 1∶500 in PBS/BSA 2%/Triton 2%. After being washed in PBS (3×10 min), tissues were mounted in Mowiol (Polysciences Europe, Eppelheim, Germany) and analyzed using a Zeiss confocal laser scanning microscope and a ×63 oil immersion objective (N.A. 1.4). For quantitative analysis of co-localization of HcA with other markers, images of 512×512 pixels were taken from series of optical sections of 0.8 µm thickness. The colocalisation was analyzed with the Zeiss LSM Image Browser software, on 100 HcA-immunoreactive varicosities for each marker using the merged image from the different experiments. For analysis of BoNT/A internalization into intestinal cells, total fluorescence intensity was quantified in serotonin-positive cells in at least 40 optical fields taken from three independent experiments. Data are presented as the mean ± SD.
Segments of the ileum were removed and immersed in oxygenated Krebs-Ringer solution. After extensive wash to flush out the luminal contents, the tissues were embedded in OCT embedding medium (Tissue-Tek, Miles Laboratories, Naperille, IL, USA), and stored at −80°C. Sections (5 µm) were cut with a cryostat-microtome and thaw, mounted onto SuperFrost glass slides (Fisher Scientific, Illkirch, France). For HcA binding experiments, sections were removed from the freezer and incubated for 30 min with 10 µg/ml Alexa- or Cy3-labeled HcA in PBS/BSA (1%). After 3 washes in PBS, tissue sections were fixed with 4% PFA for 20 min at RT (22°C), rinsed in PBS and autofluorescence was quenched with 50 mM NH4Cl (15 min, 22°C). After permeabilization with 1% triton X-100 in PBS for 10 min, tissues were stained with TRITC-phalloidin (Sigma, 0.4 µg/ml), TRITC-Urex europaeus agglutinin type 1 (UEA1, 1 µg/ml), immunostained with anti-neurofilament 200 (Sigma; mouse, diluted 1∶500) and Cy3 coupled goat anti-mouse IgG (Sigma, diluted 1∶250). After 3 washes in PBS, sections were mounted in Mowiol and observed with a Zeiss confocal laser scanning microscope and a ×25 objective. For quantitative analysis of co-localization of HcA with markers, images of 512×512 pixels were taken from serial optical sections of 1 µm thickness. Data are presented as the mean ± SD.
The mouse neuroendocrine intestinal cell line STC-1 [45], m-ICcl2 [77] and Caco-2 (human colon) cells were grown on filter (Transwell, Corning) in Dulbecco's modified Eagle's medium (DMEM, Invitrogen) supplemented with 10% fetal calf serum (FCS, Invitrogen) until confluence. Integrity of tight junctions was confirmed by ZO-1 labeling and non-permeability to FITC-labeled dextran (4300Da, Sigma-Aldrich) (data not shown). BoNT/A as prepared previously [30] was added to the apical chamber. After 60 min incubation at 37°C, medium from the basal chamber was collected and BoNT/A was assayed by the mouse bioassay as previously described [30]. In the experiments with anti-proteases, culture medium in the apical and basolateral chambers was replaced with Dulbecco's medium containing 1% bovine serum albumin (BSA) and 1× anti-protease coktail without EDTA (Calbiochem) 30 min prior addition of BoNT/A into the apical compartment. BoNT/A transcytosis was monitored in basolateral medium samples after 60 and 120 min incubation at 37°C by a biological assay as previously described [30].
m-ICcl2 and STC-1 cells grown on glass coverslips (coated with poly-ornithine, Invritrogen, for STC-1 cells) were exposed to HcA-Cy3 (2.5 µg/ml), alone or in combination with a 10-fold more molar concentration of SV2C/L4-GST for 10 min at 37°C. Cells were washed twice with PBS, fixed with 4% PFA and mounted in Mowiol. The number of HcA fluorescent patches per µm2 was evaluated in m-ICcl2 and STC-1 cells by counting in images of 512×512 pixels (10 optical fields), taken from serial optical sections of 1 µm thickness from 3 experiments. Data was evaluated using ImageJ software (http://rsbweb.nih.gov/ij/).
Cells were lysed with boiling Laemmli buffer and lysates were homogenized by passages through a 26-gauge needle. An equal amount of protein (100 µg) from each boiled sample was loaded on a SDS-polyacrylamide (10%) gel. Samples separated by SDS-PAGE were transferred to nitrocellulose membrane (Amersham) and blocked in phosphate saline buffer containing 5% dried milk, washed with Tris-buffered saline containing 0.1% Tween20 (TBST), incubated with primary antibodies diluted in TBST overnight at room temperature, then washed 5 times for 10 min, and incubated for 1 h at room temperature with HRP-protein A diluted in TBST. After membrane washing in TBST, the specific signal was detected by enhanced chemiluminescence.
Specificity of the anti-SV2A antibodies was tested using lysate from SV2A-transfected cells (Santa Cruz). The lysate (10 µg) was run on a 10% SDS-PAGE, transferred on nitrocellulose, and blotted with anti-SV2A, anti-SV2B, anti-SV2C, and anti-SV2C-L4 antibodies. The specificity of anti-SV2B and anti-SV2C antibodies was tested on rat brain extract by competition with peptides, which had been used for immunization. Rat brain extract (5 µg, Santa Cruz) was run on a 10% SDS-PAGE, transferred on nitrocellulose, and blotted with anti-SV2B or anti-SV2C antibodies preincubated or not with the indicated SV2B or SV2C peptides (10 µg, Synaptic System) for 30 min at room temperature.
Values in the text are expressed as the mean ± SD, unless otherwise indicated. Differences between means were tested using Student's t-test, and p-values<0.05 were taken to indicate significance.
|
10.1371/journal.pcbi.1001118 | Multi-Scaled Explorations of Binding-Induced Folding of Intrinsically Disordered Protein Inhibitor IA3 to its Target Enzyme | Biomolecular function is realized by recognition, and increasing evidence shows that recognition is determined not only by structure but also by flexibility and dynamics. We explored a biomolecular recognition process that involves a major conformational change – protein folding. In particular, we explore the binding-induced folding of IA3, an intrinsically disordered protein that blocks the active site cleft of the yeast aspartic proteinase saccharopepsin (YPrA) by folding its own N-terminal residues into an amphipathic alpha helix. We developed a multi-scaled approach that explores the underlying mechanism by combining structure-based molecular dynamics simulations at the residue level with a stochastic path method at the atomic level. Both the free energy profile and the associated kinetic paths reveal a common scheme whereby IA3 binds to its target enzyme prior to folding itself into a helix. This theoretical result is consistent with recent time-resolved experiments. Furthermore, exploration of the detailed trajectories reveals the important roles of non-native interactions in the initial binding that occurs prior to IA3 folding. In contrast to the common view that non-native interactions contribute only to the roughness of landscapes and impede binding, the non-native interactions here facilitate binding by reducing significantly the entropic search space in the landscape. The information gained from multi-scaled simulations of the folding of this intrinsically disordered protein in the presence of its binding target may prove useful in the design of novel inhibitors of aspartic proteinases.
| The intrinsically disordered peptide IA3 is the endogenous inhibitor for the enzyme named yeast aspartic proteinase saccharopepsin (YPrA). In the presence of YPrA, IA3 folds itself into an amphipathic helix that blocks the active site cleft of the enzyme. We developed a multi-scaled approach to explore the underlying mechanism of this binding-induced ordering transition. Our approach combines a structure-based molecular dynamics model at the residue level with a stochastic path method at the atomic level. Our simulations suggest that IA3 inhibits YPrA through an induced-fit mechanism where the enzyme (YPrA) induces conformational change of its inhibitor (IA3). This expands the definition of an induced-fit model from its original meaning that the binding of substrate (IA3) drives conformational change in the protein (YPrA). Our result is consistent with recent kinetic experiments and provides a microscopic explanation for the underlying mechanism. We also discuss the important roles of non-native interactions and backtracking. These results enrich our understanding of the enzyme-inhibition mechanism and may have value in the design of drugs.
| “Intrinsically Disordered Proteins” (IDPs) are proteins that are disordered either in whole or in part. They play important roles in various cellular functions, including regulation, signaling and control processes [1]. Bioinformatic and statistical studies show that many proteins are intrinsically disordered: Of the crystal structures in the Protein Data Bank that contain no missing electron density, only about 30 percent show completely ordered structures [2], [3]. From this perspective, biological function may not require ordered structure. A key question is then, how do intrinsically disordered proteins carry out biological function?
Experiment and theory are beginning to probe the relationship between the dynamics and function of highly flexible IDPs [1], [4]–[12]. The intrinsically disordered proteinase inhibitor IA3, found in the cytoplasm of Saccharomyces cerevisiae, is an inhibitor of the protein vacuolar yeast peptidease A (YPrA). YPrA, which is also known as saccharopepsin [13], is a member of the aspartic proteinase family. The aspartic proteinases are present in many species, including vertebrates, fungi, plants and retroviruses [14], and they play a role in a range of pathologies that includes Alzheimers disease, hypertension, malaria and AIDS [15], [16]. Until recently, few peptide inhibitors of aspartic proteinases were known [17]. Even fewer structures of inhibitor-enzyme complexes have been determined. One complex that has been studied is that of the yeast peptidase A with its naturally occurring peptide inhibitor, IA3 [18]. Free IA3 is a 68-residue peptide that lacks a stable structure in solution. Upon interaction with with YPrA, the N-terminal region of IA3 folds into an amphipathic helix that blocks the active site cleft of the enzyme. [19]–[21]. Therefore, IA3 undergoes a major disordered-to-ordered transition during binding to its target enzyme. Understanding this transition and the mechanism of IA3's interaction with YPrA may provide clues as to how IDPs regulate their function through dynamics.
Narayanan and coworkers recently used laser temperature-jump fluorescence spectroscopy and fluorescence resonance energy transfer (FRET) to investigate the kinetics of the binding-induced folding of IA3 with YPrA [22]. A rapid kinetic relaxation in IA3 was observed in the presence of YPrA, whereas this process was absent in free IA3. Modeling of the kinetic data for both free IA3 and the IA3/YPrA complex indicated that unfolded IA3 binds with YPrA prior to forming its N-terminal helix. The present work uses a multi-scaled simulation approach to explore the binding of N terminal IA3 to YPrA. (The structure of the C-terminus in the bound complex is unknown.) Although molecular dynamics (MD) simulation is a powerful tool for investigating biomolecules, the time scales for the IA3/YPrA folding and binding interaction are too long for simulation in atomic detail by MD, at least at the present time. In order to bridge the gap of time scales between experiment and computation, several approaches have been developed that reduce the number of degrees of freedom. One method is to construct a structure-based energy function at a coarse-grained residue level [23]. A second method is to identify and quantify the optimal kinetic paths between the initial disordered and final ordered native states [24], [25]. The optimal paths are those paths that connect the reactant and product on the potential energy landscape surface with the largest statistical weight [26], [27]. In this work, we first carry out a structure-based coarse-grained residue level study of IA3 binding and folding. This step uncovers the underlying thermodynamics of the binding-folding free energy landscape. We then identify several optimal paths of IA3 binding to YPrA, as initiated from different starting points, based on a fully atomistic description of the protein. We address the effect of non-native and native interactions on the binding-folding of IA3. We obtain results that are consistent with the experimental findings [22]. This multi-scaled approach provides a detailed dynamic picture of the folding of a natural peptide inhibitor in the presence of its target enzyme.
In order to understand the binding-folding process from a global thermodynamic perspective, we explored the free energy landscape with a coarse-grained structure-based model by MD simulation under constant temperature. In this work, the simulation temperature is chosen to be lower than the binding transition temperature so that binding is possible and the target enzyme is stable. Meanwhile a harmonic biasing potential is introduced to accelerate the sampling. The harmonic biasing potential serves two purposes: (a) It prevents the IA3 molecule from being too physically distant from YPrA. This approach prevents the molecule from consuming too much computational time wandering in free space and searching for its interaction partner. It is analogous to simulating the system in a highly crowded cell-like enviroment where IA3 has higher chances of colliding with YPrA [5]. (b) The harmonic bias also facilitates crossing of the energy barrier by elevating the free energy basin of the complex. The biasing potential enhances sampling by minimizing trapping in less probable states. This idea is similar to the conformational flooding algorithm [28]. Finally, we can find the unbiased thermodynamic properties from our simulations by transforming back from the biased to the unbiased case, using Equation 2 of the Methods.
We take the normalized native contact fraction for folding of IA3 and the center of mass () distance between IA3 and YPrA as the order parameters that quantify the progress of the folding and binding process towards the final conformation of the YPrA-IA3 complex. The free energy profile shown in Figure 1 suggests there are two stable configurations: one is the unfolded and unbound state of IA3 and the other is the native binding-folding complex. The transition state ensemble corresponds to the region where of IA3 is in the range [0.3–0.5] and is in the range [2.2–2.5] nm. The finding that the unbound state corresponds to a nonzero , so that IA3 is not entirely disordered in unbound state, is consistent with NMR and CD measurements [17], [21], [22], which indicate that the N-terminus of IA3 is approximately folded when the peptide is free in solution. The fact that our result for the of the unbound state is larger than this value may reflect the fact that our chosen order parameter for the native contact fraction is not very sensitive to the fluctuations in the local contacts within the helical structure of IA3. We also measured the RMSD between the unbound and helical states of IA3. The average RMSD of Å (from 31 atoms) reflects the unstructured character of IA3 in unbound state. Overall, the coarse grained simulation reproduced the experimental properties of the system in a qualitative or semi-quantitative way. The free energy surface in Figure 1 indicates that binding and folding of IA3 are decoupled, with no folding occuring as the system approaches the transition state region. After the transition state however the binding and folding become strongly coupled. IA3 first approaches YPrA through binding from distant initial positions, then overcomes the transition state barrier, and finally folds itself into the structured conformation. Binding precedes folding.
From the free energy profile in Figure 1 we can conclude that IA3 binds prior to folding. Here we address the question of which regions of YPrA interact with IA3 at the transition state.
We captured the contacts between IA3 and YPrA by using the cutoff algorithm instead of counting only the native contacts . Figure 2A shows that the interfacial contacts at the transition state are distributed widely with low populations. Many of these contacts do not coincide with the native contacts (labelled by red square points) in the PDB structure of the IA3/YPrA complex. This implies that the transition state may be characterized by many non-native contacts and only a few native contacts. The important role of non-native interactions in the early stages of IA3 binding to YPrA can not be captured quantitatively by the structure based residue-level model, but it is explored in our full-atomic model, which uses a physics-based force field whose energy function combines the AMBER and OPLS force fields. Figure 2B shows the distribution of interfacial contacts in the transition state. Contacts are mostly formed at the surface of active site groove of YPrA, which is shown in blue in the cartoon representation. This distribution shows unambiguously that the first stage of the interaction involves IA3 binding to the surface of the active site groove. The highest peak, colored in red for emphasis, corresponds to the “flap” region, a hairpin loop formed by residues 72–82, which project out to cover the YPrA active site. This structural motif is commonly found in aspartic peptidases [15].
At our simulation temperature, YPrA is not a completely rigid partner in IA3 folding and binding. Figure 2C shows the effect of temperature in the RMS fluctuation in several local regions of YPrA. X-ray experiments [29] also show that the electron density is poor at the two loop regions marked with red squares in the figure. These two loop regions are the “flap” (or loop1) and a second region, named loop2. Comparing Figure 2B with Figure 2C shows a role for the “flap” region in controling IA3 binding to YPrA. The “flap” region forms the most contacts with IA3 although the RMS fluctuation data does not indicate a large capture radius. By contrast, loop2 has a largest capture radius as reflected by its structural fluctuation during binding, but it does not contribute to the interfacial contacts with IA3. Remarkably, at the tip of the flap, there is one absolutely conserved tyrosine (Tyr75) that is considered to play a crucial role in the capture and cleavage of substrates [30].
To gain further insight into the process that follows IA3 binding with the surface of active site groove, we investigated the distribution of the native interfacial contact fraction of individual IA3 residues () along the binding routes. In the crystal structure of the complex, the hydrophilic face of IA3 is oriented toward the solvent. The other face of IA3 is composed of the nine hydrophobic amino acid residues, V8, I11, F12, L19, A23, V25, V26, A29 and F30. This face is enveloped completely with the residues of the YPrA active site cleft and consists of three hydrophobic clusters: “cluster-1” (red) of V8-X-X-I11-F12 in the N-terminal, “cluster-2” (green) of L19-X-X-X-A23 in the mid region, and the C-terminal “cluster-3” (yellow) of V26-X-X-A29-F20 (see Figure 5 in Text S1). These clusters are indicated in Figure 3, which shows the evolution of the native interfacial contact fraction () of individual IA3 residues. We find that is well-distributed and less than 0.2 at the transition state region. By following the evolution of distribution along the binding routes we see that the mid region of IA3 forms native contacts with YPrA first, followed by the C-terminal region, and finally the N-terminus. However, the distribution of IA3 intrachain contacts does not show a sequential order of IA3 folding. It seems that the folding of IA3 does not necessarily occur from a particular nucleation site.
In studying protein folding and binding, the Q score (defined in the Text S1) for structural similarity has been extensively used as a structural reaction coordinate [31]–[34]. Q represents the fraction of native contacts that have been formed and it characterizes the structure's similarity to a referenced structure. Here, the referenced structure is the crystal structure of the enzyme-inhibitor complex of YPrA-IA3 (PDB code: 1DP5). To monitor the folding and binding of IA3 interacting with YPrA in a fully atomistic description, optimal kinetic paths were calculated in order to determine the most probable pathways between the beginning and ending points. The optimal paths depend on the choice of initial and end points. For the end point we use the structure of the native IA3-YPrA complex, as resolved by xray crystallography. The initial point is disordered, unfolded IA3 and uninhibited, folded YPrA. Obviously, the initial point for IA3 should consist of an ensemble of conformations with a sufficient number of degrees of freedom. Unfolded conformations of IA3 were generated by molecular dynamics with explicit solvent at high temperature. Three paths were chosen to illustrate the folding and binding process in detail. We refer to them as , and .
Figure 4 shows . and are the intrachain contact Q scores of IA3 and YPrA, respectively. The Q score of the interaction between them () represents the interfacial similarity relative to native binding complex of IA3 and YPrA. Figure 4A shows that the curve increases slowly close to 1, corresponding to the native inhibited structure. YPrA does not move much although it manifests some flexibility to accomodate the folding of IA3, mostly in the two loop regions located on the surface of the active site groove. Figure 4B shows the evolution of the folding score and binding score along the path. does not vary much when is less than 0.35. It even decreases slightly (IA3 unfolds) before grid 70 (where folding begins) due to backbone movements, not helix formation and breaking. This is consistent with the experimental indications that the pre-equilibrium of folding may not be helpful to IA3 folding through binding [22]. increases more and more along the path, especially when it exceeds the of IA3 after grid 64 and reaches 0.5 at grid 80. It implies that IA3 does not fold itself before binding tightly with YPrA. The binding score of IA3 to YPrA as a function of the folding score is shown in Figure 4B. Before entering into the hydrophobic cave of YPrA, IA3 searches the structural surface of YPrA with a continuous adjustment of its positioning, as reflected by the zigzag behavior of , until the binding score reaches 0.5. Folding proceeds once significant binding is realized on the interface, and binding and folding are subsequently coupled as the final native complex forms. The evolution of the structures and the related contact maps are shown in Figure 4c in Text S1.
The Q score measures only the residue-level similarity of the backbone to a reference structure. Therefore we introduce a Shadow Contact Map (SCP), an algorithm that calculates the interatomic contacts involving sidechain atoms [35]. The SCM algorithm describes the contact map excluding unphysical contacts. A cutoff distance is used to define the contacts, and this cutoff is set as Å in our calculation. As in the definition of Q score, the total number of contacts can be divided into monomeric folding contacts and interfacial contacts. The contacts are grouped into two categories: native and non-native contacts, as determined by whether the contacts between residue pairs exist in final conformation. For final state, the atomic interfacial contact number and the intrachain contact number of IA3 are 604 and 103, respectively. This indicates that the interfacial interactions are far stronger than the intrachain interactions in IA3. This may explain why the kinetic process proceeds as binding followed by folding.
The evolution of the number of atomic contacts and the helix formation in IA3 along the folding pathway of , and are shown in Figure 5B–5D. For exploring the relationships between atom-atom contacts and IA3 folding, the contact number curves are overlaid with the evolution of helix formation of IA3. Residues constituting helix are assigned via analysis by the DSSP program [36], according to characteristic hydrogen-bond patterns, but other secondary structure elements such as coils and turns are excluded from this plot for the sake of clarity. Note that we have not detected beta sheet elements in our model, although there are experimental reports that IA3 may bind pepsin as a beta-strand and is therefore cut and digested as a substrate [19]. From the evolution of number of interfacial contacts (blue line) and the native contacts in IA3 (black line), we can see that IA3 binds with YPrA more and more tightly before it begins to form native contacts and helix structure. The native interfacial contacts then (green line) begin increasing until most of native contacts in IA3 are formed. For and , the evolution of contact number is quite similar, but the corresponding processes of helix formation are significantly different. The long helix is formed from three nuclei located around the three hydrophobic clusters. Although there are significant differences between the three pathways, they reveal the common theme that IA3 binds to YPrA prior to folding. We also see clearly that non-native interactions are the dominant driving force in the initial stage of binding. Non-native contacts smoothly increase, while the native interface and native folding contacts only begin to appear at grid values near 70.
The average path is shown by the evolution of the average number of atomic contacts in Figure 6. A sharp increase in IA3 native contacts is observed in the black curve. This can be explained as the result of contact network forming in a highly synergistic way. Figure 6A shows that the native contacts of IA3 form together with non-native interfacial contacts while the number of interfacial native contacts remains nearly zero until grid 80. At the first stage of kinetic binding, the interactions between IA3 and YPrA are mostly contributed by non-native contacts. It is the non-native interactions between IA3 and the residues on the surface of YPrA that induce IA3 to bind with its target partner. Therefore, non-specific (non-native) interactions induce initial binding of IA3 to YPrA. After IA3 reaches YPrA, native interactions of binding set in by adjusting the conformation at the active site groove.IA3 folds into helical structure after binding with the active site groove of its target enzyme.
It is noteworthy that in (see Figure 5B) we observe a certain fraction of the helix content formed between grid 55 and 58 disappears and then later reappears. The formation and breakup of local secondary structure is observed not only in , but also in several other pathways that we calculated. In addition, we also found the formation and breaking of native contacts in (see Figure 2c in the Text S1). Remarkably, an analogous process was observed in the investigation of the folding of Interleukin-1 (IL-1) [37], [38], knotted proteins [39], CheY-like family [40], [41] and SAM-1 Riboswitch [42]. This behavior is known as “backtracking”. Here, we define it as the interim formation of local secondary structures or native contacts along the reaction pathways. Experiments in and in suggest that it is the result of topological frustration. Here, we propose that it results not only from topological factors but also from energetic contributions to the stability of IA3, as a partial compensation of entropy reduction during binding to YPrA.
During a biologically realistic interaction, binding to the special partner can help a protein to shrink the search space in the energy landscape. However, the associated free energy increases due to the rapid entropy reduction. As a partial compensation, occasional interactions may form only if they are energetically favored, irrespective of whether they are native contacts. These interactions do not form optimally but form easily in certain topologies. Therefore they are often unstable and fragile. During this process, both native interactions which are not stably formed and non-native interactions which are not included in the final state play roles in smoothing the free energy landscape. They stabilize the protein energetically, thus compensating the entropy reduction. As the molecule searches deeper in the free energy basin, the unstable native interactions will break down and reform stably in final structure. Hence backtracking is observed.
However we do not always observe non-native interactions or backtracking of native interactions along folding or binding pathways in nature. We explain this from three perspectives. First, the order parameters are usually coarse. They are not accurate enough to capture these details. Second, these interactions are transient and unstable. They may be difficult to measure. Third, not all the pathways are very rough. In this case of IA3-YPrA, it seems that non-native interactions play a more important role in IA3 binding while backtracking interactions are more significant in IA3 folding. From this, we believe non-native interactions and unstable native interactions can both play a role in protein folding and protein-protein recognition.
Like many IDPs, IA3 forms an ordered structure in the presence of its interaction partner. Its binding and folding dynamics play an essential role in the regulation of its target enzyme, YPrA. Molecular dynamics simulations can help us to explore the interaction at a level of detail that is difficult to obtain in laboratory experiments. However, standard MD is often limited by the temporal range it can probe at atomic detail. In this work, we developed a multi-scaled approach to provide a comprehensive picture of the protein binding-folding dynamics, including both global thermodynamic landscape and atomic details of structural evolution paths.
Several reaction pathways were generated from different starting points to the final conformation of the protein-inhibitor complex. Although there are significant differences between the multiple pathways, which reflect the multidimensional nature of the underlying energy landscape [23], [25], [27], [43], all reveal a common theme that IA3 binds to its target enzyme prior to folding itself into a helix. This finding is consistent with that of a coarse-grained free energy landscape from a structure-based MD simulation. In summary, the following folding and binding mechanism emerges. In the first step IA3 moves close to YPrA and binds to the surface of the active site groove via non-native interactions, through the long range electrostatic attraction. Before overcoming the free energy barrier, most of IA3 remains unstructured. Once IA3 enters into the cleft, its motion is greatly restrained, due to the lack of space for motion. In this highly hydrophobic environment, IA3 finally folds into an amphipathic helix at the long cleft. In addition, we found that the mid region of the IA3 sequence, consisting of hydrophobic , forms native interactions with YPrA earlier than the two terminal regions. This may be the result of stabilization by the interactions with the YPrA “flap”. During binding, YPrA plays the role of a template to induce IA3 folding into the characteristic structure that blocks the active site of the enzyme. In other words, the mechanism of saccharopepsin inhibition by IA3 as revealed by our simulation is in favor of the “induced-fit” model [12]. In this context, an “induced-fit” mechanism refers to a target enzyme that induces in its inhibitor a significant conformational change.
We also examined the non-native interaction by classifying the atomic contacts, as calculated by a new algorithm (SCM). At the first stage, the interactions between IA3 and YPrA are under the control of non-native interfacial contacts. The recognition process of the inhibitor-enzyme complex is dominated by these non-native interactions, which have been reported to play a role in protein assembly [9], [44]–[48]. The great success of simulating protein folding using structure-based models [49] which depend on the native topology suggests that the native contacts govern the folding of a protein that is well-designed by evolution. In the conventional view, non-native interactions are the major factor contributing roughness to the energy landscape [31], [46]. Why does non-native interaction seem to play a facilitating role for binding in the IA3/YPrA system? It is easy to explain the results in the view of the structure of the enzyme-inhibitor complex. As the target binding site is located deep in the groove, IA3 has to search the molecular surface of YPrA to find an entropically favored and energetically optimized path to the hydrophobic cleft. Experimental studies have already hinted that non-native contacts from the helix-forming (enzyme-inhibiting) N terminus as well as the disordered C terminus (not included in this study) of IA3 assist the kinetics during early stages of the interaction without affecting the final stability of the complex [17], [20], [22]. The importance of nonnative interactions was also observed in pKID and KIX binding experiments [1], [4], [9] and DNA-binding proteins [48]. These earlier findings support our conclusion here as well as the fly-casting mechanism [50].
From the energy landscape perspective, the underlying landscape of the entire binding process must be funnel-like in order to guarantee biological recognition and native binding complex formation. There are several ways to guarantee the underlying landscape to be funnel-like [23], [51]. One way is to enhance the native interactions or native bias. The other way is to reduce the non-native interactions or the roughness of the landscape. Those two ways are natural and conventionally emphasized. However there is another way to help the formation of the funneled landscape. A reduction in the entropy can significantly shrink the search space of the landscape. Here we see non-native interactions, even if not energetically favored, can contribute significantly to forming the binding funnel by reducing the entropy, bringing IA3 closer to the target YPrA interface). In this sense, non-native interactions can help the binding process.
Here we reveal the interaction mechanism of an aspartic proteinase and its endogenous inhibitor. Our studies provide a greater understanding of this unprecedented mode of enzyme inhibition. The results demonstrate the success of the multi-scaled approach for explorinng the interaction of IA3 and YPrA, and they are consistent with the conclusions from time resolved experiments, which suggest non-specific binding followed by folding [14]. The combined method may be useful in understanding other enzyme-inhibitor systems. It also may offer valuable insights into the design of drugs inhibitors for the aspartic proteinases generated by pathogenic organisms.
We performed the molecular dynamics simulations using a structure-based Hamiltonian to describe the energy of the protein in a given configuration. A structure-based Hamiltonian takes into account only native interactions, and each of these interactions enters into the energy balance with the same weighting. Therefore the model does not have heterogeneity in energy and it includes only topological frustration. Each amino acid is described by a single bead on a polymer chain located on the position [52]. The structure-based Hamiltonian is given by the expression:
The total energy is divided into bond stretching, angle bending, torsion and nonbonded interactions. , and are the virtual bond length, bond angle, and torsion angle defined by position. , and are the corresponding native values from the PDB structure. Nonbonded interactions are considered when two atoms i and j are separated sequentially by at least three residues on a chain or when they come from different chains, are subdivided into native interactions and nonnative interactions. For native contacts, is the distance between the positions of contacting residues i and j. For non-native contacts, provides excluded volume repulsion. We treat the nonlocal interactions within a chain and between the chains with the same strength. The native contact map is derived from a shadow contact map (SCM) [35]. Parameters , , , , weight the relative strength of each kind of interactions contributed to energy, and , , , , , . In order to sample more binding transitions we added a bias potential into the Hamiltonian. The bias potential is intended to make binding transitions more frequent by raising the free energy of the bound state. Here, we choose a harmonic form where the bias potential energy depends on the center of mass distance () between YPrA and IA3.(1)Here is the force constant, is the equilibrium position, is the COM distance between the two chains. Then the Hamiltonian has a new form . In the native complex, where IA3 is bound to YPrA, (only ) is 1.063 nm, while for larger than 3.5 nm, we consider the system to be in the unbound state. We choose to ensure that the bias potential lifts the free energy of bound states more than unbound states. Through many trials we found that an optimized value for the force constant is .
The unbiased thermodynamic average of a function can be calculated as follows:where is the reaction coordinate. The free energy of the system at is given by , where is the equilibrium probability.The unbiased free energy can be calculated as(2)The constant does not change in constant temperature simulations, so in our simulations we can select the value that sets the lowest to zero.
The simulations were performed using the Gromacs software package [53]. We put the protein system in a 50 nm cubic box corresponding to a low protein concentration. In fact, the effective box length is about 8.4 nm (the largest in coarse grained MD simulation). Nonbonded interactions are cut off at 3 nm. The time step was 0.5 fs. Stochastic dynamics were used with a drag coefficient . We started our trajectories with 9 different configurations in either native or nonnative state. The actual total constant temperature simulation time is . The total data include 214 binding and dissociation transitions, allowing us to observe how the dynamics change during the folding and binding. We calculated the free energy from the trajectories using WHAM (Weighted Histogram Analysis Method) [54], and using the formula 2 to get unbiased free energy. The simulation temperature is set at 176 K. Plotting the 2-D free energy surface for the binding/folding behavior requires two independent reaction coordinates, representing binding and folding respectively. From the transformation equation 2 we know that has an explicit expression only when contains . can describe the binding behavior. For folding, we choose , which is defined as the fraction of native spatial tertiary contacts. A native contact is formed if the distance between the two atoms is shorter than 1.2 times their native distance .
The free and inhibited states of YPrA were generated from crystal structures taken from the Protein Data Bank (1FMU and 1DP5, respectively). The unfolding of IA3 was generated by Langevin dynamics by NAMD with the Charmm32 force field. Then, the initial and final structures of the complex were modelled. Crystallographic water molecules and carbohydrate moieties were removed. After modeling of the reactant and product, the paths connecting these states were calculated with the MOIL package [55]. The MOIL energy function combines the AMBER and OPLS force fields [56], [57]. We can solve the minimum energy path if the pre-specified initial and final states are known. Given the minimized endpoint structures, the initial guesses for the trajectory are determined by the minimum-energy-path self-penalty walk (SPW) [58] functional embedded in the CHMIN module. Then these paths were optimized in the SDP module with steepest descent. The solvation effects are described by the Generalized Born model [59], [60]. The high-frequency modes from the trajectories are filtered and modeled as Gaussian white noise. The cut-off distance for van der Waals interactions is Å.
The steepest descent path is widely used in qualitative interpretation of chemical reactions [61], [62]. In analogy to the classical action, an action as a function of length in a discrete representation is defined to represent a most probable Brownian trajectory as follows:(3)
We split the path by N grids to approximate the path by a set of discrete conformations. is the entire vector of conformational coordinates at grid . The initial conformation is and the final conformation is . The potential energy is a function of the mass-weighted coordinate vector. The constant is an arbitrary positive value that mimics the energy in classical mechanics. Optimal paths with different thermal energies are generated by tuning this parameter. The steepest descent path is the limiting path that optimizes the action for . The shortest path between and as generated by linear interpolation is the optimal path for . Here we used .
Given the two end structures, the SDP module will minimize the target function:(4)(5)(6)where C is a restraint to ensure that configurations 's are distributed approximately uniformly along the pathway. The target function is minimized by conjugate gradient local minimization. is the arc-length of the path in mass weighted coordinates between conformation and . is the strength of a penalty function that restrains the step length to the average length . For further details see Ref. [63].
SDP is a continuous curve with a low-energy barrier that connects the reactants and the products. An important advantage of an SDP is that it allows testing of a concrete mechanism. The disadvantage is that it gives no information about the properties of the system far from the steepest descent path. Other non-native interactions which are away from the binding groove cannot be sampled on the steepest descent path. As we know, the Milestoning method [64], [65] has been developed by Ron Elber and coworkers to solve this problem. However, it is still an open question how to calculate kinetics and thermodynamics of long-time biological processes, which are typically not accessible by straightforward MD simulation.
Based on boundary conditions, the initial and final coordinates must be specified. We take the crystalline structure of YPrA complexed with IA3 mutant inhibitor (PDB code: 1DP5) as the endpoint of the transitional trajectories. We assume that the trajectories start with the uninhibited enzyme and unfolded IA3 that is far from the binding site of proteinase A. For free YPrA the coordinates in trigonal and monoclinic crystal forms are accessible from Protein Data Bank under accession codes 1FMU and 1FMX, respectively. Here, we adopt the coordinates of the trigonal crystal form, not only for the clarity of the electron density in the “flap” consisting of a hairpin loop extending over the active site, but also because of the possible presence of some hydrolysis products in the monoclinic crystal [29]. However, in the crystal structure of the trigonal form, there are two disordered regions in which the electron density is relatively poor. The two highly flexible regions, located at the peptide segments 162–165 and 243–245, are considered to make less contribution to the inhibitor binding as their locations on the molecular surface are far from the active site. The missing segments were modelled by structure prediction. Although YPrA is glycosylated its covalently binding carbohydrate moieties are not considered in the simulation. The structural difference between the initial and final states of YPrA is shown in Supplemental Figure 1A in Text S1. Obviously, the unfolded state of the inhibitor can not be represented by a single structure. It should be an ensemble of conformations having a sufficient number of degrees of freedom. These were generated by Langevin dynamics using NAMD [66] with Charmm22 force field. The initial conformation of the system was constructed in stages, starting with the complex of folded IA3 and uninhibited YPrA with a center of mass distance of 3 nm from each other, followed by packing the complex with a 4 nm thick water sphere. We then carried out minimization using the conjugate gradient algorithm with 1000 steps. The initial distance between IA3 and YPrA was set to about 3 nm as hinted by the coarse grained model. The region of most interest is the range from 1.06 nm to 3.0 nm, where folding and binding occur, which is indicated by the yellow region in Figure 7. In order to generate the conformation of the complex, the minimized system was heated to 500 K in the canonical ensemble. The procedure employed a Langevin thermostat with a damping parameter. Constraints were applied to the lengths of all bonds involving hydrogen atoms, thus allowing a 2 fs time step. A spherical boundary condition was used to control the 4.5 nm thick water sphere from the center of mass of the complex. YPrA was fixed during the high temperature dynamics. After heating IA3 for 1 ns, we extracted intermediate structures (without water) whose RMSDs from the helical structure were larger than Å and that were separated by interval steps larger than 1 ps.
|
10.1371/journal.ppat.1003912 | Viral MicroRNA Effects on Pathogenesis of Polyomavirus SV40 Infections in Syrian Golden Hamsters | Effects of polyomavirus SV40 microRNA on pathogenesis of viral infections in vivo are not known. Syrian golden hamsters are the small animal model for studies of SV40. We report here effects of SV40 microRNA and influence of the structure of the regulatory region on dynamics of SV40 DNA levels in vivo. Outbred young adult hamsters were inoculated by the intracardiac route with 1×107 plaque-forming units of four different variants of SV40. Infected animals were sacrificed from 3 to 270 days postinfection and viral DNA loads in different tissues determined by quantitative real-time polymerase chain reaction assays. All SV40 strains displayed frequent establishment of persistent infections and slow viral clearance. SV40 had a broad tissue tropism, with infected tissues including liver, kidney, spleen, lung, and brain. Liver and kidney contained higher viral DNA loads than other tissues; kidneys were the preferred site for long-term persistent infection although detectable virus was also retained in livers. Expression of SV40 microRNA was demonstrated in wild-type SV40-infected tissues. MicroRNA-negative mutant viruses consistently produced higher viral DNA loads than wild-type SV40 in both liver and kidney. Viruses with complex regulatory regions displayed modestly higher viral DNA loads in the kidney than those with simple regulatory regions. Early viral transcripts were detected at higher levels than late transcripts in liver and kidney. Infectious virus was detected infrequently. There was limited evidence of increased clearance of microRNA-deficient viruses. Wild-type and microRNA-negative mutants of SV40 showed similar rates of transformation of mouse cells in vitro and tumor induction in weanling hamsters in vivo. This report identified broad tissue tropism for SV40 in vivo in hamsters and provides the first evidence of expression and function of SV40 microRNA in vivo. Viral microRNA dampened viral DNA levels in tissues infected by SV40 strains with simple or complex regulatory regions.
| The recent discovery of virally encoded microRNAs (miRNAs) raises the possibility of additional regulatory processes being involved in viral replication, immune recognition, and host cell survival. In this study, we sought to characterize the effect of SV40-encoded miRNAs and the structure of the viral regulatory region on infections in outbred Syrian golden hamsters. Results revealed that SV40 has a wide tissue tropism, including liver, kidney, spleen, lung, and brain, with kidney the preferred site for long-term persistent infection. Significant increases in tissue-associated viral DNA loads were observed with miRNA-negative mutant strains, whereas the presence of SV40 miRNAs had no effect on tumor induction and little effect on viral clearance. Our results provide the first evidence for SV40 miRNA expression and function in an in vivo animal model and highlight the complexity of regulation of SV40 viral replication and persistent infections.
| Definition of steps in the pathogenesis of a viral infection and of the response of the host to the viral assault is important to understanding the genesis of associated diseases and to guiding the development of viral diagnostics, therapeutics, and preventative measures. Viral microRNAs (miRNAs) are predicted to be involved in the infection process.
Polyomaviruses are small, nonenveloped, DNA-containing viruses that establish persistent infections in susceptible hosts and induce tumors under certain conditions. The Polyomaviridae family contains several viral species able to infect humans, some of which have been linked with disease, especially in immunocompromised hosts. Little is known of the patterns and dynamics of acute and chronic infections by polyomaviruses in vivo, except that kidneys appear to be a common site of long-term persistence [1].
Simian virus 40 (SV40) is the type species of the virus family. Since its discovery in 1960, it has served as a model cancer virus and has revealed many insights into fundamental cell processes and basic mechanisms of cell transformation [2]–[8]. SV40, whose natural host is the rhesus macaque monkey, is able to infect humans. It is currently the only polyomavirus that can infect humans that has a small animal model amenable to mechanistic studies of early steps in viral pathogenesis. Syrian golden hamsters (Mesocricetus auratus) have been widely utilized for studies of SV40 infections and oncogenesis [9]–[17]. It has been shown that SV40 isolates differ in oncogenic potential in vivo, with those having a regulatory region (RR) with a single enhancer region being more oncogenic in weanling hamsters than those with a complex, rearranged RR. These RR differences, however, had no effect on transforming frequencies in vitro, suggesting that differences in tumor induction reflected viral interactions with the host [16]. Conversely, vertical transmission of virus from pregnant dams to their offspring was more frequent when infections involved an SV40 strain with a complex RR [15]. In addition, viruses with a complex RR induce more frequent antibody responses to the SV40 large T-antigen (T-ag) in tumor-free animals than those with a simple RR following intraperitoneal inoculation [10], [16]. We interpret these observations as indications that viruses with complex RR replicated better in vivo, increasing the likelihood of vertical transmission and of elevated synthesis of T-ag, the latter leading to more detectable T-antibody responses. Despite these findings, the effect of viral genetic variation among virus strains on the natural history of acute and persistent infections is not known.
miRNAs are small, noncoding RNAs that are key regulators of gene expression, able to control large networks of genes by binding and repressing mRNAs; miRNAs can regulate cell cycle genes and host response genes, modulate downstream gene expression, and play a role in tumorigenesis [18], [19]. SV40, the first polyomavirus found to encode viral miRNAs, expresses a precursor miRNA giving rise to two derivative effectors during infection in cultured cells. Both SV40 miRNA effectors (5p and 3p) reduce expression of the viral early proteins [20]. The presence or absence of viral miRNA appeared to have no effect on replication of the virus in tissue culture, although there were higher levels of T-ag in mutant-infected cells, and it was proposed that the miRNAs might function in vivo to help infected cells avoid killing by the host immune response. To date, no animal studies addressing SV40 miRNA function in vivo have been reported.
This project was undertaken to characterize patterns of acute and chronic infections by polyomavirus SV40 in a susceptible host. Specific objectives were (i) to identify tissues that can be infected by SV40 and those that are sites of long-term persistence, (ii) to measure quantitative changes in viral DNA levels over time in different tissues, (iii) to determine if SV40 genetic differences, including the presence of viral miRNA and the structure of the viral RR, affect the patterns of infections in vivo, and (iv) to evaluate the influence of viral miRNA on tumor induction.
Young adult, outbred Syrian golden hamsters were inoculated by the intracardiac route with 1×107 plaque-forming units (PFU) of two wild-type (WT) strains of SV40, one having a complex (2E) RR (776-WT) (GenBank accession No. J02400) and one a simple archetypal (1E) RR (SVCPC-WT) (GenBank accession No. AF156108), and with mutants unable to produce SV40 miRNAs derived from each WT strain (776-SM1, SVCPC-SM2).
The intracardiac route of inoculation was used to distribute the virus throughout the body in order to identify susceptible tissues. Animals were sacrificed at times ranging from 3 to 270 days postinoculation (p.i.), organs were harvested, and the presence of SV40 DNA was determined by real-time quantitative polymerase chain reaction (RQ-PCR) assays as detailed in Materials and Methods. The total numbers of tissues analyzed and the percentages that were positive for detectable levels of viral DNA are summarized (Table 1). All four viruses were efficient at establishing persistent infections. Animals were frequently virus-positive in liver, kidney, and spleen through day 45 and in kidney through day 270. In fact, all the liver and kidney samples collected through day 45 from infected animals were virus positive with only two exceptions. Spleen, lung and brain less commonly contained detectable viral DNA and virus disappeared more quickly. Muscle and cheek pouch specimens from 15 animals inoculated with the 776 viruses were tested at days 3, 7, and 45 p.i. and all were negative for viral DNA (data not shown).
Polyomaviruses are known to establish persistent infections and these results illustrate the highly efficient nature of the process by SV40 (Table 1). At this level of analysis in which tissue samples were either positive or negative for viral DNA, no obvious effect of viral miRNA or the viral RR on the dynamics of viral presence in different tissues could be discerned.
Viral DNA levels in hamster tissues were quantitated using an SV40-specific RQ-PCR assay. The single-copy hamster vimentin gene (GenBank accession No. AH001833.1) was also measured by RQ-PCR to determine the number of cell equivalents of DNA in tissue extracts; viral loads were then expressed as SV40 DNA copies per 104 cells. Observed viral loads (log2) in liver and kidney samples from individual animals are shown for each of the four viruses (Figure 1). Geometric mean values for each set of experimental animals are represented by horizontal lines. Decreases in viral loads tended to occur between days 14 and 28 with all four viruses, although viral DNA remained readily detectable. There was often a range in viral loads among the animals within a virus group harvested at each time point, presumably reflecting at least in part individual variation typical of outbred animals.
The relative amounts of virus found in liver, kidney, spleen, lung, and brain early during infection (days 7 and 14) are shown (Table 2). Data for the four viruses are presented as averages of observed values (SV40 DNA copies/104 cells). During the early stages of infection SV40 viral loads were higher in liver and kidney than the other tissues with all the viruses, ranging from several fold to 100-fold in excess. Higher levels of SV40 DNA copies were usually detected with the miRNA-negative mutants than with the WT viruses in the liver and kidney samples. This difference was observed despite the fact that fewer viral genome copies were inoculated per 107 PFU with the mutant viruses than with the WT strains (Materials and Methods). There was a trend for viral loads of WT viruses to increase between days 7 and 14 in liver and kidney, in contrast to decreases in viral loads of the miRNA mutant viruses in most cases. Subsequent detailed analyses focused on liver and kidney specimens.
Analysis of viral load data over time quantitated the preferential persistence of SV40 in the kidney compared to the liver and spleen. Virus levels, expressed as a ratio of day 45 to day 3 values for each virus (SV40 DNA copies/104 cells), are shown (Table 3). Between 0.41 and 1.29 of day 3 levels were detected at day 45 in kidneys (equivalent to 41% and 129% of day 3 levels), in contrast to ratios of 0.10 to 0.17 in livers and even smaller ratios (≤0.10) in spleen. These day 45∶day 3 ratios show a higher relative persistence of SV40 in kidney than in liver by day 45 p.i. with all four viruses. The ratios of viral DNA levels in kidney∶liver and in kidney∶spleen at each time point are detailed in supplemental information (Table S1). These data suggest there was a reproducible trend for long-term viral persistence in the kidney, consistent with the idea that the kidney is the major reservoir for persistent infections in natural hosts by the well-characterized polyomaviruses. However, the retention of detectable viral DNA in the liver suggests that cells in the liver may be an unrecognized site of persistent infection by SV40 (Table 3). We noted there was a higher viral load ratio for 776-WT virus in kidney relative to SVCPC-WT virus (approximately 1.5-fold higher) as well as a higher ratio for the two miRNA mutants in kidney compared to the respective parental WT virus (about 2-fold higher) (Table 3). These results suggest that both a complex RR and the absence of viral miRNA can lead to higher chronic levels of SV40 viral DNA in the kidney of susceptible hosts. Virus was frequently detected in the kidney at day 270, ranging from 62% to 100% of animals tested (Table 1). However, only occasional animals (0–43%) had detectable virus remaining in the liver at day 270, and it is noteworthy that the 776 miRNA mutant (776-SM1) was not detected.
The expression of the SV40 miRNA in hamster liver and kidney tissues was analyzed as described in Materials and Methods. Briefly, total RNA was harvested and then size-fractionated to enrich for small RNAs. The presence of the SV40 3p miRNA was determined via RT-PCR. Specimens were coded and tested without knowledge of sample identity. A sample was scored positive if its CT was ≤35, which is within the linear range of this assay (as determined using a standard dilution series).
Expression of SV40 miRNA was detected in tissue samples from animals infected with WT SV40 strains, but not in tissues infected with miRNA-negative mutants (Table 4). Compared to cultured African green monkey BSC40 cells infected with SV40 at a high multiplicity of infection (MOI) (CT 12.2), or to an abundant host miRNA in BSC40 cells (miR-Let7a, CT 24.6), the levels of the SV40 miRNA in hamsters were lower (CT 31–<35); likely reflective of the fact that only a small fraction of the assayed cells was infected. Twelve of 15 (80%) WT-infected samples tested positive for SV40 miRNA, including 7 of 7 (100%) kidney specimens and 5 of 8 (62%) liver samples. All control samples tested were negative for SV40 miRNA (0/6, 0%), including 4 miRNA mutant virus-infected tissues and 2 samples from uninfected hamsters. These results represent the first detection of SV40 miRNA expression in vivo in intact animals.
SV40 miRNA effects on the dynamics of SV40 infections in hamsters were then examined in detail. To directly compare the infection patterns of paired WT and miRNA-negative mutant viruses, observed viral genome counts were normalized as described in Materials and Methods, geometric mean titers were calculated, and expressed as viral DNA copies (ln) per 10,000 cells (Figure 2). In the 776 system, the viral loads for the SM1 mutant were consistently higher than those of the WT virus in both liver and kidney samples from day 3 through day 45 (Figure 2A). These differences between the two viruses were statistically significant at p<0.05 in both tissues at each time point, except for day 3 in the kidney. Similarly in the SVCPC system, viral DNA levels for the SM2 miRNA-negative mutant were higher than those of the WT virus in both liver and kidney from day 3 through day 45 (Figure 2B). The differences between these two viruses in both tissues were also statistically significant (p<0.05) at each time point, except for day 7 in the liver.
Viral DNA copies remained detectable at low levels at day 270 in most kidney specimens and less frequently in liver samples (Tables 1, 5). In contrast to the SVCPC-SM2 miRNA mutant, the 776-SM1 mutant was not detected in the liver of any hamsters tested at this late time point, suggesting that the SM1 infections may have been cleared from the liver. The differences in viral DNA loads at day 270 were statistically significant between the miRNA mutant and its WT virus in the SVCPC system in both tissues and between the two miRNA mutants in both tissues (Table 5). The difference in viral loads between the two WT viruses in the liver was also statistically significant.
The relative influence of the structure of the SV40 RR and of the presence or absence of viral miRNA on the dynamics of virus infection in both liver and kidney was next evaluated (Figure 3). The bars represent the ratios of the means of normalized viral load titers between the two viral strains listed below each set; individual bars within a set reflect days 3 to 270 p.i. The two miRNA-negative mutants had relatively higher viral DNA loads than the matched parental WT viruses in the liver, reaching as much as >100-fold higher (Figure 3A). No bars are shown for day 270 in the liver for the two comparisons involving the 776-SM1 mutant because 776-SM1 viral DNA was not detected in the livers of infected animals at day 270 (Tables 1, 5). The effect in the liver of the absence of miRNA was more dramatic with the 776 system than with SVCPC. The mutant viruses also displayed higher viral DNA loads than the parental WT viruses in kidneys, differences that reached >20-fold (Figure 3B). The WT and mutant viruses with complex RR (776-WT, 776-SM1) reached higher viral loads (≤5-fold) in kidney as compared to the viruses with simple RR (SVCPC-WT, SVCPC-SM2). The loss of viral miRNAs had a larger impact than the structure of the viral RR on the amount of viral DNA in infected tissues.
Expression levels of SV40 early (T-ag) and late (VP1) regions in liver and kidney were determined as described in Materials and Methods. The single copy hamster vimentin gene was used to normalize viral transcript levels; results are presented as SV40 mRNA copies per 106 vimentin mRNA molecules (Table 6). The viral early region was expressed at higher levels than the late region in both liver and kidney samples at time points ranging from 3 to 45 days p.i. The WT viruses and miRNA-negative mutants displayed similar expression patterns, with transcript numbers routinely higher in liver than in kidney. It was noted that the T-ag mRNA levels from 776-SM1 consistently exceeded those of the 776-WT virus in the liver. There appeared to be a trend for T-ag transcript levels in the liver to decrease over time, whereas those in the kidney increased.
Recovery of infectious virus was attempted from liver and kidney specimens from hamsters exposed to the four virus strains. Tissue lysates were prepared and tested as described in Materials and Methods. Tissues included kidney samples from animals sacrificed at 7, 14, 45 and 270 days p.i. and liver samples from animals harvested at 7 and 14 days p.i. (2 or 3 animals each). Infectious virus was detected after passage of kidney specimens collected on day 7 (776-SM1, 3 of 3; SVCPC-SM2, 3 of 3; and SVCPC-WT, 1 of 2). Other kidney samples were negative and no infectious virus was recovered from liver materials tested. These results suggest that the majority of the viral DNA detected in these tissues was not present in the form of infectious virus particles by 14 days p.i.
Focus-forming assays using primary mouse embryo fibroblasts were carried out as described in Materials and Methods to determine if the absence of SV40 miRNA affected transformation by the virus in vitro. Three sets of WT and miRNA-negative mutant viruses were tested, including the 776 and SVCPC strains. Also tested in the transformation studies were SVCPC-based recombinants carrying either the 776-WT large T-ag gene (SVCPC-776-WT) or the T-ag gene with the SM1 mutation (SVCPC-776-SM1). Replicate infected cultures were harvested and transformed foci counted at 3 and 5 weeks p.i. For each matched pair of viruses, numbers of foci were expressed as the percentage of WT virus transformed foci at 5 weeks (set as 100%). All six viruses were able to transform mouse cells in vitro (Table 7). The foci induced by the mutant viruses were somewhat slower growing and appeared later, but by 5 weeks p.i. there was an insignificant 2-fold difference in the relative number of foci produced by the 776-derived mutants. The SVCPC-WT transformed foci were slower at appearing than those induced by the 776-WT viruses, and the SVCPC-SM2-induced foci had a slower growth rate. Results showed that the lack of miRNA neither enhanced nor abolished the transforming ability of SV40 in cultured mouse cells.
The WT and miRNA mutants were tested for tumor induction in hamsters to determine if miRNA effects were manifest during in vivo oncogenesis. Weanling (21-day-old) animals were inoculated intraperitoneally with 1×107 PFU of virus and observed for 12 months. A third pair of viruses (SVCPC-776-WT and SVCPC-776-SM1) was included in the in vivo tumorigenicity experiment. All six viruses induced tumor formation. There was no significant difference in tumor incidence between a given WT virus and its miRNA-negative mutant for any of the three virus pairs (Table 8). There was also no difference in the latency period (time to tumors) between the WT and miRNA mutant viruses. As observed in earlier studies, virus strains with a simple RR (SVCPC) were more oncogenic than virus strains with a complex RR (776) [16]. Susceptibility to tumor induction by SV40 is age-related in hamsters. Tumors were not observed in the animals inoculated by the intracardiac route in this study because those animals were older at the time of virus exposure (Materials and Methods).
Tumors induced by WT and mutant viruses were characterized for the content of SV40 DNA and for viral gene expression. The following tumor-associated SV40 DNA levels were determined by RQ-PCR [copies/cell, mean (range)]: SVCPC-776 = 32 (1–50) [5 tumors], SVCPC-776-SM1 = 14 (1–48) [5 tumors]; SVCPC = 15 (1–23) [3 tumors], and SVCPC-SM2 = 10 (6–14) [4 tumors]. Tests were not carried out to determine the integrated or episomal status of those genome copies in the tumor cells. Total RNA was isolated from the same tumors, reverse transcribed, and SV40 transcripts quantitated by RQ-PCR (Figure 4). Results are expressed as the number of SV40 transcripts per 106 copies of 18S ribosomal RNA (GenBank accession No. X03205.1). Both early and late viral mRNAs were detected in the tumors, with the SV40 early mRNAs being more abundant than the late mRNAs.
Antibody responses to SV40 T-ag as well as neutralizing antibody responses to SV40 were measured as described in Materials and Methods (Table 9). All but one of the 33 tumor-bearing animals (97%) produced T-antibodies with similar median titers. Similarly, all (100%) of the tumor-bearing animals produced SV40 neutralizing antibody with comparable neutralization titers, regardless of the virus strain inoculated. Results were more variable among the 70 animals that failed to develop tumors. The frequency of T-antibody responses in tumor-free animals ranged from 0% to 72% per group with median titers that tended to be lower than those in animals with tumors. Animals exposed to the 2E viruses (776) tended to produce T-antibody more frequently than those exposed to the 1E (SVCPC) viruses. There were no discernible differences in patterns of T-antibody responses in tumor-free animals related to the inoculation of miRNA-positive or -negative viruses. Those responses indicate there was sustained T-ag expression in the animals adequate to induce the production of stable T-antibodies. From 42% to 80% of a given group produced neutralizing antibody with median titers that tended to be lower than those in the tumor-bearing animals in the same group.
One goal of this study was to identify tissues susceptible to SV40 infection and those that became persistently infected. Hamster liver, kidney, and spleen samples were frequently viral DNA positive through day 45 p.i.; lung and brain tissues were less often virus positive and muscle specimens were consistently virus negative (Table 1). Higher viral loads (SV40 DNA copies/104 cells) were routinely detected in liver and kidney samples as compared to other tissues, with virus titers in the liver often exceeding those in the kidney during the early phases of infection (Table 2, Figure 1, Figure 2). These results show that tissues varied in their susceptibility to infection by SV40. Both SV40 miRNA in WT infections and viral transcripts were detected in liver and kidney samples (Tables 4 and 6), evidence of viral gene expression in infected cells. Most kidney samples were virus-positive through day 270, whereas liver tissues remained infected in most animals for as long as 45 days p.i. and in some animals as late as day 270. It is possible that livers are an unrecognized site of polyomavirus persistence and it will be important in future studies to identify the target cells in the liver that support long-term SV40 infections. Evidence suggests that polyomaviruses can be transmitted by the fecal/urine–oral route and can be found in lymphoid-rich samples from hamsters and humans, so cells in the gastrointestinal tract and in lymphoid tissues are of interest as well [2], [12], [14], [21]–[26] (Tables 1, 2). Results from an earlier study of SV40 vertical transmission in hamsters are compatible with the tissue susceptibilities described here [15]. In that investigation pregnant hamsters were inoculated intraperitoneally with 1×107 PFU of SV40 strains and were sacrificed at different times up to 24 days p.i. Kidney and spleen were infected in all 16 maternal animals, and lung in some animals, confirming that cells in these tissues are susceptible to infection by SV40.
Viral miRNAs have been discovered for several virus families, including the herpesviruses and the polyomaviruses [19], [20], [27]–[33]. The functions of many of the viral miRNAs have not yet been determined but are proposed to be involved in prolonging the viability of infected cells and in promoting immune evasion to increase virus survival [19], [30], [31], [34]. Viral miRNA effects are likely to be of significance to the establishment, maintenance, and/or reactivation of persistent infections in susceptible hosts by viruses that commonly establish chronic infections. Very few studies have examined the effect of viral encoded miRNAs on viral infection in an intact host. This study provides the first evidence to our knowledge of the detectable expression and functional effects of SV40 miRNAs on polyomavirus SV40 acute and persistent infections in vivo in a susceptible host. Hamsters infected with SV40 mutants unable to express the viral miRNAs frequently contained higher levels of viral DNA in liver and kidney tissues than animals exposed to the WT viruses. This finding is compatible with the original report that SV40 miRNAs reduce the expression of SV40 T-ag in infected cells [20], as higher levels of T-ag in mutant-infected cells could be expected to support enhanced viral DNA replication.
SV40 strains tested in this study differed in the structures of their RR. Variants with a duplication of the enhancer element in the RR (i.e., complex RR) can replicate to somewhat higher levels (∼2-fold) in DNA replication assays and in cell cultures than variants with a simple RR [35], [36]. It has also been shown that overexpression of T-ag stimulates replication of human JCV and BKV isolates with simple RR in vitro [37]. In the current study, the SV40 776 strains with complex RR exhibited approximately 2- to 5-fold higher viral loads in hamster kidney than the SVCPC strains with simple RR (Figure 3B), suggesting an RR effect on virus replication in vivo in kidney cells. A complex RR would likely have a more pronounced enhancement on virus replication in immunocompromised hosts. It has been observed that natural SV40 variants with rearranged RR arise de novo in infected immunocompromised rhesus monkeys [38] and this same phenomenon has been observed with human polyomaviruses. JCV isolates with a simple RR are shed in the urine, whereas isolates with a complex RR are recovered from the brain and cerebrospinal fluid of patients with progressive multifocal leukoencephalopathy [39]–[41]. The emergence of BKV with a rearranged RR is linked to increased viral replication and increased BKV-associated nephropathy disease in kidney transplant recipients [42].
Notably, our results showed that the loss of viral miRNA had a greater effect on viral DNA loads than that of the complex RR structure. However, the absence of viral miRNA resulted in a relatively larger increase of viral DNA in the 776 system than it did with the SVCPC strain (Figure 3). We speculate that this stronger effect might reflect the combination of the complex viral RR and the absence of viral miRNA. Another virus distinction is that there are several nucleotide differences between strains 776-WT and SVCPC-WT in the T-ag gene, plus a 9-bp insertion in the SVCPC gene, changes which result in 6 amino acid differences between the two T-ags (Figure S1) [43], [44]. However, we do not believe these changes affected the level of viral DNA in hamster tissues. We compared the average titers of virus stocks of 776-WT (n = 18) and SVCPC-WT (n = 15) prepared in our laboratory at different times over the last decade. There was no statistical difference (p>0.05) between the groups, indicating no obvious differential T-ag effects on virus replication.
SV40 miRNA was detected (Table 4) when levels of SV40 late transcripts were low (Table 6) and infectious virus was not recovered. This raises the question of how functional SV40 miRNAs were expressed in the absence of abundant lytic infections of hamster cells. The most likely explanation is that some cells in a given tissue were supporting complete cycles of virus replication and the late transcription in those cells produced adequate amounts of miRNA to yield the observed results. Alternatively, it has been speculated there may be a mechanism to produce miRNAs in polyomavirus chronically infected cells that is not dependent on late viral gene expression [19]. Importantly, it has been shown recently that BKV miRNA is expressed in BKV-infected cultured cells before the onset of viral DNA replication [45]. The BKV study observed increased replication of miRNA-null mutants with an archetypal RR in renal proximal tubule epithelial cells, similar to our findings with SV40 in hamsters. However, the absence of BKV miRNA had no stimulatory effect on replication of BKV variants with a rearranged RR, in contrast to our SV40 results [45]. Whether this reflects a fundamental difference between BKV and SV40, the influence of cell culture vs. in vivo conditions, or some other experimental variable is a topic for future study.
It has been proposed that SV40 miRNA may function to mediate evasion of the host response to infected cells [20]. However, in this study the mutant viral loads were not reduced more quickly than those of WT viruses over the first 45 days, suggesting that mutant virus-infected cells were not cleared more quickly by the host immune response (Figure 2). There was a statistically significant difference between 776-WT and 776-SM1 in the number of animals that retained viral DNA in the liver at day 270, but not between SVCPC-WT and SVCPC-SM2. There are several possible explanations for the inability to detect pronounced miRNA effects on viral clearance. One possibility is that the original model does not encompass the most important role for the miRNA during infections. Perhaps SV40 has an unrecognized non-miRNA mechanism for interfering with the host immune response to allow establishment of chronic infections, making the presence or absence of viral miRNAs irrelevant. Another possibility is that the SV40 miRNAs may have been unable to target a cellular mRNA critical to the host response because of sequence mismatches between the primate virus and the hamster genome. (It has been proposed that human JCV and BKV miRNAs target a stress-induced human cell ligand recognized by natural killer cells to mediate killing of virus-infected cells [46].) Alternatively, the Syrian golden hamster immune system may possess some unusual characteristic that makes the animals prone to establishment of persistent infections by microbial agents, thereby over-riding or negating possible miRNA effects. For example, hamsters are known to be susceptible to infections by numerous viruses and a parasite able to infect humans [47]–[56]. Finally, although viral miRNA expression was detected and differences were observed in viral loads between the WT and viral miRNA mutant viruses in hamster tissues, perhaps the levels of the viral miRNAs were insufficient to mediate effects on viral clearance. Because of these unknown factors, we cannot conclude that the viral miRNA never affects the host immune response. It is noteworthy that our findings are similar to the observation that a miRNA-negative mutant of murine polyomavirus generally behaved like the WT virus in experimentally infected mice with respect to viral clearance and cellular immunity. Although expression of miRNAs was not directly demonstrated, the genome copy loads of the mutant tended to exceed those of the WT polyomavirus in adult mouse kidneys [29].
There are several limitations to this study. Hamsters are not the natural host for SV40, but they represent one of the few systems that can be used for investigations of polyomavirus infections in intact animals. The intracardiac route of inoculation that was used to allow determination of the breadth of susceptible tissues and the sites of viral persistence was not the natural route of infection. It was beyond the scope of this study to compare different routes of inoculation. In addition, the dose of virus inoculated was high compared to the low level of virus encountered in a natural infection. It is conceivable that the nonnatural route of inoculation coupled with the high level of inoculum might have masked possible phenotypes. The virus did not appear to produce infectious progeny beyond the early stages of infection, although it is possible that additional tests of more tissue samples might have detected extended replication. Alternatively, it is possible that SV40 enters a persistent, nonproductive state relatively soon after infection.
The study raised several challenging subjects for follow-up investigations. These include a potential role of viral miRNA in reactivation of SV40 persistent infections, the molecular basis for the difference between viral miRNA inhibitory effects on BKV and SV40 viruses with complex RR, the mechanism of expression of SV40 miRNA in the absence of abundant viral replication, and the interaction of SV40 with susceptible cells in the liver.
In summary, this study demonstrated for the first time that SV40 miRNAs are expressed and are functional in vivo in infected hamsters. We showed that SV40 has a broad tissue tropism, identified tissues that support viral DNA replication (kidney, liver, spleen, lung and brain), and found that kidneys are the major site for long-term persistence with livers a possible secondary site. One role for the miRNA appears to be to reduce viral loads in infected tissues. We also showed that viral miRNA does not appear to affect SV40 tumorigenicity in hamsters, the traditional assay for SV40 oncogenic potential.
Outbred Syrian golden hamsters (M. auratus) were obtained from Harlan Laboratories and were housed in the biohazard facility in the Center for Comparative Medicine at Baylor College of Medicine. The animals were maintained in accordance with established national guidelines as outlined in the Guide for the Care and Use of Animals [57] and the Animal Welfare Act. The Center for Comparative Medicine is fully accredited by the Association for Assessment and Accreditation of Laboratory Animal Care International (Animal Welfare Assurance Number A3823-01). The studies were approved (Protocol Number AN668) by the Institutional Animal Care and Use Committee of Baylor College of Medicine.
Procedures for intracardiac and intraperitoneal injections have been described [10], [16]. We have excellent survival outcomes (≥99%) with both routes of inoculation. Five- to 7-week-old female hamsters were used for intracardiac inoculations for the pathogenesis studies, whereas 3-week-old male and female hamsters were employed for intraperitoneal injections for the tumorigenicity studies. Animals were inoculated with 1×107 PFU of virus in 0.1 ml by the intracardiac route or 1×107 PFU of virus in 0.5 ml intraperitoneally. Control animals were inoculated with uninfected cell lysate under the same conditions. Animals were sacrificed at designated days p.i., at evidence of neoplasia or debility, or at the termination of experiments. Euthanasia involved isoflurane overdose and exsanguination by cardiac puncture. Necropsies were carried out on all the animals inoculated by the intracardiac route and some of those inoculated by the intraperitoneal route. Tissue samples were collected and were stored frozen or preserved in RNAlater.
Two natural SV40 strains from different phylogenetic groups were studied [43]. Reference strain 776 (complex RR) was isolated from an adenovirus type 1 vaccine seed stock [58]. SVCPC (simple RR) was recovered from a pediatric brain tumor [59]; the same virus has been detected in other human cancers [21], [59]–[61] and in the Russian oral poliovaccine [62]. The structures of the 776 and SVCPC viral RR are shown in reference [15]. The miRNA-negative mutant derived from 776-WT virus (776-SM1) was obtained from C. Sullivan [20]. The construction of recombinant virus SVCPC-776-WT was described previously [16]. This recombinant contains the RR, late region, the N-terminal part of T-ag, and the small t-ag coding region of strain SVCPC with the C-terminal part of T-ag (BstXI and BamHI restriction fragment) from 776-WT. Recombinant SVCPC-776-SM1 was constructed following the same strategy and contains the C-terminal T-ag fragment from 776-SM1.
Due to nucleotide polymorphisms in the miRNA region, the predicted sequence of the pre-miRNAs of SVCPC-WT and 776-WT differ slightly. Consequently, the amino-acid sequences of the T-ags of SVCPC-WT and 776-WT also slightly diverge. The miRNA-negative mutant derived from SVCPC-WT virus (SVCPC-SM2) was created by mutagenesis of 18 nucleotide positions in the pre-miRNA region analogous to that of 776-SM1. These mutations disrupt the hairpin structure of the SVCPC-WT pre-miRNA on the late strand while leaving intact the amino acid-coding potential of T-ag on the early strand (Figure S1) [20]. Three overlapping fragments covering the miRNA region were PCR-amplified with six mutagenic primers using SVCPC-WT or 776-SM1 DNA as a template. Three PCR primer sets were MA1 5′-TTCCTGGGGATCCAGACATGATAAG-3′ and MIRA2 5′-CTCCGCAGCCTTCGCAGTCCT-3′ on the SVCPC-WT template, MIRB1 5′-AGGCTGCGGAGCTTGAAACGAAC-3′ and MIRB2 5′-AGCCAGGAAAATGCTGATAAAAATG-3′ using the 776-SM1 template, and MIRC1 5′-AGCATTTTCCTGGCTGCTGTCATCATCA-3′ and SVPST 5′-AAAACACTGCAGGCCAGATTTG-3′ on the SVCPC-WT template. Three mutated fragments were subsequently joined together by overlapping PCR using outer primers. The resulting mutated fragment was cleaved with restriction enzymes PstI and BamHI and used to replace the corresponding part of the genome of SVCPC-WT also cleaved with the same restriction enzymes. The final miRNA-negative mutant virus was denoted SVCPC-SM2.
Each recombinant construct was sequenced to confirm that no accidental mutations had been introduced and was transfected into TC-7 cells to confirm its infectivity. The mutant viruses were shown to be defective for production of viral miRNAs using a test designed to detect early mRNA cleavage fragments that are generated by functional miRNA [20]. TC-7 cells were infected at a multiplicity of infection of 2 PFU/cell with each virus and total RNA was harvested at 65 hours p.i. using the RNAqueous-4 PCR kit (Ambion) and treated with DNase I. Total RNA was Northern blotted with a probe specific to the 3′ region of SV40 early mRNA (nucleotides 2681–2845, prepared by PCR), radiolabelled with α32PdCTP using the DECAprimeII kit (Ambion). The characteristic early mRNA 3′ cleavage fragment about 260-nt in size was generated by each of the three WT viruses (776-WT, SVCPC-776-WT, and SVCPC-WT) that produce intact miRNAs. In contrast, the miRNA-negative mutant viruses (776-SM1, SVCPC-776-SM1, and SVCPC-SM2) produced only full-length early transcripts without any detectable miRNA-generated 3′ cleavage fragment.
Virus stocks were prepared in TC-7 monkey kidney cells. Infectious virus titers were determined by plaque assay [63]. The numbers of viral genome copies in virus stocks were determined using an RQ-PCR assay.
Small pieces of frozen tissues were minced and digested overnight with Proteinase K and nuclei lysis solution. Proteins were precipitated and removed and then the DNA was precipitated with isopropanol, washed, resuspended in Tris-EDTA buffer (pH 8.0), and stored at −20°C [16]. Two or three random fragments from each tissue sample were processed and assayed independently; those values were averaged to represent the results from that tissue.
SV40 DNA was detected and quantified by an RQ-PCR assay [64]. Five µl of DNA sample were assayed in a final reaction volume of 25 µl using 40 cycles of amplification. The single-copy hamster vimentin gene was measured by RQ-PCR to evaluate the quality of each DNA extract and to normalize viral gene copy numbers to cell numbers [15]. Vimentin copies per reaction were divided by 2 to determine cell equivalents per reaction. A sample was omitted from viral analysis if less than 10,000 cell equivalents were amplified in a vimentin reaction. For those samples adequate for further analysis, RQ-PCR reactions were considered virus positive if ≥10 viral genome copies per reaction were detected. Viral loads are expressed as SV40 DNA copies per 104 cells. Total RNA was extracted from samples preserved in RNAlater (Ambion) using the RNAqueous-4 PCR kit (Ambion) with DNase 1 treatment. RNA was reverse transcribed as described and then quantitated by RQ-PCR [14]. As a control, vimentin RNA or 18S ribosomal RNA was amplified and used to normalize SV40 RNA. Data are expressed as mRNA copies per 106 vimentin molecules (liver and kidney tissues) or copies per 106 18S molecules (hamster tumors). Primer and probe sequences have been reported [14].
The number of SV40 genome copies per PFU varied among the four virus stocks for unknown reasons: 776-WT, 1700 genome copies/PFU; 776-SM1, 130 genome copies/PFU; SVCPC-WT, 1800 genome copies/PFU; and SVCPC-SM2, 590 genome copies/PFU. To directly compare viral loads in hamster tissues between virus strains (based on PCR assays which detect viral DNA copies), adjustments were made to account for the different number of viral genome copies in each 1×107 PFU inoculum given the hamsters. SV40 DNA levels determined by PCR for a given sample were normalized by dividing the number of observed SV40 DNA copies by the number of viral DNA copies (×10−9) inoculated for that virus (i.e., if a viral load value was “X” copies of SV40 DNA per 104 cells for an SVCPC-WT sample, “X” would be adjusted by dividing by 18). Normalized values were used when the dynamics of infection were compared between different viruses. Observed (non-adjusted) viral load values were used for analyses involving a single virus strain.
Portions of frozen livers and kidneys from infected or uninfected hamsters were coded and assayed without knowledge of sample identity. The samples were lysed in 1 mL of TRIzol reagent (Life Technologies) containing Lysing Matrix B tissue grinding tubes (MP Biomedicals) by subjecting the tubes to 3 cycles of a 1-minute duration of tissue grinding at the maximum speed using a Mini-Beadbeater-16 (Biospec Products) followed by 1 minute of cooling on ice. The lysed samples were transferred individually to an additional 4 mL of TRIzol followed by RNA isolation according to the manufacturer's standard protocol. To enrich for small RNA, gel fractionation was performed. Total RNA was first subjected to electrophoresis on a Tris-borate-EDTA-Urea-15% polyacrylamide gel. The gel portion between the bromophenol blue and xylene cyanol markers (approximately 10–35 nucleotides) was excised and then cut into smaller pieces. The gel pieces were soaked in 15 mL of 1 M sodium chloride solution to allow RNA elution overnight at 4°C, with gentle rocking. The eluted small RNA supernatant was transferred to a Vivaspin 15R 2000 MWCO centrifugal concentrator (BioExpress) and spun at 3000 g until the supernatant was concentrated down to approximately 500 µL. The small RNAs were precipitated overnight at −20°C in one-tenth volume of 3 M sodium acetate, pH 5.2, and an equal volume of isopropanol. The small RNA was reverse transcribed using the manufacturer's (Taqman, Life Technologies) primers (designed to recognize the SV40 3p miRNA) and the SuperScript III reverse transcriptase, using a modified version of the manufacturer's protocol: 16°C for 30 minutes, followed by 60 cycles of 30°C for 30 seconds, 42°C for 30 seconds, and 50°C for 1 second, and a final incubation at 85°C for 5 minutes. The reverse transcription product was used in the Taqman Small RNA assay for detection of the SV40 3p miRNA (Life Technologies) and the assay was performed on a ViiA 7 real-time PCR system (Life Technologies) according to the manufacturer's protocol. The linear range of the assay was determined using a dilution series of total RNA (from SV40 776-infected BSC-40 cells at an MOI of 10) as the standard. This analysis revealed linearity down to CT 39. We also estimated the sensitivity of our assay. The total number of cells assayed was held constant at 300,000 while altering the ratio of infected (776- infected BSC-40, MOI of 10) to uninfected cells. This resulted in an estimate of sensitivity ranging between 1 in 8,000 and 1 in 100,000 infected to uninfected cells. To minimize the chances of false positives, a conservative cutoff of CT 35 was chosen to score positive for miRNA detection. For comparison, we also calculated the SV40 3p miRNA and miR-Let7a (Taqman Life Technologies) levels in BSC40 African green monkey cells infected with SV40 (strain 776, MOI 10) and harvested RNA at 24 hours postinfection. The average of at least three independent technical replicates was used to calculate CT for all analyses.
Antibody responses to SV40 T-ag were detected and titered by indirect immunofluorescence as described [10]. SV40 neutralizing antibodies were measured using a plaque-reduction assay in TC-7 cells [10].
Focus-forming assays to quantitate SV40 transformation used primary mouse embryo fibroblasts infected with different strains of SV40 at 5 PFU/cell [16]. Replicate plates (usually 4–6) were harvested at 3 and 5 weeks p.i. and the number of transformed foci determined. For each WT and mutant virus pair, numbers of transformed foci per 1×105 cells infected were calculated as the percentage of WT foci at 5 weeks (set as 100%). The transforming frequencies overall were about 1 per 2000 infected cells.
Tissue fragments were minced, frozen and thawed two times, and the cell debris pelleted by centrifugation. The pellets were resuspended in 200 µl buffer and both the lysate supernatants and the lysate pellets were inoculated into TC-7 cell cultures. These cultures were incubated for 12 days, at which time the samples were harvested and assayed for infectious virus by plaque assay [63].
Logarithm base 2 (log2 transformed) values were used for evaluation of observed viral load data to reduce the effect of very low or very high values which might bias the mean. Statistical analysis indicated log normal distribution of the viral load data; the two-sample T-test was used to determine the difference in geometric means for adjusted viral loads between groups according to viral strain. The Wilcoxon rank-sum test was used to determine the difference in persistent infections at day 270 by different viral strains. The Z test for the comparison of proportion was used to evaluate differences in the frequency of tumor formation between groups exposed to different viral strains and the proportional hazards regression test to assess differences in survival times of animals that developed tumors between groups. Statistical analyses were performed using statistical software SAS version 9.3. A p-value of ≤0.05 was considered statistically significant.
|
10.1371/journal.pntd.0005953 | Estimating sensitivity of the Kato-Katz technique for the diagnosis of Schistosoma mansoni and hookworm in relation to infection intensity | The Kato-Katz technique is the most widely used diagnostic method in epidemiologic surveys and drug efficacy trials pertaining to intestinal schistosomiasis and soil-transmitted helminthiasis. However, the sensitivity of the technique is low, particularly for the detection of light-intensity helminth infections. Examination of multiple stool samples reduces the diagnostic error; yet, most studies rely on a single Kato-Katz thick smear, thus underestimating infection prevalence. We present a model which estimates the sensitivity of the Kato-Katz technique in Schistosoma mansoni and hookworm, as a function of infection intensity for repeated stool sampling and provide estimates of the age-dependent ‘true’ prevalence. We find that the sensitivity for S. mansoni diagnosis is dominated by missed light infections, which have a low probability to be diagnosed correctly even through repeated sampling. The overall sensitivity strongly depends on the mean infection intensity. In particular at an intensity of 100 eggs per gram of stool (EPG), we estimate a sensitivity of 50% and 80% for one and two samples, respectively. At an infection intensity of 300 EPG, we estimate a sensitivity of 62% for one sample and 90% for two samples. The sensitivity for hookworm diagnosis is dominated by day-to-day variation with typical values for one, two, three, and four samples equal to 50%, 75%, 85%, and 95%, respectively, while it is only weakly dependent on the mean infection intensity in the population. We recommend taking at least two samples and estimate the ‘true’ prevalence of S. mansoni considering the dependence of the sensitivity on the mean infection intensity and the ‘true’ hookworm prevalence by taking into account the sensitivity given in the current study.
| The World Health Organization (WHO) has defined a roadmap for schistosomiasis and soil-transmitted helminthiasis morbidity control and interruption of transmission with targets to be reached by 2025. Control efforts require reliable estimates of at-risk populations, number of infections, and disease burden estimates in population subgroups in terms of age and location. Intervention guidelines are based on insensitive diagnostic techniques, such as the Kato-Katz method and do not take into account the effect of sampling effort and infection intensity. Our proposed methodology estimates the infection intensity-dependent sensitivity and the ‘true’ age-prevalence of the blood fluke Schistosoma mansoni and hookworm. We also provide recommendations on the number of stool samples required and the methodology to be used to reliably estimate the ‘true’ prevalence of parasitic worm infections.
| Soil-transmitted helminthiasis (STH) and schistosomiasis are two of the most prevalent neglected tropical diseases with more than 1 billion and over 250 million people affected worldwide, respectively [1, 2]. Their collective global burden is 6 million disability-adjusted life years [3], with school-aged children at the highest risk of associated morbidity. Control efforts have intensified over the past 15 years, with preventive chemotherapy serving as the main pillar [4–6].
Information about the spatial and temporal distribution of STH and schistosomiasis are important to guide interventions. Moreover, it is necessary to know which age groups contribute most to transmission, in terms of helminth egg output, in order to effectively use the available resources. There are two approaches to obtain this information. First, large-scale, standardized studies including all age groups with intense sampling which, however, is difficult to pursue due to the high cost. Second, reanalysis of data from previously published studies. Inference is hampered by the paucity of quality data [7–9]. Individual-level data are usually not reported; instead, only the number of participants tested positive for a specific helminth infection is considered. In addition, diagnosis relies on the Kato-Katz technique [10], i.e., counting of helminth eggs in a small amount of stool. This approach, however, has a low, setting-dependent sensitivity, which is governed by variation in day-to-day production of eggs per worm, non-random distribution of eggs within a stool sample, decay of eggs in the sample due to methods and duration of the experimental procedure, transportation, and storage [11–15]. Collecting multiple stool samples over consecutive days increases the accuracy but there are no guidelines on the optimal number of samples [16, 17]. Consequently, the comparison of studies that employed different sampling efforts, which is necessary for monitoring progress of control programs, is hampered.
Statistical modeling can help studying the age-prevalence and its dependence on the diagnostic error but has been restricted by the aforementioned limitations that compromise the quality of the data [7, 18]. Although the qualitative shape of helminthiasis age-prevalence curves is known, there has been little progress in the application of quantitative transmission models, especially for STH infections [19–22]. Furthermore, the dependency of the intensity of the infection on the diagnostic sensitivity has been largely neglected. The negative binomial distribution has commonly been used to fit helminth egg count data. For example, De Vlas and Gryseels [12] and Levecke et al. [23] proposed models that separate the measurement process from the true underlying infection intensity distribution. However, none of the preceding models are able to infer on the dependence of the sensitivity of the Kato-Katz method for repeated stool sampling on infection intensity.
We developed a model for fecal egg-count (FEC) data, which quantifies the relation between sampling effort, infection intensity, and diagnostic sensitivity. The model separates the infected from the non-infected individuals and the measurement process from the infection status. Variability due to egg output, experimental conditions, and aggregation within a population are taken into account. We calculate the ‘true’ prevalence and other biological and transmission-related parameters based on the probability of false-negatives. Our model improves estimation of the age-related disease burden and provides inputs for mathematical transmission models.
This study consists of a secondary analysis of published data. Ethics approval, written informed consent procedures, and treatment of infected individuals have been described elsewhere [24–27].
We tested our model performing a secondary analysis of individual level FEC data from three separate studies in medium and high transmission settings in Côte d’Ivoire conducted in 1998, 2002, and 2011, respectively. All data used were from baseline surveys with no previous mass drug administration in the area. The studies took place in Fagnampleu [24, 25], Zouatta [26], and Azaguié [27]. Based on the Kato-Katz assay, hookworm prevalence varied from 11.4% to 59.0%, and mean or infected population based infection intensity from 280 eggs per gram of stool (EPG) to 396 EPG. For S. mansoni, prevalence varied from 35.6% to 76.3%, and infection intensity from 152 EPG to 307 EPG. Between two and four stool samples were collected and analyzed on consecutive days of a total of 1423 participants. Azaguié and Zouatta were surveys performed in the full age range from 0 to 90 years, while Fagnampleu only included school-aged children. Prevalence of Ascaris lumbricoides and Trichuris trichiura were too low to be analyzed. Summary measures are included in Table 1, a more detailed description can be found in S1 Appendix, and the individual level data used is included in S1 Table.
We utilized a hierarchical Bayesian model to address the objectives given in the introduction. Let Yij be the FEC, i.e., the number of helminth eggs found in sample j in individual i, ki the number of stool samples from individual i, and xi the age of individual i.
We assumed that a population consists of a proportion of infected individuals p, i.e., people that carry at least one pair of worms, and that of uninfected individuals. Thus p is interpreted as the prevalence. Each infected individual has a characteristic infection intensity λi, measured in units of mean eggs per sample and assumed to be distributed within the population according to a shifted gamma distribution, given by
λ i = v i + μ m v i ∼ Gamma ( μ f · α , α ) = α μ f · α Γ ( μ f · α ) v i μ f · α - 1 exp ( - α · v i ) (1)
with a mean number of eggs μf + μm in an infected individual, variance μ f α that corresponds to the aggregation of infection intensities, and hence, the aggregation of worms within the population. The shift parameter μm is the mean number of eggs per sample that can be expected from an individual carrying exactly one female worm and thus the minimal possible infection intensity. Direct inference on the worm load is not possible in this frame as the dependence on mean egg output is non-linear and not well known [28].
The process of taking ki samples from an infected individual i with infection intensity λi is modeled by a negative binomial distribution with mean λi and a variance given by λ i + λ i 2 / r. r reflects the additional variation, due to changes in the day-to-day helminth eggs output, the aggregation of eggs in stool, and the precise experimental procedure but not the within-population variation which is given by α. If a perfectly random distribution of the eggs and perfect measurement is assumed r → ∞, the measurement process becomes a Poisson process. By including the uninfected, the model is written as
Y i ∼ { ( 1 - p ) + p · NB ( 0 , λ i , r ) k i , if I i = 0 p · ∏ j = 1 k i NB ( Y i j , λ i , r ) , if I i = 1 (2)
which corresponds to a zero-inflated negative binomial model with p corresponding to the mixing proportion. NB is the negative binomial distribution and Ii the result of the Kato-Katz test over all samples from an individual. They are defined as follows:
Ii={0,ifmax(Yi)=01,ifmax(Yi)≠0NB(y,λ,r)=(y+r−1y)(λλ+r)y(rr+λ)r (3)
False-negatives are included in the model as repeated zero measurements for an infected individual. Thus, the sensitivity depending on the number of repeated measurements becomes
si[ ki ]=1−NB(0,λi,r)ki=1−(rλi+r)ki·r
(4)
where s is the sensitivity, and k and λ vary for each individual.
Low-rank thin-plate splines are used to study the age dependence of p, μf, r, and α. For a detailed derivation of the spline model, see Crainiceanu et al. [29]. The representation of p is
logit ( p i ) = β 0 + β 1 x i + ∑ m = 1 M u m ( x i - κ m ) 3 (5)
where Θ1 = (β0, β1, u1, …, uM)T is the vector of regression coefficients and κ1 < ⋯ < κM are the fixed knots. The other parameters can be represented analogously using logarithmic or linear spline models. The spline regression makes only a few very general assumptions about the shape of the curve, e.g., continuity and differentiability, and is therefore able to infer without requiring prior knowledge about the biology of the process, i.e., the transmission model.
The minimum eggs per sample μm is fixed to the average egg output of a worm divided by an average amount of feces per day, multiplied by the weight of a sample. For hookworm, μm is 5 eggs in a sample which corresponds to 120 EPG and for S. mansoni, μm is 0.03 eggs which corresponds to 0.72 EPG [30, 31]. We choose the following semi-informative priors for the model: gamma for r with mean 1 and variance 1; normal for log(α) with mean 0 and variance 1; gamma for μf with mean 2 and variance 4; normal for β0,1 with mean 0 and variance 1; normal for u1,…,M with mean 0 and variance τ, where τ is distributed as a gamma with mean 2 and variance 4. The results were not sensitive to the specific shape of the priors.
Bayesian inference was performed using Markov chain Monte Carlo (MCMC) simulations implemented in Stan [32]. Validity of the model was checked using simulated data. Models with splines on p, μf, r, and α were run to check for age dependence. μf, r, and α showed no significant age dependence and were set as independent of age for the simulations presented in the results section. μm was varied from 1 egg to 6 eggs for hookworm and from 0.01 eggs to 0.1 eggs for S. mansoni, which also showed no significant influence. The lower limit of 0.01 eggs per slide for S. mansoni corresponds to roughly 100 eggs in 500 g of stool. The upper limit of 0.1 eggs per slide corresponds to 1000 eggs per 500 g of stool therefore any value larger than 0.1 is most likely unrealistic for a single worm pair. The model was run with a total of 25 chains, with 20,000 iterations each, of which 2,000 where used as warm up and adaption, for each study and for each of the two infections separately. Convergence was achieved, and assessed using Gelman + Rubin diagnostics and visual inspection of the chains [33].
We applied a Bayesian hierarchical model to FEC data from three studies carried out in Zouatta, Azaguié and Fagnampleu in Côte d’Ivoire, as described in the “Materials and methods” section. Parameter posterior mean estimates and 95% Bayesian credible interval (BCI) are summarized in Table 1.
The three studies are from different hookworm transmission settings with observed prevalence of 11.4% and mean infection intensity of an infected individual of 396 EPG for Azaguié, 35.4% and 331 EPG for Zouatta, and 59.0% and 283 EPG for Fagnampleu. Based on our model, we estimated the ‘true’ hookworm prevalence at 14.3% (95% BCI 10.9–18.5%), 43.7% (95% BCI 38.6–49.2%), and 62.2% (95% BCI 56.6–67.6%), for Azaguié, Zouatta, and Fagnampleu, respectively. The estimated mean infection intensity does not significantly differ from one study to another and mean estimates ranged from 220 EPG to 262 EPG (see Table 1). Age-prevalence curves in Fig 1 from the three studies show similar features such as a steep increase from birth till an equilibrium is reached at ages of around 20 years for Zouatta, and 45 years for Azaguié. The prevalence stays constant till an age of about 60 years from where the rate of infection declines. For Fagnampleu only the initial steep increase is visible due to the fact that no individuals older than 15 years were included.
For S. mansoni, the Azaguié and Zouatta studies show similar transmission levels with an observed prevalence of 35.6% and 40.8% and observed mean infection intensity of 179 EPG and 152 EPG, respectively. In contrast, the study in Fagnampleu had a prevalence of 76.3% and a mean infection intensity of 307 EPG. We estimated a ‘true’ prevalence of 49.3% (95% BCI 40.4–61.2%) and a mean infection intensity of 132 EPG (95% BCI 101–167 EPG) for Azaguié, 59.6% (95% BCI 50.7–69.3%) and 104 EPG (95% BCI 84–128 EPG) for Zouatta, and 83.8% (95% BCI 78.3–89.3%) and 282 EPG (249–321 EPG) for Fagnampleu. The estimated age-prevalence curves displayed in Fig 2 show similar qualitative features. The prevalence increases up to a peak between the ages of 15 and 20 years, and subsequently declines slowly up to an age of 60 years, followed by a stronger decrease. The lower prevalence after the peak is not significant but it appears both in the Azaguié and Zouatta data.
For hookworm, the day-to-day variation given by r is consistent across study sites, ranging from 0.15 (95% BCI 0.13–0.17) to 0.25 (95% BCI 0.15–0.37) (see Table 1), indicating strong overdispersion. The aggregation of egg output within the population is also consistent across the studies with α estimates ranging from 0.19 (95% BCI 0.05–0.68) to 0.32 (95% BCI 0.06–1.23).
For S. mansoni the day-to-day variation is consistent across studies and significantly different from hookworm with values ranging from 0.83 (95% BCI 0.67–1.02) to 1.10 (95% BCI 0.80–1.46). The aggregation of infections within the population shows no significant differences between studies with α ranging from 0.05 (95% BCI 0.04–0.07) to 0.09 (95% BCI 0.05–1.13), which indicates a significantly higher variance than for hookworm.
For hookworm, the estimates of the diagnostic sensitivity of Kato-Katz did not vary between locations. Based on a single Kato-Katz thick smear, sensitivity estimates were in the range of 47% to 57%, for two samples obtained from different days from 72% to 81%, for three samples estimates were within the range of 85% to 90%, and for four samples around 95%. For S. mansoni, data from Azaguié and Zouatta revealed similar sensitivity estimates within the range of 48% to 59% for one Kato-Katz thick smear, 62% to 73% for two samples, and 69% for three samples. Fagnampleu has a higher sensitivity of 70%, 84%, 88%, and 91% for one, two, three, and four samples, respectively (see Table 1).
The sensitivity of the Kato-Katz technique for different infection intensities was calculated using eq 4 and is plotted in Fig 3. For hookworm the dependence on infection intensity is weak, e.g., only increasing from 40% to 55% from a very light infection of 120 EPG to a still light infection of 500 EPG. For moderate and heavy infections (>2000 EPG) the sensitivity did not significantly improve with infection intensity. However, the sensitivity can be greatly increased by examining several stool samples, e.g., for an infection intensity of 360 EPG the sensitivity can raise from 50% based on a single sample to 75% for two samples, and 92.5% for three samples.
The sensitivity was strongly associated with S. mansoni infection intensity. In particular, for very light infections (<5 EPG), it was below 50% even after three samples. For light infections (<100 EPG), it was still heavily dependent on infection intensity. For moderate infections (100–399 EPG), two samples gave a high sensitivity above 90%. Heavy infections (>400 EPG) were reliably detected (i.e. >99%) by testing two samples.
Fig 4 shows the overall sensitivity in a population with a day-to-day variation of r = 1.0 and a population aggregation of α = 0.07 as a function of the mean infection intensity in the population. For lower transmission settings with 100 EPG comparable to Zouatta, the sensitivity after four samples is still below 75%. However, sensitivity rose to more than 95% for a setting with a mean infection intensity of over 300 EPG.
We present a model which determines the relation between the sensitivity of the Kato-Katz technique and intensity of S. mansoni and hookworm infections. The model takes into account day-to-day variations in helminth egg output and within population aggregation of worms. Additionally, we were able to test various parameters for age-dependence, especially the age structure of the prevalence.
The overall sensitivity point estimates for hookworm corroborate with estimates derived from latent class modeling approaches. Tarafder et al. [34] give an estimate of 65% sensitivity in a setting with 35% prevalence where no infection intensity is given, and Nikolay et al. [35] estimated the sensitivity for one sample at 59.5%, for two at 74.2%, and for three samples at 74.3%. For S. mansoni in a setting with an observed prevalence of 33%, Lamberton et al. [36] predict a sensitivity of 29.6% for one sample, 51.9% for two samples, 70.4% for three samples, and 77.8% for four samples in agreement with our estimates for Zouatta and Azaguié. The same authors predicted a much higher sensitivity of 83.5%, 97.8%, and 100% for one, two, or three samples, respectively in a setting with 95% prevalence in agreement with our estimates for Fagnampleu. Glinz et al. [37] also predicted a sensitivity of 70% for a single Kato-Katz thick smear in a high transmission setting with prevalence close to 90%, which is again consistent with our estimates for Fagnampleu.
Our estimates are in agreement with those given in the literature. However, the added value of our modeling approach is that we can predict the sensitivity depending on infection intensity and the ‘true’ prevalence in various settings and we can understand the factors influencing the sensitivity for hookworm and S. mansoni.
There are three main parameters that directly influence the sensitivity, i.e., the minimum infection intensity, the day-to-day variation, and population aggregation. These parameters are specific for S. mansoni and hookworm. The minimum infection μm, which is fixed to 0.03 eggs per sample for S. mansoni and 5 eggs per sample for hookworm, is based on estimates of egg output for a single worm. The estimates carry a large uncertainty which were shown to be non-critical for the results. The low output of eggs per worm in S. mansoni makes it almost impossible to detect light infections with only a few female worms. The low sensitivity for light infections can also be seen in Fig 3 where the three S. mansoni curves show very low sensitivity when approaching 0 EPG. For hookworm, the output of a single female worm is about 5 eggs per sample, which already leads to a reasonable probability for detection.
The parameter r is determined by the day-to-day variation. The three studies agree on a value between 0.17 and 0.19 for hookworm and between 0.8 and 0.99 for S. mansoni. Both values conclusively contradict a Poisson process which has r → ∞ and could be interpreted as worms producing and excreting eggs randomly. Thus, our model clearly indicates that a worm produces eggs in a clustered fashion. A single pair of S. mansoni produces in the order of 100 eggs per day [30], while a female hookworm sheds around 10,000 eggs [31]. Thus, an infection that manifests itself with a similar egg count for both diseases indicates a much higher number of S. mansoni compared to hookworm. This difference in worm numbers has important implications, because the variation in egg output of a single worm is partly averaged out over the population of worms that produce eggs independently in an individual. The much lower variation in day-to-day output of S. mansoni can therefore be explained by the, on average, larger number of worms in an infected individual. The larger variation for hookworm leads to a lower sensitivity of the Kato-Katz technique for this helminth species for a similar infection intensity. Thus, the increase in sensitivity due to repeated sampling is much larger for hookworm (see Fig 3).
The population aggregation parameter α describes the distribution of worm output within the population. A small value indicates a relatively wider distribution, while a larger value indicates a more uniform distribution. The estimate of α has larger uncertainty for hookworm compared to S. mansoni, most likely due to the large day-to-day variation for the former species. Furthermore, α is overall larger for hookworm compared to S. mansoni indicating that egg output is more evenly distributed among individuals infected with hookworm than for S. mansoni. These results are in agreement with findings by Krauth et al. [15].
The severity of disease burden at a location is generally given by two parameters, prevalence and intensity of infections. If negative individuals are included in the calculation of the mean infection intensity the high number of zeros will skew the mean downwards, especially in lower prevalence settings. Thus, a strong correlation to the prevalence will be induced, making the mean infection intensity to be a mixed measure of prevalence and infection intensity. We decided to only include positive individuals in the mean infection intensity to separate the effects of the sensitivity on observed prevalence from that on infection intensity. Therefore, we decided to estimate the mean infection intensity from only positive individuals.
It is evident from our estimates that the sensitivity of the Kato-Katz technique in hookworm is dominated by the large day-to-day variation in egg output and has only a weak dependence on infection intensity. Thus, increasing the number of samples is an effective strategy to increase sensitivity even in low transmission settings for hookworm. In contrast, for S. mansoni the infections that give false-negative results are largely those with light intensity. Increasing the number of samples to more than two does only marginally improve the sensitivity because the sensitivity is limited by the low density of eggs and not the variation in excretion and production. However, taking two thick smears from the same samples at 100 EPG mean infection intensity would increase the sensitivity from 50% to 70% for a comparably low additional effort. For mean infection intensities below 100 EPG no directly representative data was included in this model study. Nevertheless, extrapolations to low prevalence and infection intensity settings are valid because the model considers individual-level data. In a high intensity setting there are still many individuals with low intensity infections, and therefore our model includes information on the full range of infection intensities. Thus, inference about a low intensity and prevalence population consisting primarily of light infections is possible without making unreasonable extrapolations. For infections with an intensity above 240 EPG, the strong relation between infection intensity and sensitivity suggests that the sensitivity for two samples is close to 100%, ensuring that the most heavily infected individuals are detected and can be treated. Still, the results indicate a possibility of bias when comparing different study sites. The infection intensity-dependent sensitivity will become increasingly important in the new era with the goal to eliminate STH and schistosomiasis [38, 39].
In the study sites of Zouatta and Azaguié, the observed mean infection of S. mansoni was moderate compared to Fagnampleu. Hence, it is conceivable that there was a comparably larger share of light infections that were missed due to the low sensitivity of the Kato-Katz technique, and therefore, the ‘true’ mean infection intensity will be even lower. Moreover, the difference between estimated and observed prevalence is also larger due to the higher number of missed cases. For Fagnampleu, the observed mean infection intensity and therefore the overall sensitivity are high. Thus, the estimates agree with the observations and the ‘true’ prevalence is only slightly higher than the one observed. For hookworm, the likelihood of correctly diagnosing a heavy infection is still larger than for a light infection. Accordingly, the difference between the observed and the estimated mean infection intensity is larger the fewer the number of samples taken.
Our estimates of the underlying ‘true’ age-prevalence for S. mansoni and for hookworm are in agreement with those obtained from transmission models [19, 40] and from latent class statistical models [41]. For example, the S. mansoni age-prevalence is comparable to those obtained by the Yang et al. [42] models, which differentiate between the influence of water contact patterns and the acquired immunity. However, due to large uncertainty in older age groups, our results do not allow choosing the transmission model, which resembles best. For hookworm, a peak shift, as proposed by Woolhouse [43], is clearly visible in our age-prevalence curve. The decreasing prevalence at older age is apparent in all locations but it has yet to be discussed in greater detail in the literature. It could indicate a significantly lower life expectancy for infected individuals although other explanations and confounding factors are conceivable but cannot be tested with the available data.
The proposed model succeeds in predicting the intensity-dependent sensitivity of the Kato-Katz technique directly from the day-to-day variation in helminth egg output. Hence, the model is able to explain the differences between the sensitivity of hookworm and S. mansoni. The sensitivity of Kato-Katz for hookworm is dominated by a high day-to-day variation. We recommend collecting at least two stool samples over subsequent days combined with the given sensitivity values to estimate ‘true’ prevalence. For S. mansoni infection the sensitivity is largely driven by light infections that are hard to detect by a single Kato-Katz thick smear. We also recommend collecting two samples due to almost perfect sensitivity for moderate and heavy infections and low benefit of additional samples for light infections. We predict that improving the sensitivity for S. mansoni can be achieved more cost effectively by increasing the number of Kato-Katz thick smears from the same stool sample instead of increasing the number of samples taken. Additionally, it is necessary to take into account the infection intensity-dependent sensitivity of Kato-Katz for S. mansoni when comparing data from several studies. Including the infection dependence becomes more important when close to elimination due to the larger changes in sensitivity of Kato-Katz with infection intensity.
A further consequence of the results is due to the fact that the guidelines of WHO are defined in terms of observed prevalence. An observed prevalence of e.g. 10% for S. mansoni, which is the lower limit for yearly MDA, is indicative of a ‘true’ prevalence of roughly 14%, 20%, and 29% for 200 EPG, 100 EPG, and 50 EPG, respectively. Hence, the observed prevalence is a measure of both, the ‘true’ prevalence and the infection intensity. We advise the disentanglement of these two components by defining thresholds separately for ‘true’ prevalence and infection intensity. The results also suggest that the current disease burden estimates underestimate the true prevalence.
The spline model for age-dependence used in this study can be replaced by appropriate transmission models to determine which age groups should be treated and how frequently that has to happen to increase the intervention effectiveness. The model can be further extended to analyze studies with multiple Kato-Katz thick smears performed per stool sample and thus separate day-to-day from within-sample variation. This would enable us to address the question of how repeated testing of the same sample compares to taking several samples in order to reduce cost and increase compliance.
|
10.1371/journal.pcbi.1005736 | Differential temperature sensitivity of synaptic and firing processes in a neural mass model of epileptic discharges explains heterogeneous response of experimental epilepsy to focal brain cooling | Experiments with drug-induced epilepsy in rat brains and epileptic human brain region reveal that focal cooling can suppress epileptic discharges without affecting the brain’s normal neurological function. Findings suggest a viable treatment for intractable epilepsy cases via an implantable cooling device. However, precise mechanisms by which cooling suppresses epileptic discharges are still not clearly understood. Cooling experiments in vitro presented evidence of reduction in neurotransmitter release from presynaptic terminals and loss of dendritic spines at post-synaptic terminals offering a possible synaptic mechanism. We show that termination of epileptic discharges is possible by introducing a homogeneous temperature factor in a neural mass model which attenuates the post-synaptic impulse responses of the neuronal populations. This result however may be expected since such attenuation leads to reduced post-synaptic potential and when the effect on inhibitory interneurons is less than on excitatory interneurons, frequency of firing of pyramidal cells is consequently reduced. While this is observed in cooling experiments in vitro, experiments in vivo exhibit persistent discharges during cooling but suppressed in magnitude. This leads us to conjecture that reduction in the frequency of discharges may be compensated through intrinsic excitability mechanisms. Such compensatory mechanism is modelled using a reciprocal temperature factor in the firing response function in the neural mass model. We demonstrate that the complete model can reproduce attenuation of both magnitude and frequency of epileptic discharges during cooling. The compensatory mechanism suggests that cooling lowers the average and the variance of the distribution of threshold potential of firing across the population. Bifurcation study with respect to the temperature parameters of the model reveals how heterogeneous response of epileptic discharges to cooling (termination or suppression only) is exhibited. Possibility of differential temperature effects on post-synaptic potential generation of different populations is also explored.
| Focal cooling of the epileptic brain region has been shown to consistently suppress epileptic activity and it is hoped that this treatment can be developed in the future into an implantable cooling device. However, it is still not clearly understood how cooling suppresses epileptic activity. This study uses a computational approach to identify and investigate possible mechanisms. First, we used a neural mass model to reproduce epileptic discharge activity. Next, we simulate the effect of cooling by introducing temperature dependence in the model. Based from evidences reported from in vitro and in vivo studies, we formulated two temperature-dependent mechanisms that can reproduce the effect of cooling on the epileptic discharge activity. Parameter estimation of the model was performed using EEG recordings of focal brain cooling experiments with rats in vivo. Our model involves a synaptic mechanism that results in a reduced frequency of discharges and an intrinsic excitability mechanism that compensates such reduction in frequency of discharges resulting in persistent discharges during cooling but suppressed in magnitude. The temperature dependence is in the form of Q10 temperature coefficients which determine whether suppression or termination of epileptic discharges can be achieved.
| The World Health Organization identifies epilepsy as one of the most common neurological diseases affecting approximately 50 million people across all ages across the world [1]. According to the International League Against Epilepsy, a patient has epilepsy if he has had a seizure and his brain activity demonstrates a pathologic and enduring predisposition to have recurrent seizures [2]. Because of the risks involved with unanticipated seizures, treatment of the disease is required to improve long-term quality-of-life of the patients. Antiepileptic drugs such as anticonvulsants are usually given as first line treatment after being diagnosed with epilepsy. Pharmaceutical researches continually seek antiepileptic drugs that are more effective and have less side effects [3, 4]. However, 20%-40% of patients diagnosed with epilepsy are found refractory to antiepileptic drug treatment [5, 6]. Thus, alternative treatments are still being sought after [7, 8]. Surgical treatment is done by performing a resection of the epileptic foci of the brain. Absolute remission however is not guaranteed, let alone possibilities of unintended outcomes since the method is largely invasive [9, 10]. Although the success rate of surgical treatment is high, limitation in indication and cost significantly hinder intractable epilepsy patients in acquiring it. Another increasingly attractive treatment option involves electrical neurostimulation of specific neural region such as vagus nerve stimulation and deep brain stimulation [11, 12].
In the previous decade, focal cooling of the epileptic brain area has been pursued as an alternative therapeutic treatment for epilepsy and other seizure-inducing brain injuries [13–15]. Studies in animals have shown that reversible cooling to a temperature as low as 15°C using an implantable cooling device is able to terminate epileptic discharges without affecting the normal brain tissue [16–19]. Earlier experiments even noted that focal cooling of the cortex for one hour above 0°C did not induce any irreversible histological change or motor dysfunction [20]. Focal cooling at 25°C was also demonstrated to suppress epileptic discharges in a human brain [21]. Epileptic seizures arising from post-traumatic brain injuries were also shown to be suppressed and can be further prevented by moderately cooling the brain down by a temperature reduction of 2°C [22]. In other studies, focal brain cooling has found potential use for treatment of other brain diseases such as ischaemia, stroke, and neonatal encephalopathy [23–26]. The ultimate goal especially for epilepsy studies is to develop a technique for epileptic seizure suppression by a temperature control, when detected, via an implantable cooling device as a solution for thermal neuromodulation. This is feasible if we have precise knowledge of how temperature can suppress or terminate seizures. While temperature effects on physiological properties of animal neurons have been well-studied in vitro [27–31], mechanisms of how cooling suppresses epileptic discharges especially in vivo are still not clearly understood.
In this study, we aim to identify prospective mechanisms and investigate them using a computational modelling approach. Neural mass models have been widely utilized to study brain activities [32–37] and gain relevant physiological insights from them. The model introduced by Wendling et al. [34] in particular was shown to produce different types of brain activities similar to intracranial EEG recordings. We explore prospective mechanisms of cooling on epileptic discharges by introducing temperature dependence in the neural mass model of Wendling et al. in light of findings observed in in vitro and in vivo experiments published in literature. In particular, changes in synaptic dynamics were reported from in vitro cooling experiments such as reduction in the efficacy of neurotransmitter vesicle release [38], loss of dendritic spines [39] and reduced glutamate concentrations [40, 41], suggesting a possible synaptic mechanism. A recent study with patients with intractable epilepsy also reports reduced extracellular glutamate and GABA concentrations during focal brain cooling [42]. We then formulated temperature dependence in our chosen neural mass model by introducing a temperature factor in the post-synaptic impulse response function. Parameter estimation of the model is performed using EEG recordings from in vivo cooling experiments on an animal model of epilepsy. Although the model is able to reproduce termination of epileptic discharges reported in in vitro studies [43, 44], the results of modeling our experimental data (in vivo) reveal that this synaptic mechanism is not sufficient to explain epileptic discharges that are persistent during cooling although suppressed in magnitude. We propose that another mechanism is required to compensate the effect of this synaptic mechanism to be able to reproduce observed suppression of epileptic discharges during cooling in terms of reduction in both frequency and magnitude of discharges. We suggest some biological plausibility of this compensatory mechanism based from published results from cooling experiments. The temperature dependence is in the form of a temperature coefficient (Q10) which represents the factor by which the rate of a process increases for every ten-degree rise in the temperature at which it takes place [45]. In this study, the Q10 values determine whether suppression or termination of epileptic discharges can be achieved. Such heterogeneous response of epileptic discharge activity to cooling is revealed by bifurcation patterns with respect to the temperature parameters of the model.
All experiments were performed according to the Guidelines for Animal Experimentation of Yamaguchi University School of Medicine. The animals were anesthetized with urethane (1.25 g/kg, i.p.). Lidocaine, a local anesthetic, was applied at pressure points and around the area of surgery.
Focal brain cooling experiments were performed at Yamaguchi University School of Medicine. In this study, we utilized their data for parameter estimation of our model. Details of the experiments can be found in [46]. Briefly, anaesthetized male Sprague-Dawley rats were induced with epilepsy using Penicillin G potassium. Continuous EEG recordings of the epilepsy-induced region of the brain were made before and during cooling. An Ag/AgCl electrode for recording EEGs (Unique Medical Co., Fukuoka, Japan) was positioned stereotactically 2 mm below the cortical surface at the left sensorimotor cortex just beneath the cooling device. Five different rat experiments each were done at cooling temperatures 25°C, 20°C, and 15°C. To remove high frequency components and also match the represented frequencies in the model, the raw recordings underwent a 40-Hz low-pass filter using a fifth order Butterworth filter in Matlab. One-minute steady-state intervals before and during cooling were identified by an expert and were taken from the filtered data for the study. For the model estimation procedure, first, the data is further downsampled to 2kHz corresponding to a step size of 0.5 ms in the simulation. Next, both the downsampled data and the simulated EEG are normalized by dividing by their respective standard deviations of activity before cooling, thus, they are reported in arbitrary units (au) unless otherwise stated.
Fig 1 shows a summary of the preprocessed data in which we concatenated one-minute steady state activities before and during cooling. Suppression of epileptic discharges during cooling was observed especially with 15°C cooling temperature (Fig 1). Epileptic discharges were suppressed in terms of magnitude (lower magnitude during cooling) in all cases. In most cases, frequency of epileptic discharges is lower during cooling although slightly higher in some cases. The average magnitude and frequency of epileptic discharges before and during cooling are summarized in Fig 2 with error bars indicating minimum and maximum values from five rats. In general, we can say that epileptic discharges are suppressed during focal cooling at all three cooling temperatures. Surprisingly, significant termination of epileptic discharges was observed only in two out of five rats with 15°C cooling temperature compared to most in vitro recordings reported in literature; epileptic discharges were generally persistent during cooling from these in vivo recordings.
Different intracranial EEG activities such as spike-wave discharges and low-voltage high-frequency activity, have been widely explained using neural mass models—a class of models based on a mean-field approximation of the activity of a population of neurons. Neural mass models involve two major processes described by two functions: a firing response function and a post-synaptic impulse response function. The firing response function approximates the average firing rate of a population in response to an average input potential (the average membrane potential of the population). Assuming a unimodal distribution of threshold potentials, the firing response function of a population of neurons can be described by a sigmoid function [47] given by
S ( v ) = 2 e 0 1 + e ( v t h - v σ t h ) , (1)
where vth is the average threshold potential at which the population fires at half the maximum firing rate e0. The steepness of the sigmoid curve 1/σth is inversely related to the variability in thresholds of excitation of neurons in the population [47]. On the other hand, the average post-synaptic potential (PSP) input of a neuronal population to other populations to which it provides excitation or inhibition is given by the convolution of the post-synaptic impulse response function h(t) of the population and its average firing rate u(t). Originally, the post-synaptic impulse response function is modelled using a sum of two exponentials [32] as compared from experimental data but was later simplified to
h X ( t ) = G X g X t e - g X t ; t ≥ 0 , (2)
where GX is the average post-synaptic gain and gX is the reciprocal of the average synaptic time constant of population X. Finally, the convolution vX(t) = hX(t) * u(t) is equivalent to the solution of the following second-order differential equation using Green’s Formula [48]:
v X ′ ′ + 2 g X v X ′ + g 2 v = G X g X u . (3)
The primary cell population also receives additional noisy input from subcortical afferents or other neural masses which makes the differential equation stochastic. Such can be solved numerically using stochastic methods such as Euler-Maruyama scheme. Finally, the average membrane potential of a population, which is the input to Eq (1), is taken as the weighted summation of the average post-synaptic potentials of the afferent populations (inhibitory populations have negative contribution). The weights are determined by the number of synaptic connections. The average membrane potential of the primary cell population is taken as representative of cortical EEG activity [32].
Different neural mass models vary in terms of the types of neurons that comprise a population and the interconnections among the populations (feedback loops). Da Silva et al. [32] tried to explain alpha rhythm of brain activity by considering two populations: excitatory thalamocortical neurons as primary cell population and and a population of inhibitory interneurons. Jansen and Rit [33] extended this model using pyramidal cells as the primary excitatory neurons and two types of interneurons—excitatory and inhibitory. They also estimated the relations among the number of synaptic interconnections among the neuronal populations using animal records of cortical synapses found in literature. Wendling et al. [34] further differentiated slow and fast inhibitory interneurons based on the studies of [49, 50] based from hippocampal connections. In their model, slow inhibitory interneurons project to the dendrites while fast inhibitory interneurons project to the soma or near the soma of pyramidal cells. Moreover, slow inhibitory interneurons provide inhibition to fast inhibitory interneurons. Although the model was patterned after neuronal connections in hippocampus, similar architecture has been seen in the neocortex (see [51] for an extensive review). The block diagram of the model is shown in Fig 3. The parameters of the model are summarized in Table 1 together with the standard values adopted in this study.
Wendling et al. showed that their model is able to capture different brain activities observed in intracranial EEG recordings. By fixing the value of average excitatory synaptic gain, an activity map (Figure 4 of [34]) shows regions of different brain activities by varying the average synaptic gains of slow and fast inhibitory neuronal populations. They used their model to explain that fast epileptic activity can arise due to impaired GABAergic inhibition by slow inhibitory interneurons. They demonstrated this by estimating average synaptic gains in the model from intracranial EEG recordings of temporal lobe epilepsy (Figure 5 and 6 of [34]). In this study, we used the same model and show that it strongly captures the discharge activity of the animal model of epilepsy used in the experiments.
In this study, we try to explain how cooling works in suppressing epileptic discharges by introducing temperature dependence in the neural mass model of Wendling et al. particularly for epileptic discharges. Our formulation starts with reduction in concentration of neurotransmitters as reported in in vitro studies. We model this effect as an attenuation factor in the post-synaptic impulse response function particularly the average synaptic gain variable. Specifically, we assume a temperature dependence in terms of a Q10 factor as follows:
h X ( t ) = Q 10 , s y n ( T - T 0 ) / 10 G X g X t e - g X t ; t ≥ 0 . (4)
Here, T0 is the baseline temperature which is 31°C in the experiments. This temperature dependence attenuates the average synaptic gain and thus reduces the average PSP (Fig 4) which makes up the average membrane potential of the population to which it provides excitation or inhibition. For excitatory and slow inhibitory interneurons, their average membrane potentials are solely contributed by the average PSP from pyramidal cell population, thus, are also attenuated and consequently yield reduced firing frequency. For the pyramidal cell population and fast inhibitory interneurons, negative inhibitory PSP contributes to their average membrane potential. If the weighted (in terms of synaptic connections) effect of temperature on inhibitory PSP is less than that on excitatory PSP, a net decrease in average membrane potential results. With the parameter values chosen in the model (Table 1), this is more likely the case. In Fig 5, we can see that as Q10,syn is increased from unity, frequency of discharges during cooling is decreased until termination. However, the value of Q10,syn at which termination is nearly achieved (Q10,syn = 1.085) does not significantly attenuate PSP magnitude (Fig 4), consequently the magnitude of isolated discharges. In contrast, persistent discharges were observed during cooling in the experiments (Fig 1). These are suggestive that another mechanism is involved. To model persistent discharges during the cooling period, we conjecture that the reduction in the average frequency of firing caused by the first temperature dependence should be compensated. This can be achieved through the firing response function negating the effect of Q10,syn (see Discussion). A second temperature dependence is thus put forward involving a reciprocal Q10 factor multiplied to the average membrane potential:
S ( v ) = 2 e 0 1 + e ( v t h - Q 10 , i n t - ( T - T 0 ) / 10 v σ t h ) . (5) Fig 4 illustrates the effect of this temperature dependence in the original firing response curve. The modified firing response curve is translated to the left and has steeper slope. In summary, two temperature parameters are introduced in this study—Q10,syn and Q10,int. The latter part of this study also looks at the possibility that Q10,syn varies for different populations in their respective PSP generation.
Since the cooling experiments were performed on five rats, model parameters were estimated per rat using three cooling temperatures. Modified from [52], the objective function involved in the estimation is given by
J ( θ ) = ∑ E I D I + E E f f M a g + P ( θ ) , (6)
where Ex is the mean absolute percentage error (MAPE) of feature x computed as |xmodel − xdata| / |xdata|. The summation is over the three cooling experiments per rat. The features used in the estimation are the average inter-dischrage interval (IDI) over the one-minute series and the effective magnitude (EffMag) of epileptic discharges. IDI is computed as
I D I = 1 N D ∑ i = 1 N D - 1 t i + 1 - t i , (7)
where ti is a time at which a discharge (exceeding three standard deviations of the activity) occurs, and ND is the number of discharges within the one-minute activity. EffMag, on the other hand, is defined as
E f f M a g = P 99 - P 1 , (8)
where Pn denotes nth percentile of the activity. A penalty term P(θ) is also included in the objective for the estimation of the temperature parameters of the model from the epileptic discharge activity during cooling:
P ( θ ) = K ( [ max { v D C } - max { v B C } ] + + [ min { v B C } - min { v D C } ] + ) , (9)
where [⋅]+ = max{0, ⋅}, {v} is the simulated discharge activity centered with respect to the baseline, and K is penalty strength set to 1000. This term imposes the constraint that the range of discharge activity during cooling (DC) is contained within the range of the discharge activity before cooling (BC), that is, epileptic discharges are indeed suppressed during cooling. Since the model is stochastic, ten different simulations were taken for each set of parameters from which the MAPE is computed against the experimental data. Finally, after we are able to narrow down the parameter space to optimize the objective function, a global search is employed [53]. We used Dividing Rectangle (DiRect) method [54], a deterministic global optimization method that is less computationally expensive than stochastic evolutionary methods such as Genetic Algorithm which was used in [52]. Moreover, estimation was performed using a one-minute steady-state activity in contrast to dynamic estimation procedures such as Kalman Filtering [55] and Dynamic Causal Model [56].
A two-part estimation is performed for each experiment. The first part estimates the parameters of the Wendling model (no temperature-dependent parameters) that describes the activity of epileptic discharges before cooling. The second part estimates the temperature-dependent parameters (Q10 factors) during cooling using the result of the first part describing the pathological activity of the brain. Simultaneous estimation of all model parameters (before and during cooling) can be done, however, the two-part approach circumvents search issues in high-dimensional space. Furthermore, to address possible over-fitting, estimation of the Q10 values was done using the first 40 seconds of the one-minute activity during cooling. The next 20 seconds of the activity were used for validating the model estimates from which statistical tests are performed.
It is generally accepted that epileptic activity results from changes in excitation-inhibition ratio. In the neural mass model, keeping the average excitation gain constant, excitation-to-inhibition ratio increases as GSIN or GFIN is decreased thereby simulating epileptic discharge activity. Exploration of the model shows that EEG recordings from the animal model of epilepsy used in the study is best explained by high average fast inhibitory gain GFIN and low average slow inhibitory gain GSIN (Table 2). This is consistent with previous findings that epileptic activity can arise when dendritic inhibition is impaired [34]. Fig 6(a) shows a reproduction of epileptic discharge activities before cooling for two of the five rats. We observe that lower values of GSIN reproduce a discharge activity that is asymmetric with respect to baseline while higher values of GSIN reproduce a discharge activity that tends to be symmetric with respect to baseline. On the other hand, increasing both GSIN and GFIN reduces the frequency of epileptic discharges by effectively reducing the average membrane potential of the primary cell population which is basically the simulated EEG.
The estimation of average slow inhibitory gain and fast inhibitory gain of Wendling et al. model was aimed to reproduce epileptic discharge activity recorded from the animal model of epilepsy used. Next, we estimate the parameters involved in the temperature dependence of the model from the activity during which focal cooling is applied in the epileptic brain area. To assess our temperature-dependent formulation, three models were estimated from the experimental data namely: SYN (synaptic mechanism only: estimate Q10,syn with Q10,int = 1.0), INT (intrinsic mechanism only: estimate Q10,int with Q10,syn = 1.0), and SYN_INT (synaptic and intrinsic mechanisms: estimate Q10,syn and Q10,int). The results of the estimation were compared to no-temperature dependence (NTD) model (Q10,syn = 1.0, Q10,syn = 1.0). As discussed earlier, SYN captures changes in the frequency of epileptic discharges but not their magnitude (Fig 5). On the other hand, INT, as expected, yields estimates that are almost unity (like in the case of NTD) since the model does not have anything to compensate for having Q10,syn = 1.0, i.e. no changes in average PSP yield no changes in the average firing rate. These suggest that temperature dependence in the post-synaptic impulse response function or firing response function alone does not capture the effect of cooling on the epileptic discharges (Fig 7). In fact, when both functions have temperature dependence as formulated (SYN_INT), we see that suppression of epileptic discharges is reproduced. Fig 8 shows the boxplots of the mean absolute percentage error (MAPE) of the different models from fifteen cooling experiments. Note that the MAPE are computed from the last twenty seconds of the epileptic discharge activity during cooling which is apart from that used for the estimation (see Materials and Methods). A Wilcoxon signed rank test shows that SYN_INT is significantly different from NTD model (p = 0.0034).
It is also interesting to look at the estimated values of Q10,syn and Q10,int using SYN_INT model (Table 3). We can clearly see that Q10,int is only slightly less than Q10,syn. This is consistent in all estimations performed from experiments on five rats. We also performed estimation of Q10 factors from each cooling experiment per rat where we find cases in which Q10,int is slightly greater than Q10,syn. These cases correspond to experiments where there are slight increases in the frequency of epileptic discharges during cooling. However, in the results that we present here, Q10 factors are estimated from three cooling experiments per rat which yield Q10,int values that are all slightly less than Q10,syn.
Fixing Q10,syn at 1.8, we vary Q10,int from 1.0 to 2.0 at intervals of 0.01 and performed ten simulations of SYN_INT model with different random generator seeds. We find that the magnitude and frequency of simulated activity during cooling exhibit bifurcation behavior for different temperatures (Fig 9). There are three apparent bifurcation regions found for cooling temperatures 15°C and 20°C. From baseline activity, a bistable region occurs at around Q10,int = 1.5 and vanishes at around Q10,int = 1.66 going back to baseline activity until a sudden transition to discharge activity at around Q10,int = 1.8 which is the same value at which Q10,syn is fixed. The results of estimation from experiments lie around the third region where Q10,int values are only slightly less than Q10,syn values. This region corresponds to termination of epileptic discharges or suppression of epileptic discharges to a fixed magnitude. The bistable region, on the other hand, correspond to two possible activities depending on initial condition of the simulation- a baseline activity and an activity characterized by low-amplitude high frequency oscillations. This region, however, was not realized in the experiments. Hypothetically though, this suggests that seizure may occur with cooling when the compenstatory mechanism that involves the intrinsic excitability of neurons operates with Q10,int values in this region. This bistable region vanishes at weaker cooling temperatures (Fig 10) indicating that such possibility of seizure may be prevented. Similar pattern of bifurcation is also observed with a bistability region that is wider at higher values of Q10,syn and vanishes at lower values of Q10,syn (Fig 10).
To gain more insight about the bifurcation behavior observed in the model, we performed a numerical continuation of the deterministic version of the model (standard deviation of input is zero) using MatCont [57]. Similarly, we fixed Q10,syn at 1.8. Continuing from a fixed point with Q10,int = 1.0, two saddle node bifurcations are found at around Q10,int = 1.7996 and Q10,int = 1.1702 (Fig 11(a)). From the second bifurcation point, a Hopf bifurcation is found at around Q10,int = 1.566175 with negative first Lyapunov coefficient. This implies that a stable fixed point transitions into a stable limit cycle. These bifurcation points explain the observed bistable region in the original stochastic model above where low-amplitude high-frequency oscillations or a baseline activity can be observed depending on the initial state of the system. (Note that stationary state in the noiseless model corresponds to baseline activity in the stochastic model.) Furthermore, continuing from the Hopf bifurcation point, a limit point of cycles (LPC) is found at around Q10,int = 1.68. A LPC is a saddle node bifurcation for periodic orbits where two limit cycles coalesce and annihilate each other. This explains the recovery of stationary state until the first bifurcation point at which the system exits the bistable region and goes back to stable periodic orbits (discharge activity). The transition point observed in the stochastic model (termination to suppression of discharge activity) is then a sudden jump from baseline activity resulting in a magnitude of suppressed discharge activity that is proportional to the width of the hysteresis loop for a particular temperature and does not gradually increase from the magnitude of a baseline activity. At weaker cooling temperatures, such bifurcation is not observed at least in the physiologically explicable region of Q10 values.
We also explored the possibility that cooling has differential effect on PSP generation of different neuronal populations. We investigate this by assuming that Q10,syn is not homogeneous for different populations with different average synaptic gains. (Q10,int is not differentiated across different subpopulations as we assumed that the temperature effect is the same across different populations in their intrinsic excitability mechanisms.) SYN_INT assumes homogeneous effect of cooling across different populations. Two more models were estimated to account for the possibility of such differential effect of cooling. In EXC_INH, we assume differential effect of cooling on excitatory and inhibitory PSP generation involving production of glutamate and GABA respectively. In EXC_SIN_FIN, we further assume differential effect of cooling on slow and fast inhibitory PSP generation involving slow GABA and fast GABA respectively. Estimation of these two models were also found to yield significant difference from NTD (p = 0.0034 and p = 0.0034 respectively). The two models however are not significantly different from SYN_INT (p > 0.01, Fig 8). It is interesting to note that EXC_SIN_FIN is able to capture termination of epileptic discharges from rat 1 under cooling temperature of 15°C which is roughly captured using SYN_INT or EXC_INH. Estimated Q10 values in Table 3 present some general observations. In EXC_INH model, Q10,syn values are now slightly less than Q10,int values except for rat 1 in which termination of epileptic discharges was observed. In EXC_SIN_FIN, higher Q10,syn,FIN values were estimated especially with rats 3 and 4. On the other hand, lower Q10,syn,EX values are observed for rats 1 and 2 in which termination of epileptic discharges were found while lower Q10,syn,SIN values for rats 3 and 4 in which epileptic discharges are only suppressed during cooling. These observations suggest that termination or suppres sion of epileptic discharges can result from different synaptic responses of different neuronal populations to cooling. Figs 12 and 13 show how the different models reproduce termination or suppression of epileptic discharges in rats 4 and 1, respectively.
Finally, it can also be observed that the estimated Q10 values are between 1.7 and 2.0 except those estimated from rat 5 in which case the estimated values are less than 1.2. The estimation result from rat 5 can be substantiated by observing the activities during cooling of rat 5 at different temperatures showing less evidence of suppression of epileptic discharges (Fig 1).
Our study confirms the ability of Wendling et al. model to capture different brain activities particularly epileptic discharge activity induced in the animal model of epilepsy used. After a brute-force search in the GSIN and GFIN space (with GPY = GEX = 5.0), we find that the epileptic discharge activities from our animal model of epilepsy are best estimated in the range [24.0, 31.0] mV for GSIN and [80.0, 110.0] mV for GFIN, the latter of which is not explored in the original model. Alternatively, we can keep GFIN in physiological range [40.0, 60.0] mV but would entail that the number of synaptic connections from fast interneurons to pyramidal cells is twice than the standard value or that the maximum average firing rate of fast inhibitory interneurons is twice than that of the others (see Eq 3). This is still consistent with the findings of Wendling et al. [34] suggesting that impaired dendritic inhibition alters excitation-inhibition balance giving rise to rhythmic discharge activity capturing the effect of Penicillin G potassium in cortical tissues inhibiting GABA receptors [58]. Nevertheless, the estimated parameters indicate that our animal model of epilepsy can be best explained by much lower dendritic inhibition and much higher perisomatic inhibition compared to the standard range of values reported. High GSIN values in fact supports [59] which reported high somatic inhibition together with impaired dendritic inhibition in experimental epilepsy. Meanwhile, asymmetric epileptic discharge activities with respect to baseline activity as seen from experiments with rat 3 can be reproduced with lower value of GSIN (25.012 mV) and higher value of GSIN (101.44 mV). On the other hand, symmetric discharge activity with respect to baseline is observed when dendritic inhibition is increased. Fig 6(b) illustrates that this symmtery (asymmetry) of the discharge activity (which is the summation of the PSP from excitatory and inhibitory interneuronns) is largely due to the PSP response of excitatory interneurons showing faster (slower) repolarization while the PSP responses of the inhibitory interneurons do not show significant changes.
Neurotransmitters play a central role in the generation of PSP [60]. They are released in response to Ca2+ influx after depolarization of pre-synaptic terminal and bind to their receptor molecules at the post-synaptic membrane opening or closing ion channels thereby generate excitatory or inhibitory PSP. It has long before suggested that neurotransmitter release has temperature dependence which causes changes in PSP generation [61]. This was confirmed by experimental observations of reduced efficacy of neurotransmitter vesicle release and reduced extracellular glutamate concentration during cooling [38, 40] that imply lower neurotransmitter concentration at the synapses to bind at the post-synaptic receptor and generate PSP. In light of this, it was straightforward to assume a temperature dependence on the post-synaptic impulse response function in a neural mass model. Similar to temperature-dependent formulation of Hodgkin-Huxley type neurons [62, 63], temperature dependence in Wendling et al. model is modelled using a temperature coefficient given by a Q10 factor. This factor accounts for mean-field effect of temperature to several processes occurring during PSP generation across the neuronal population. For example, diffusion of neurotransmitters, Ca2+, and receptor proteins [64–66] are slowed down at different rates at decreasing temperatures affecting efficacy of neurotransmitter vesicle release and the binding of neurotransmitters at the post-synaptic terminal receptors which regulate the activities of specific ion channels. Simply, Q10,syn is added to the post-synaptic impulse response function and can be interpreted as direct attenuation of the average post-synaptic gain of the population or synaptic conductance of one neuron. This yields lower average PSP values when temperature is decreased from a baseline temperature. This decreases or increases the average membrane potential of the populations to which the population provides excitation or inhibition respectively. Reduced average membrane potential yields lower frequency of firing. In fact, we saw that termination of epileptic discharges results when the firing frequency approaches zero with Q10,syn ≈ 1.085 with nonsignificant decrease in the magnitude of isolated discharges. In contrast, what was actually observed from experiments is that epileptic discharges are persistent during cooling but suppressed in magnitude. This is not reproduced by the model because of the nonlinearity of the firing response function. A Q10 value of 1.085 does not significantly suppress the magnitude of discharges but its effect on attenuated PSP responses significantly lowers the firing rate of the receiving population. Interestingly, in some cases in the experiment, slight increases in frequency of epileptic discharges were observed (Fig 1). These lead us to assume that a concomitant mechanism plays a role during cooling which may involve the intrinsic excitability mechanism of neurons compensating for the effect of reduced PSP on the average firing activity of the populations. Thus, a reciprocal Q10 factor was formulated as put forward in (Eq 10). Similarly, the Q10 factor involved here accounts for mean-field effect of temperature to several processes occurring during action potential generation such as diffusion of ions and ion channel gating across the neuronal population. Fig 4 shows how the average firing rate is compensated by the second temperature dependence. The firing frequency of a positive average membrane potential in the original firing response curve corresponds to an increased firing frequency at the same value of average membrane potential in the temperature dependent curve. The effect is opposite for negative average membrane potentials and rather minimal. A direct physiological interpretation of this mechanism can be examined if we write the equation in its equivalent form
S ( v ) = 2 e 0 1 + e ( Q 10 , i n t ( T - T 0 ) / 10 v t h - v Q 10 , i n t ( T - T 0 ) / 10 σ t h ) , (10)
where the Q10 factors are now with the parameters vth and σth. Recall that vth is the average threshold of firing of neurons and σth is the variability in the thresholds of excitation of neurons. This then implies that as a compensatory mechanism, cooling lowers both the average and variance of the distribution of the firing thresholds of neurons in the population. Hence, even if the average PSP is reduced resulting in lower average membrane potential, epileptic discharges can still be persistent since lower average threshold of firing allows for subthreshold activity before cooling to become suprathreshold during cooling. This can be seen as a form of homeostasis in the firing activity of the neuronal population involving both synaptic and intrinsic excitability mechanisms. Surprisingly, the combined mechanisms result in suppression of epileptic discharges in terms of magnitude which is not captured if we assume temperature dependence in the post-synaptic impulse response function alone. This is because higher Q10,syn values now significantly reduce PSP responses (Fig 11(b)) but the effect of which is compensated by the reciprocal of Q10,int. Slight increase in frequency of discharges observed in some of the experiments can be realized if Q10,int is made slightly greater than Q10,syn. Further increasing Q10,int proportionately increases the frequency of discharge activity (Fig 9).
Reduced threshold potential of firing during cooling has been reported on an early experiment with squid axons [27]. Experiments with mammalian brains [28, 67, 68] reported that cooling depolarizes cell membrane potential and increases input resistance. In [68], Volgushev et al. noted that cooling-induced depolarization of cell membrane occurs with an even higher gradient giving a marked decrease in the difference between the spiking threshold and the actual resting membrane potential. Thus, cooling brings the cells closer to spiking threshold, increasing excitability and decreasing variability in excitation levels across neuronal population. They proposed that such cooling-induced depolarization of the cell membrane may be attributed mainly by reduction of partial K+ conductance. Variability in threshold potential of firing has also been reported to increase with recent spiking activity [69]. We suppose that the opposite happens during cooling. As discussed earlier, cooling can decrease average firing rate of neurons which can imply less recent spiking activity. Henze and Buzsaki [69] suggested that prior action potentials cause Na+ channel inactivation that recovers with approximately a one-second time constant, increasing action potential threshold during this period. On the other hand, a study by Yu et al. [70] suggests that firing threshold variability can be explained by backpropagation of action potentials. Moreover, cooling was shown to strongly inhibit A-type K+ channels [71] in DRG neurons while these channels are reported to regulate action potential backpropagation in CA1 pyramidal neurons [72]. This might be in conflict with our finding that cooling reduces variability in firing thresholds since inhibited A-type K+ channels enhance backpropagating action potentials which in turn increases variability in firing thresholds. Then again, it is also possible that a net decrease in action potential backpropagation results as cooling can attenuate other critical factors such as density of Na+ at axon initial segment [73] and ion transport at nodes of Ranvier [74].
Estimation of the model from cooling experiments indicated that Q10,int is only slightly less than Q10,syn. This means that during cooling, the intrinsic excitability mechanisms of neurons just balance out the effect of temperature change on PSP generation. At first, it seemed that when Q10,int ≈ Q10,syn, discharge activity is suppressed but not terminated and when Q10,int ⪇ Q10,syn, discharge activity is terminated. To verify this generalization, we simulated the model for different values of Q10,int fixing Q10,syn = 1.8. This led us to discover bifurcation patterns in the model which were confirmed using numerical continuation on the noiseless version of the model. First, we have verified that when Q10,int ≈ Q10,syn, discharge activity is suppressed but not terminated. At this point, the intrinsic mechanism “fully” compensates the effect of the synaptic mechanism resulting to a discharge activity that has approximately the same frequency but reduced in magnitude (Fig 11(b)). However, we found out that when Q10,int ⪇ Q10,syn, discharge activity is terminated only up to a certain value of Q10,int and a seizure activity can arise with a wide range of intermediate Q10,int values. As far as the authors are knowledgeable, there has been no report that seizure activity was ever observed in focal cooling of epileptic discharges. Moreover, Q10,int values are not interpretable in terms of how intrinsic firing mechanisms can give rise to such values which would allow experiments to verify such finding. In theory, this should guide the design of implantable cooling devices which would necessitate a feedback control law to terminate cooling when a possible seizure can arise. Similar bifurcation patterns were observed for arbitrary values of Q10,syn other than 1.8. Our estimation results indicated Q10 values around 1.8 which was, surprisingly, also reported in previous studies involving voltage-gated Na+ channel (VGNC) dynamics [13]. Then again, in vitro studies [68, 75] suggest that involvement of VGNC might be ruled out as abortion of epileptiform discharges were seen to be associated with a depolarization block. Perfect depolarization is against changes in the gating property as initially hypothesized, i.e., cooling is not inducing a liquid phase transition in phospholipid bilayer of the membrane thereby distorting the channel’s property, rather through other mechanisms.
Another interesting study by Motamedi et al. [75] with an in vitro epilepsy model showed that cooling has differential effect on the firing rates of pyramidal cells and interneurons. This actually motivated the models where we included more temperature dependent parameters to investigate possible differential effect of cooling on PSP generation. This relies on the assumption that cooling may have differential effect on different neurotransmitters responsible for generating PSP. However, in this study, the model parameters were estimated from in vivo EEG recordings which have clear departures from the aforementioned in vitro study. We can speculate though that it may be possible to reproduce such differential effect of cooling on the activity of pyramidal cells and inhibitory interneurons if we had isolated EEG recordings from pyramidal cell population and interneuronal population activities and from which we could estimate the model parameters with an appropriate modification of the objective function (Eq 9). Nevertheless, when the effect of cooling on inhibitory interneurons is much less than on excitatory interneurons, reduced average membrane potential of pyramidal cell population results and consequently, reduction in the average firing frequency of the population is observed as reported in the study.
The results presented in this paper only considered the steady-state effect and does not include transient dynamics of cooling on epileptic discharges although some experiments have noted the effect of rate of cooling on termination of epileptic discharges. For instance, an in vitro study [44] reported that during slow cooling, epileptic discharges persist with decreasing amplitude until termination is achieved with further temperature drop. In contrast, rapid cooling achieves immediate termination of the discharges. The gradual decrease in amplitude of epileptic discharges during slow cooling can be captured by the model using an appropriate model for temperature dynamics (e.g. Newton’s Law of Cooling). In its present form, immediate termination of discharges by rapid cooling can be explained by our model as a case where Q10,int ⪇ Q10,syn, i.e. reduction in average and variance of firing thresholds across neuronal population is not able to compensate reduction in discharge frequency due to reduced average membrane potential resulting from attenuation of post-synaptic activity. Alternatively, such transient effect may be modeled by a Q10 that decays from a non-steady state value to a steady state value proportional to the rate of cooling. In most in vitro studies that we reviewed, steady-state termination of epileptic discharges was achieved using either slow or rapid cooling down to a constant temperature. In contrast, termination may not be always possible in in vivo setting. We surmise that the compensatory mechanism put forward by the model is more concomitant in in vivo than in in vitro environment.
Recent studies on epilepsy and epilepsy models have involved the role of non-neuronal cells such as astrocytes and microglia in mechanisms of seizure development such as reactive astrogliosis, glial-mediated inflammation, and Ca2+ signalling dysfunction [76, 77]. It may also be possible that cooling can attenuate activation of both neuronal and non-neuronal cells that will consequently impair their involvement in one or several hyperexcitability mechanisms. While there have been recent attempts at modelling the interaction of neuronal and non-neuronal cells [78, 79], formulation of temperature dependence on the models may involve multimodal recordings other than EEG (extracellular GABA and glutamate concentrations, cerebral blood flow) in focal brain cooling experiments to estimate the model parameters. This is an interesting direction which we hope to pursue in the future.
|
10.1371/journal.pcbi.1005173 | MD/DPD Multiscale Framework for Predicting Morphology and Stresses of Red Blood Cells in Health and Disease | Healthy red blood cells (RBCs) have remarkable deformability, squeezing through narrow capillaries as small as 3 microns in diameter without any damage. However, in many hematological disorders the spectrin network and lipid bilayer of diseased RBCs may be significantly altered, leading to impaired functionality including loss of deformability. We employ a two-component whole-cell multiscale model to quantify the biomechanical characteristics of the healthy and diseased RBCs, including Plasmodium falciparum-infected RBCs (Pf-RBCs) and defective RBCs in hereditary disorders, such as spherocytosis and elliptocytosis. In particular, we develop a two-step multiscale framework based on coarse-grained molecular dynamics (CGMD) and dissipative particle dynamics (DPD) to predict the static and dynamic responses of RBCs subject to tensile forcing, using experimental information only on the structural defects in the lipid bilayer, cytoskeleton, and their interaction. We first employ CGMD on a small RBC patch to compute the shear modulus, bending stiffness, and network parameters, which are subsequently used as input to a whole-cell DPD model to predict the RBC shape and corresponding stress field. For Pf-RBCs at trophozoite and schizont stages, the presence of cytoadherent knobs elevates the shear response in the lipid bilayer and stiffens the RBC membrane. For RBCs in spherocytosis and elliptocytosis, the bilayer-cytoskeleton interaction is weakened, resulting in substantial increase of the tensile stress in the lipid bilayer. Furthermore, we investigate the transient behavior of stretching deformation and shape relaxation of the normal and defective RBCs. Different from the normal RBCs possessing high elasticity, our simulations reveal that the defective RBCs respond irreversibly, i.e., they lose their ability to recover the normal biconcave shape in successive loading cycles of stretching and relaxation. Our findings provide fundamental insights into the microstructure and biomechanics of RBCs, and demonstrate that the two-step multiscale framework presented here can be used effectively for in silico studies of hematological disorders based on first principles and patient-specific experimental input at the protein level.
| Red blood cells (RBCs) and their mechanical properties play a crucial role in the dynamic and rheological behavior of blood in normal and disease states. However, the precise determination of RBC membrane properties is hard to be achieved experimentally. To this end, accurate numerical modeling can be used to provide valuable information for quantifying the biomechanical properties of RBCs. In this paper, we have developed and validated a two-step multiscale framework for RBC modeling, by performing molecular dynamics simulations to compute the shear modulus, bending stiffness and network parameters of a small RBC patch, which we then use as input to dissipative particle dynamics simulations to predict the stress field and morphology of defective RBCs, including Plasmodium falciparum-infected RBCs as well as RBCs in hereditary spherocytosis and elliptocytosis.
| Blood is a biological fluid that delivers nutrients and oxygen to living cells and removes their waste products. Due to its particulate nature, blood is considered as a complex non-Newtonian fluid exhibiting intriguing dynamic and rheological behavior depending mainly on flow rate, vessel geometry, and volume fraction of suspending particles, especially red blood cells (RBCs).
The deformability of a RBC is determined by the geometry, elasticity, and viscosity of its membrane [1, 2]. A healthy RBC (H-RBC) has a biconcave shape when not subject to any external stress and is approximately 8.0 μm in diameter and 2.0 μm in thickness. The membrane of a RBC consists of a lipid bilayer contributing to the bending resistance, an attached spectrin network (cytoskeleton) responsible for the shear stiffness, and transmembrane proteins such as band-3 and glycophorin C, bridging the connections between lipid and spectrin domains. Experimental and numerical observations of RBC behavior in flow mimicking the microcirculation reveal dramatic deformations and rich dynamics. The extreme deformability allows RBCs to squeeze through narrow capillaries in microcirculation without any damage. However, this feature of RBCs can be critically affected by parasitic conditions such as malaria [2, 3], or genetic factors, e.g., in sickle cell disease (SCD) [4, 5], hereditary spherocytosis (HS) [6], and hereditary elliptocytosis (HE) [7]. For many diseases involving RBCs, it is known that the membrane damage and interactions associated with the lipid bilayer and the cytoskeleton including elastic strength, viscous friction, and integral protein linkages strongly influence the biomechanics of RBCs [8]. For example, RBCs infected with Plasmodium falciparum (Pf-RBCs) become progressively less deformable and more spherical during the intraerythrocytic cycle. Remarkable nanoscale protrusions (knobs) have recently been identified causing significant stiffening effects on the cell membrane [9, 10]. Knobs mainly composed of two exported parasite proteins, knob-associated histidine-rich protein (KAHRP) and parasite-derived erythrocyte membrane protein (PfEMP1), deposit at the cytoplasmic face of RBC membrane and form vertical links to spectrin network. Consequently, both the lipid bilayer and the spectrin network are stiffened by the knobs as it is implied by the notable enhancement in shear resistance. In addition to the presence of knobs, actin mining introduced by the parasite invasion may result in an enhanced spectrin network at the trophozoite stage or in a deficient spectrin network at the schizont stage [10, 11].
SCD is an inherited blood disorder exhibiting heterogeneous cell morphology and abnormal rheology, due to the polymerization of sickle hemoglobin at high enough concentrations, forming long fibers that distort the RBC shape and dramatically alter their biomechanical properties [4, 12]. HS RBC is usually caused by the defects in anchoring proteins involved in vertical interactions between lipid bilayer and spectrin network, whereas HE RBC is a result of defects in spectrin filaments related to lateral interactions in the spectrin network [13]. Protein mutations associated with membrane defects subsequently lead to aberrant cell shape and impaired deformability. Thus, quantifying the deformability of RBCs could play a key role in understanding RBC related diseases.
Recent advances in computational modeling and simulation enables us to tackle a broad range of dynamics and rheology problems associated with RBCs [14–16]. Several computational models at the whole-cell level, including spectrin-level and multiscale RBC models [17–23], have been developed and employed to quantify the biomechanical properties and dynamic behavior of RBCs in malaria and other hematological diseases [24–29]. Examples include dynamic cell deformability for various stages of the Pf-RBCs [24–26, 30–33] and cell morphological sickling [34–36] and vaso-occlusion phenomena in SCD [27, 37]. In these existing models, the membrane is usually considered as a single shell with effective properties that represent the combined effects of lipid bilayer and spectrin network. Under normal conditions, the cytoskeleton is attached to the lipid bilayer from the cytoplasmic side. However, under certain conditions, such as RBCs in SCD and other hereditary disorders, the cytoskeleton may become dissociated from the lipid bilayer [12, 13]. Also, in malaria disease, the Pf-RBC undergoes irreversible structural modifications with the deposition of knob structures on the membrane surface and the actin remodeling in the skeletal network [38, 39]. The biomechanical properties associated with the bilayer-cytoskeleton interactions strongly influence biorheology, cell function and the onset and progression of RBC diseases. However, the one-component whole-cell models cannot facilitate detailed whole-cell exploration of diverse biophysical and biomechanical problems involving RBCs, such as the bilayer loss in HS, the bilayer-cytoskeleton uncoupling in SCD and the aforementioned multiple stiffening effects of knob structures in Pf-RBCs. For these reasons, there is a compelling need to develop a more realistic RBC representation, e.g., to endow the spectrin-based RBC models with more accurate structure, hence considering separately the lipid bilayer and cytoskeleton but also include the transmembrane proteins.
Recent efforts have been directed towards this approach. For example, a two-component composite model of RBC membrane with explicit descriptions of lipid bilayer, cytoskeleton, and transmembrane proteins has been developed and implemented using coarse-grained molecular dynamics (CGMD) [10, 41, 42]. This CGMD membrane model has been successfully applied to study the membrane-related problems in RBCs such as protein diffusion and vesiculation in defective RBC membrane [43, 44], and the stiffening effects of knobs on Pf-RBCs [10]. Although changes on the biomechanics of RBC membrane, including bending rigidity and shear modulus, in certain diseases can be evaluated by modeling a small piece of cell membrane with the two-component composite model, the whole-cell characteristics strongly related to RBC biomechanics and biorheology are not efficiently depicted by modeling only a portion of the RBC membrane. Recently, a two-component whole-cell model has been developed and implemented using dissipative particle dynamics (DPD) [45, 46]. The DPD RBC model also accounts separately for the lipid bilayer and cytoskeleton, but it includes implicitly the transmembrane proteins, Thus, the DPD model is computationally more efficient than CGMD model for RBC modeling at the whole-cell level, which has been applied to investigate RBC response and dynamics in flow. However, the lack of the molecular details in this two-component whole-cell model may limit its predictive capacity in identifying the key factors that cause the reorganization of the RBC membrane. An effective way to address this problem is to incorporate only the necessary molecular information from a molecular-detailed composite membrane model into a more coarse-grained whole-cell model.
In this work, we develop a two-step multiscale framework by employing the two-component models; see Fig 1. The only experimental input required is information about the structural defects of the lipid bilayer, the cytoskeleton, and their coupling via the transmembrane proteins. This information can be obtained by Scanning Electron Microscopy (SEM) [47] or Transmission Electron Microscopy (TEM) [40]. We then apply this framework to study the biomechanical characteristics of the healthy and diseased RBCs under static and dynamic tensile forcing. Specifically, we probe the alterations in cell deformability from the development of knob structures and remodeling of the spectrin network of Pf-RBCs. We also assess the implication of dynamic deformation and recovery response of defective RBCs from the bilayer-cytoskeleton dissociation in hereditary disorders such as HS and HE.
The rest of the paper is organized as follows. In Section 2, we briefly describe the RBC model and simulation method. In Section 3, we present and discuss our numerical results. Finally, in Section 4, we summarize the findings and present the conclusion. We also provide supplementary material in order to better explain the simulation results.
In the first step of the two-step process, we compute the shear modulus, bending stiffness, and network parameters by employing CGMD on a small RBC patch. In the second step, by passing the aforementioned parameters as input to a whole-cell DPD model, we simulate the stretching deformation, stress field, and shape relaxation of the normal and diseased RBCs subject to tensile forcing. For completeness, the simulation methods and models are briefly reviewed below, whereas details on the construction of RBC models are available elsewhere [10, 42, 45, 46].
In the two-component CGMD model of a normal RBC membrane, the lipid bilayer and cytoskeleton as well as the transmembrane proteins are explicitly represented by CGMD particles [41]. Specifically, three types of CG particles are introduced to represent the lipid bilayer of the RBC membrane (Fig 2A and 2C). The red particles represent clusters of lipid molecules with a diameter of 5 nm, which is approximately equal to the thickness of the lipid bilayer; the black-color particles signify band-3 complexes; the light blue particles immersed in the lipid bilayer are glycophorin C. The volume of one black particle is similar to the excluded volume of the membrane domain of a band-3 protein. However, when band-3 proteins interact with the cytoskeleton, the effect of the cytoplasmic domain has to be taken into account and thus the effective radius is ≈ 12.5 nm. One third of band-3 particles, which are connected to the spectrin network, are depicted as yellow particles (Fig 2C).
The spectrin tetramer is modeled by a chain of 39 beads (grey particles) connected by spring bonds [48]. The corresponding potential has the form,
V c y s - s = 1 2 k 0 ( r - r e q s - s ) 2 (1)
where k0 and r e q s - s are the spring constant and equilibrium distance between two spectrin particles, respectively. Spectrin particles that are not connected by the spring potential interact with each other via the repulsive term of the Lennard-Jones potential as follows
V r e p ( r i j ) = 4 ϵ σ i j r i j 12 - σ i j r i j 6 + ϵ r i j < r e q s - s 0 r i j ≥ r e q s - s (2)
where ϵ is the energy unit and σij is the length unit, and rij is the distance between spectrin particles.
To couple the lipid bilayer and spectrin network, actin junctional complexes (blue particles) are connected to the glycophorin C and the middle beads of the spectrin network are bonded to the band-3 complexes (black particles) which are specifically rendered in yellow particles, as shown in Fig 2C. These bonds are modeled as harmonic springs, given by
V c y a - s = 1 2 k 0 ( r - r e q a - s ) 2 (3)
where r e q a - s = 10 nm is the equilibrium distance between an actin and a spectrin particle. For a detailed description of the configuration of the cell membrane and the employed potentials, we refer to Ref. [42].
In the two-component whole-cell model, the cell membrane is modeled by two distinct components, i.e., the lipid bilayer and the spectrin network (Fig 2B). Specifically, through the DPD approach, each component is constructed by a 2D triangulated network on a membrane surface that is characterized by a set of points with Cartesian coordinates xi, i ∈ 1 ⋯ Nv which are vertices of the 2D triangulated network. Different from the one-component whole-cell model [20, 21], the lipid bilayer of the two-component whole-cell model has no shear stiffness at healthy state but only bending stiffness and a very large local area stiffness, whereas the cytoskeleton has no bending stiffness but possesses a finite shear stiffness.
The whole-cell DPD model takes into account the elastic energy, bending energy, bilayer-cytoskeleton interaction energy, and constraints of fixed surface area and enclosed volume, hence
V ( x i ) = V s + V b + V a + V v + V i n t (5)
where Vs is the elastic energy that mimics the elastic spectrin network, given by
V s = ∑ j ∈ 1 . . . N s k B T l m ( 3 x j 2 - 2 x j 3 ) 4 p ( 1 - x j ) + k p ( n - 1 ) l j n - 1 , (6)
where lj is the length of the spring j, lm is the maximum spring extension, xj = lj/lm, p is the persistence length, kBT is the energy unit, kp is the spring constant, and n is a specified exponent. The shear modulus of the RBC membrane, μ0, is determined by
μ 0 = 3 k B T 4 p l m x 0 x 0 2 ( 1 - x 0 ) 3 - 1 4 ( 1 - x 0 ) 2 + 1 4 + 3 k p ( n + 1 ) 4 l 0 n + 1 , (7)
where l0 is the equilibrium spring length and x0 = l0/lm. The bending resistance of the RBC membrane is modeled by
V b = ∑ j ∈ 1 . . . N s k b 1 - c o s ( θ j - θ 0 ) , (8)
where kb is the bending constant, θj is the instantaneous angle between two adjacent triangles having the common edge j, and θ0 is the spontaneous angle.
Constraints on the area and volume conservation of RBC are imposed to mimic the area-preserving lipid bilayer and the incompressible interior fluid. The corresponding energy is given by
V a + v = ∑ j ∈ 1 . . . N t k d ( A j - A 0 ) 2 2 A 0 + k a ( A cell - A cell , 0 tot ) 2 2 A cell , 0 tot + k v ( V cell - V cell , 0 tot ) 2 2 V cell , 0 tot , (9)
where Nt is the number of triangles in the membrane network, A0 is the triangle area, and kd, ka and kv are the local area, global area and volume constraint coefficients, respectively. The terms A cell , 0 tot and V cell , 0 tot represent the specified total area and volume, respectively.
The bilayer-cytoskeleton interaction potential, Vint, is expressed as a summation of harmonic potentials given by
V i n t = ∑ j , j ′ ∈ 1 . . . N b s k b s ( d j j ′ - d j j ′ , 0 ) 2 2 , (10)
where kbs and Nbs are the spring constant and the number of bond connections between the lipid bilayer and the cytoskeleton, respectively. djj′ is the distance between the vertex j of the cytoskeleton and the corresponding projection point j′ on the lipid bilayer, with the corresponding unit vector njj′; djj′,0 is the initial distance between the vertex j and the point j′, which is set to zero in the current simulations. Physical view of the local bilayer-cytoskeleton interactions include the major connections via band-3 complex and ankyrin, as well as the secondary connections via glycophorin C and actin junctions (Fig 2A), here we consider them together as an effective bilayer-cytoskeleton interaction and model it as a normal elastic force, f j j ′ E, and a tangential friction force, f j j ′ F (Fig 2B). The corresponding elastic force on the vertex j of the cytoskeleton is given by
f j j ′ E = k b s ( d j j ′ - d j j ′ , 0 ) n j j ′ d j j ′ < d c 0 d j j ′ ≥ d c (11)
where dc ≈ 0.2 μm is the cutoff distance. The tangential friction force between the two components is represented by
f j j ′ F = - f b s [ v j j ′ - ( v j j ′ · n j j ′ ) n j j ′ ] , (12)
where fbs is the tangential friction coefficient and vjj′ is the difference between the two velocities.
The RBC membrane interacts with fluid particles through DPD conservative forces, which are defined as
F i j C = a i j ( 1 - r i j / r c ) s n i j , (13)
where rij = ri − rj, rij = |rij|, nij = rij/rij, aij is the maximum repulsion between particles i and j, rc is the cutoff distance, s = 1 is the most widely adopted for the classical DPD method. However, other choices (e.g., s = 0.25) for the envelopes have also been used. Detailed description of these interations can be found in Ref. [21]. In addition, in combination with the total energy V(xi) expressed in Eq 5, we are able to derive the total force acting on particle i of the RBC membrane,
F i = - ∂ V ( x i ) ∂ x i + ∑ j ≠ i F i j C . (14)
The RBC model is multiscale, as the RBC can be represented on the spectrin level (NvDPD,S = 23,867), where each spring in the network corresponds to a single spectrin tetramer with the equilibrium distance between two neighboring actin connections of ∼ 75 nm [20, 59], equal to the average length of one edge of triangular mesh of RBC membrane in the CGMD model. In such case, both CGMD and DPD simulate the RBC membrane at the spectrin level [21, 44]. Thus, the size and density of nanoscale knobs of Pf-RBCs applied in both CGMD and DPD models can be directly derived from measured data in experiments. In addition, the RBC membrane properties, such as shear modulus μ0 and bending stiffness kc can be passed directly from CGMD to DPD,
l 0 DPD , S = l 0 CGMD , S , μ 0 DPD , S = μ 0 CGMD , S , k c DPD , S = k c CGMD , S . (15)
On the other hand, for more efficient computation, the RBC network can also be coarse-grained by using a smaller number of vertices. The equilibrium spring length of the coarse-grained RBC model is then estimated as:
l 0 DPD , C = l 0 DPD , S ( N v DPD , S - 2 N v DPD , C - 2 ) . (16)
Using a similar geometric argument, the spontaneous angle is adjusted as,
θ 0 DPD , C = θ 0 DPD , S ( l 0 DPD , S / l 0 DPD , C ) . (17)
In addition, the property parameters of the RBC membrane can be estimated as,
μ 0 DPD , C = μ 0 DPD , S , k c DPD , C = k c DPD , S . (18)
In this study, we adopt a whole-cell DPD model with NvDPD,C = 9128 in order to simulate the stretching behavior of the whole RBC more effectively. We model an H-RBC using the whole-cell DPD model with the following properties: A cell , 0 tot = 134 μm2, V cell , 0 tot = 94 μm3. Based on the CGMD and previous one-component whole-cell simulations of [21], we choose μ0 = 4.73 μN/m for H-RBC. Regarding the elastic contribution to the interaction energy, we use the value of the bending modulus derived directly from CGMD simulations, i.e., kc = 31.9 kBT for H-RBC, which is approximately 1.3 ×10−19 J. The corresponding bending constant is set to k b = 2 3 k c = 36.8 kBT. It has been shown that the bending resistance contributes little to the cell deformation in the stretching test in previous studies [23, 26, 33], so we keep kb constant in all simulation cases.
The simulations are performed using a modified version of the atomistic code LAMMPS. The time integration of the motion equations is computed through a modified velocity-Verlet algorithm [60] with λ = 0.50 and time step Δt = 1.0 × 10−5 τ ≈ 0.23 μs. It takes 2.0 ×106 time steps for a typical simulation performed in this work.
Optical tweezers have been used successfully in the studies of RBC elasticity because of the finer-scale in describing whole-cell deformation [2, 64]. Numerical simulations have mimicked this experimental setup by directly applying stretching forces on the opposite sides of a RBC [23, 33, 59]. In a normal RBC, proper transmembrane protein and protein-to-lipid linkages in the membrane could sustain the cell elasticity and mediate the association between the lipid bilayer and the spectrin network even though a relatively large tensile force is imposed on the cell membrane. However, additional vertical linkages between the lipid bilayer and cytoskeleton domains introduced by parasite proteins in Pf-RBC will reduce cell viscoelasticity and enhance cell stiffness. On the other hand, defects in the membrane proteins of RBCs in HS and HE will weaken the bilayer-cytoskeleton interactions and facilitate the detachment of the lipid bilayer from the spectrin network. In this study, we have focused on two types of pathological RBCs, Pf-RBCs and hereditary diseases with protein defects, and quantitatively investigate their morphological and biomechanical properties during the stretching test by the two-component whole-cell model.
Here we investigate the elastic properties of the modeled RBC in health but also at different stages of malaria. To probe the RBC mechanical response and the change of its mechanical properties at different malaria stages, we subject the cell to stretching deformation analogously to that in optical tweezer experiments. As shown in Fig 2D, the T-RBC has a lower knob density (ρknob,DPD = 7 knobs/μm2) and enhanced spectrin network, whereas the S-RBC bears a higher knob density (ρknob,DPD = 12 knobs/μm2) and deficient spectrin network. The total stretching force, Fs, is applied in opposite direction to ϵNv (ϵ = 0.05) vertices of the cell membrane at diametrically opposite directions. First, we examine the response of H-RBCs and Pf-RBCs to large deformation in comparison with previous experimental measurements [2] and computational simulations based on the one-component whole-cell model [24]. We analyze the changes in DA and DT. Our simulation results show that both DA and DT are in agreement with previous experimental and computational results, see Fig 4A. Since the model has separate components for the cytoskeletion and lipid bilayer, we can get the DA and DT values for each component. The consistency of the DA (or DT) values calculated from the lipid bilayer and the cytoskeleton indicate the strong association between the two components.
To investigate the influence of the knob density on RBC deformability, we change the knob density from 4 to 9 knobs/μm2 for T-RBC and 10 to 18 knobs/μm2 for S-RBC. For an H-RBC under tensile force Fs = 110 pN, we obtain that the EI value is about 0.51. For a less deformable Pf-RBC, we find a considerable decrease in EI value, i.e., EI ≈ 0.20 for T-RBC at ρknob,DPD ≈ 4 knobs/μm2 and EI ≈ 0.15 for S-RBC at ρknob,DPD ≈ 10 knobs/μm2. The decrease in EI values from H-RBC to T-RBC then to S-RBC is associated with a reduction of RBC deformability as the progression of the parasite maturation in Pf-RBC. With the increase of knob density, a further decrease in EI values for both T-RBC and S-RBC is obtained, see Fig 4B, which indicates a further reduction in cell deformability of the Pf-RBC. Our simulation results demonstrate that the knobs, being rigid, contribute to cell membrane stiffness.
In other words, we model the Pf-RBC with normal spectrin network by removing the influence of actin remodeling. We find that the EI values increase slightly for T-RBC but decrease slightly for S-RBC (Fig 4B). Nevertheless, in comparison with the change of EI values resulting from the stiffening effects of knobs, the difference of EI values between the modified (enhanced or deficient) and normal spectrin network is relatively small. These simulation results further demonstrate that the presence of the knobs in Pf-RBCs is the primary stiffening factor in the loss of cell deformability, which is also consistent with the recent CGMD simulation study by Zhang et al. [10]. More importantly, we efficiently scale up the particle-based RBC model from a portion of cell membrane to a whole-cell structure leading to more realistic simulations of RBC dynamics and blood rheology.
Using the virial theorem, we can obtain the average virial stress tensor over a volume Ω,
Π = 1 Ω ∑ i ∈ Ω - m i ( v i - v ¯ ) ( v i - v ¯ ) + 1 2 ∑ j ∈ Ω r i j · F i , (20)
where mi and vi are the mass and velocity of particle i, respectively, and v ¯ is the average velocity of particles in the volume Ω. The stress contours of H-RBC, T-RBC, and S-RBC at Fs = 110 pN are shown in Fig 4C. In general, large stress response is observed around the longitudinal axis of the stretched cell, by contrast, small stress response emerges at the two sides of the transverse axis [65]. For H-RBC, the lipid bilayer has little shear resistance but large bending stiffness and helps to maintain the membrane surface area. The cytoskeleton is primarily responsible for the shear elastic properties of the RBC. Consequently, the principal stress of the stretched RBC is essentially reflected on the cytoskeleton, while only a small stress response is observed at the ends of lipid bilayer where the tensile force is imposed. As shown in Fig 4C, more than 90% of RBC’s total tensile stress comes from the cytoskeleton component, while the rest (less than 10%) from the lipid bilayer component. For T-RBC and S-RBC, the presence of the knobs immersed in the lipid domain stiffens the cell membrane, which is reflected in the stress contour of the lipid bilayer of the stretched RBC, see Fig 4C. The lipid bilayer with stiff knobs contributes ∼ 33% of the total tensile stress of T-RBC membrane. Compared with the results from T-RBC, the knobs with a higher density in S-RBC withstand an elevated stress of the stretched RBC. The stiffened lipid bilayer bears ∼ 45% of the total tensile stress of the S-RBC membrane.
The associations between the lipid bilayer and the cytoskeleton mediated by specific molecular interactions are essential for the mechanical stability of RBCs [66]. The primary interaction is the connection between the transmembrane protein band-3 and the spectrin via the ankyrin binding sites. The secondary interaction is supported by another transmembrane protein glycophorin C and the actin junctional complexes. To involve these bilayer-cytoskeleton interactions into the two-component whole-cell model, we have simply considered them together as normal (elastic) and tangential (friction) interactions. Two parameters, the elastic interaction coefficient, kbs = 46 pN/μm, and the tangential friction coefficient, fbs = 0.194 pN⋅μm−1s−1, are accordingly introduced for the normal RBC membrane [45].
We examine cell biomechanics and deformation of H-RBCs subject to stretching tests with the default values of kbs = 46 pN/μm and fbs = 0.194 pN⋅μm−1s−1 of the bilayer-cytoskeleton interactions. First, we probe the stretching deformation of both the lipid bilayer and the cytoskeleton under different stretching force (Fs), see S3 Fig. From this figure, we find that the values of DA and DT obtained from the two-component whole-cell model are in agreement with those from experimental measurements [2], and the one-component whole-cell model [21]. In addition, we find that there is a small difference in the detachment length, which is defined as the distance along the longitudinal axis from the rightmost part of the lipid bilayer to that of the cytoskeleton. However, the deviations are sufficiently small or purely due to statistical fluctuations under normal stretching forces (Fs ≤ 200 pN); hence, there is no significant bilayer-cytoskeleton detachment in these cases. A visible detachment of the lipid bilayer from the cytoskeleton appears in extreme cases (for example, Fs > 300 pN) due to abnormal stretching forces exerted on the cell membrane. Nevertheless, under normal conditions, the large deformability and the strength of the bilayer-cytoskeleton interactions of the H-RBC can be computationally described by the two-component whole-cell model.
To investigate the effect of weakened bilayer-cytoskeleton interactions in defective cell membrane on the deformability of RBCs with stretching test, we present an extensive simulation study on the biomechanical behavior of defective RBCs by varying the values of bilayer-cytoskeleton elastic interaction coefficient, kbs, and tangential friction coefficient, fbs, see Figs 5 and 6. First, we consider the cell membrane with the default value of kbs but different fbs (Fig 5A). Specifically, we chose the value Fs = 140 pN, for which the RBCs undergo large biomechanical deformation without bilayer-cytoskeleton detachment under normal physiological condition. We find that when using the default value of fbs = 0.194 pN⋅μm−1s−1 or other values of similar magnitude, there is a strong coupling between the lipid bilayer and the cytoskeleton. This is indicated by the overlapped black solid lines (with black squares) and red dashed lines (with red spheres) in Fig 5A. However, assuming a pathological RBC state where fbs is decreased by one or two orders of magnitude, an apparent uncoupling between bilayer and cytoskeleton occurs. Compared to the DA value of cytoskeleton, the DA value of the lipid bilayer is relatively larger when fbs is smaller than 0.01 pN⋅μm−1s−1, and the detachment length between the lipid bilayer and the cytoskeleton increases as fbs reduces. Moreover, the DT values obtained from the lipid bilayer and cytoskeleton remain almost the same regardless of the changes in fbs, which can also be observed in Fig 5A. It seems that the DT value is insensitive to the variation of fbs. Next, we study the effect of the bilayer-cytoskeleton elastic interaction coefficient as shown in Fig 5B. Similarly, there is no obvious distinction in the deformation of RBC for different kbs at the default value of fbs (0.194 pN⋅μm−1s−1). However, the bilayer-cytoskeleton uncoupling along the stretching axis occurs when we decrease the values of both fbs and kbs by one or two orders of magnitude. Consistent with the previous simulations on RBC traversing across microfluidic channels and tank-treading dynamics [45, 46], the strength of the bilayer-cytoskeleton interaction responsible for the association between lipid bilayer and cytoskeleton is again verified by the RBC stretching tests.
Next we compare our simulation results with experimental data for normal and defective RBCs. The results on the functional dependence of the stretching response of normal and defective RBCs on the stretching force, Fs, are shown in Fig 6A. In general, DA grows while DT decreases with increasing Fs. For H-RBC, no bilayer-cytoskeleton detachment is observed due to the strong bilayer-cytoskeleton interactions. The corresponding stress contours of deformed RBCs at Fs = 0, 80 and 160 pN are presented in Fig 6B. The tensile stress of H-RBC inherently coming from the cytoskeleton is enhanced as Fs rises, which is in reasonable agreement with other computational results [65]. On the contrary, the lipid bilayer is observed to separate from the underlying spectrin network when a defective RBC undergoes large deformation due to the weakened bilayer-cytoskeleton interactions (Fig 6A). The stress contours of the defective RBC at Fs = 160 pN in Fig 6C show significant shear response at two ends of the stretching sites in the lipid domains coinciding with the location for bilayer-cytoskeleton detachment.
Upon external tensile forces, a normal RBC undergoes large mechanical deformation, and it restores to its original state when the external tensile force is no longer applied. However, the weakened bilayer-cytoskeleton interactions due to the vertical or lateral defects disturb the biomechanical stability of the RBC membrane. Therefore, when a defective RBC experiences large biomechanical deformation, it may be unable to recover its original shape even though the stretching force causing the shape change is removed. Here, we examine the elastic relaxation of H-RBC and defective RBC when Fs = 110 pN is turned off. The simulation results are shown in Fig 7 and more details are available in the video clips in the Supporting Information. In Fig 7A, we find that both the lipid bilayer and the cytoskeleton are deformed simultaneously under stretching-relaxation cycles and their DA and DT values approach the initial values at equilibrium state, which demonstrates that a normal RBC with high elasticity can recover its original shape. The recovery process of the RBC can be characterized by the dynamic recovery expression, R(t), which is described by an exponential decay [67],
R ( t ) = ( λ - λ ∞ ) ( λ 0 + λ ∞ ) ( λ + λ ∞ ) ( λ 0 - λ ∞ ) = exp - ( t - t 0 ) t c , (21)
where λ = D A D T, λ0 and λ∞ correspond to the ratios at release and recovery states; t0 is the time when Fs is turned off and tc is the characteristic time. Using the relaxation results from Fig 7A into R(t), we obtain the best fitting decay as illustrated in S4 Fig and estimate the characteristic time, tc, for both lipid bilayer and cytoskeleton, i.e., tc ≈ 0.12 s. This value falls within the range of 0.1–0.3 s measured from the shape relaxation of H-RBCs in experiments [65, 67] and also computed in simulations [33].
However, for the defective RBC with weakened bilayer-cytoskeleton interactions, once the detachment of the lipid bilayer from the cytoskeleton occurs, the DA values of both lipid bilayer and cytoskeleton are always larger than those at equilibrium state, even for a sufficient relaxation time, which indicates that the defective RBC loses its ability to recover its original shape. We note that not only the lipid bilayer but also the cytoskeleton has a permanent deformation because of the slow response distortion as shown in Fig 7B and the S2 video.
In summary, our simulation results show that the elastic deformations of Pf-RBCs match those obtained in optical tweezer experiments for different stages of intraerythrocytic parasite development. We found that the developed knobs on the cell membrane can effectively stiffen the lipid bilayer and the spectrin network, leading to a decrease in RBC deformability. We also investigated the effects of bilayer-cytoskeleton interactions and our simulation results reveal that these interactions play a key role in the determination of cell membrane biomechanical properties. We analyzed the biomechanical response of the RBCs during loading and upon release of the tensile force, in order to mimic the stretching and compression that RBCs experience as they pass through small capillaries in the microcirculation. Our results indicate that H-RBCs with normal bilayer-cytoskeleton interactions recover their original discocyte shape when the stretching force is turned off. However, defective RBCs with weakened bilayer-cytoskeleton interactions lose their ability to recover their original shape due to the irreversible membrane damage. Overall, our findings demonstrate that the two-step multiscale framework we have developed, combining coarse-grained molecular dynamics (CGMD) and dissipative particle dynamics (DPD), can be used to predict the altered biomechanical properties of RBCs associated with their pathophysiological states.
|
10.1371/journal.pcbi.1004601 | Lévy Walks Suboptimal under Predation Risk | A key challenge in movement ecology is to understand how animals move in nature. Previous studies have predicted that animals should perform a special class of random walks, called Lévy walk, to obtain more targets. However, some empirical studies did not support this hypothesis, and the relationship between search strategy and ecological factors is still unclear. We focused on ecological factors, such as predation risk, and analyzed whether Lévy walk may not be favored. It was remarkable that the ecological factors often altered an optimal search strategy from Lévy walk to Brownian walk, depending on the speed of the predator’s movement, density of predators, etc. This occurred because higher target encounter rates simultaneously led searchers to higher predation risks. Our findings indicate that animals may not perform Lévy walks often, and we suggest that it is crucial to consider the ecological context for evaluating the search strategy performed by animals in the field.
| Moving agents should efficiently search for targets (e.g., food, prey, or specific locations) when lacking information about the location of the targets. For this random search problem, the Lévy walk hypothesis claims that Lévy walk movement patterns (i.e., each step length follows a distribution that is heavy-tailed) enable the searcher to capture more targets. However, most searchers may have antagonistic agents (e.g., predators) that can lead to death. Thus, the searcher needs to seek targets while avoiding encounters with antagonistic agents. Here, we show that the Lévy search strategy is less efficient in terms of total lifetime fitness when the predators are abundant, and especially when predators adopt a sit-and-wait strategy. Moreover, the results indicate that the life-cycle type of the searcher is an important fitness factor. These ecological aspects significantly influence the consequences of the random search. Therefore, it is critical to consider the ecological properties of searchers and other interacting agents when examining and estimating animal movements.
| How should we move to search for targets when we have no information about their location? This is called the random search problem, which has attracted the attention of researchers in various fields [1]. The problem can be applied to various phenomena, including molecular-level movements within an organism, cell movements, movements of an individual animal, and the movement of robots [2–4]. For example, animals search their environment for food, prey, mates, and nesting locations, and DNA-binding proteins move around to find a specific DNA sequence to initiate gene expression. The search strategy is considered to evolve to be more efficient through the process of natural selection because successful searches increase fitness, especially at the individual level in animals.
The Lévy walk search (or foraging) hypothesis was proposed to solve the random search problem [5]. A Lévy walk is a special class of random walk models in which the probability function of step length l has a power-law tail: P(l)∼l−μ(1<μ≤3), where μ is a power-law exponent, such that rare ballistic movements occur among a number of relatively short steps. Comparisons of the efficiency of random searches showed that a Lévy walk with μ ≈ 2 was a highly efficient search strategy in environments where patchy prey were sparsely distributed [1,5–7]. In dense environments, on the other hand, Lévy walks had almost the same efficiency as Brownian walks [6]. Therefore, the Lévy walk foraging hypothesis predicts that most animals should perform Lévy walks while searching unless there are abundant targets.
Although many empirical studies have reported that diverse organismal components and taxa (e.g., T cells, insects, and human beings) perform Lévy walks with μ ≈ 2 [1,3,8–15], several recent analyses demonstrated that some animals had various Lévy exponents μ, or they exhibited Brownian walks [13,14,16–18]. For example, rigorous statistical analyses of deer and bumblebees failed to provide strong evidence for Lévy walks [16]. Thus, the question changed from whether animals have Lévy walk movement patterns to when or why animals perform Lévy walks. In general, the diversity of organisms is the result of varying ecological and environmental factors as well as complex biotic interactions with conspecific and heterospecific individuals [19]. Theoretical reports of random searches have generally focused only on search efficiency to evaluate the fitness of the searcher [1,5–7,20–23]. Moreover, most of these studies paid little attention to other relevant ecological factors such as death rate with predation risk, interactions with other individuals, and the metabolic costs of foraging. A few studies have considered such factors [24–28], including predation risk [29–32]. In physics, Yuste et al. analyzed the survival rate of mortal random (Brownian) walkers surrounded by diffusing traps and revealed that the death in the course of motion dramatically affected the search efficiency [33]. Such a situation would be relevant to biological encounters. However, the relationships between search efficiency and predation risks in the context of Lévy walks are still poorly understood.
Here, we focus on the fact that search efficiency represents the probability of an encounter with anything existing in the environment. Highly efficient search strategies may correspond to more frequent encounters with predators, and thus higher death rates. Therefore, search efficiency, as defined in previous studies, may not reflect the actual fitness, because fitness is not only determined by the efficiency of searching for targets (i.e., benefits), but also by the death rate caused by predation (i.e., cost) [34]. Reynolds [30,31] reported that predation risk altered the optimal strategy, but did not consider predators [30] or the fitness of the searcher [31]. In this paper, we explicitly introduce predation risk and life-cycle types to the previous simulations, and extend the random search scenario to correctly estimate the fitness of a searcher to determine an animal’s optimal search strategy.
First, we considered a searcher performing either the Lévy walk (hereafter, LW) or the Brownian walk (hereafter, BW) at movement velocity vs(= 1) in an environment in which patchy targets were sparsely distributed. Then, Np predators were randomly placed in the environment as the initial condition. To explore the effect of the predators’ movements, we considered four cases with respect to the predators’ movement velocity vp/vs = 0 (sit-and-wait); vp/vs = 0.2 (slow); vp / vs = 1 (middle); and vp / vs = 5 (fast). If vp > 0, we assumed that a predator performed LW with μ = 2 (or BW in S1 Text). For simplicity, we assumed that if the searcher encountered a predator, the searcher died from predation.
Second, if the death effect arising from encounters with predators was considered, search time became an important factor because the length of rest during searches was associated with fitness. Thus, each searcher had a maximum searching time, Tmax, that could be cut off by an encounter with a predator.
Finally, when considering the searcher’s fitness, we assume: (1) the fitness is the lifetime reproductive success (i.e., we analyze the number of offspring reproduced within the lifetime), (2) without alternation of generations (i.e., we do not take population dynamics into account), (3) a searcher has either one of two life-cycle types as described next. In the simplest case, finding a target directly led to increased fitness in a linear fashion (life-cycle type I). For example, when a female parasitoid wasp finds and attacks a host, and then searches for another target, we presume that its fitness increases linearly. Furthermore, when a male finds a female and mates, its fitness as a searcher also increases linearly. In contrast, animals characterized by life-cycle type II would need to survive until their reproductive stage Tmax to obtain higher fitness. In life-cycle type II, individuals that die from predation prior to maturity have no offspring and have a fitness of zero.
Here, we show the general relationship between fitness and the rate of encounter with targets and predators as well as the robustness of Tmax to our results. We denote the encounter rate with a predator per unit time ΔT as γ. The probability of an encounter with the predator at the m-th time unit is expressed as
(1−γ)m−1γ.
(1)
Therefore, when Tmax is divided into n pieces by unit time ΔT (i.e., Tmax = ΔTn), the mean search time
T¯
is
T¯=∑m=1n{m(1−γ)m−1γ}+n(1−γ)n=1−(1−γ)nγ,
(2)
where n(1−γ)n indicates the case in which the searcher never encounters predators. When the mean number of encounters with predators for n is k, nγ = k and (1−γ)n ≈ e−k for γ << 1 and a large n, thus
T¯≈n(1−e−k)k.
(3)
We calculated the fitness of LW and BW strategies in the ecological context using computer simulations because it is difficult to analytically derive the encounter rate in our relatively complicated setting, even though the analytical solutions were obtained in different scenarios under much simpler assumptions (e.g., Brownian walks, 1-D field, ideal gas model) [5,33,35,36,37]. Using the methods described previously [5,7,17], we simulated one searcher roaming in a 2-D environment in which some targets (e.g., food, hosts, mates) and predators were distributed. Although the species at higher trophic levels are lower in number in real ecosystems and the population we simulated seems unsustainable, we introduced only one searcher. This is because we focused on the fitness of a single searcher by picking it up from searcher’s population, and our main results must be robust if we introduce a number of searchers. The searcher had no prior information about the locations of both targets and predators, and wandered at a constant velocity vs = 1 (per unit time) in a 2-D continuous field with length L2 = 500×500 in which the boundary condition is periodic [7].
The LW was characterized by a distribution function P(l) ∼ l−μ(1 < μ ≤ 3). In our simulations, we derived step lengths from the following equation to obtain LW, generating a uniform random number u (0 < u ≤ 1; except for u = 0):
l=l0u(1−μ)−1,
(7)
where the minimum step length l0 is 1 [7]. For the BW simulation, to obtain an exponentially decaying distribution of the move length, each successive step length was drawn from a Gaussian distribution, where the mean was the minimum step length l0 = 1 and the variance was equal to 1 [7]. In LW or BW, after walking in a straight-line motion until reaching a step length l, the searcher turns in the angle drawn from a uniform distribution [−π,π].
The center of each patch was randomly scattered, and the radius of each patch was equal to 10. The number of targets and patches in the whole field was 1000 and 50, respectively. Each target was randomly assigned to a patch so that each patch had 20 targets on average. The targets were randomly distributed within the patch. In the initial state, Np predators were randomly distributed in the whole area, i.e., the x and y position of each predator was independently drawn from a uniform distribution [0, L] (See S1 Text for the effects of initial conditions). Rt and Rs represented the radius of the targets and searcher, respectively, and
Rs′
and
Rp′
represented the radius of perception of the searcher and predators. If the distance between the searcher and a target was less than
Rt+Rs′=1
, the searcher obtained the target, and the target disappeared. Then the step length of the searcher is truncated and recalculated, and the direction is drawn from a uniform distribution. After the searcher migrated a 500 path length, the depleted target regenerated to maintain the specified target density [17]. Similarly, the searcher died if the distance between the searcher and a predator was less than
Rs+Rp′=1
. The mean free path λ, which represents the mean distance or travel time between patches or targets, is
L22RN
for the 2-D environment [7]. Hence, in our simulation, λpatch = 2500 and λtarget = 125 for encounter distance R = 1. This is equivalent to the low-resource scenario of previous studies (e.g., [22]). The maximum search time Tmax was 104. To converge the results, the total time for a single parameter set (i.e., searcher’s movement pattern and density of predators) was 107 for sit-and-wait, slow, or middle predator conditions, and 5×107 for fast conditions. Then, k,η,γ were calculated, and the relative fitness was obtained using Eqs (5) and (6).
The results of the relative fitness (ϕLW / ϕBW) calculation for life-cycle type I are presented in Fig 1. When predators were absent (Np = 0), the relative fitness ϕLW / ϕBW was >2. Thus, LW with intermediate-level μ had the highest fitness, which was consistent with the findings of previous studies [5,7]. However, as the number of predators increased, the ϕLW / ϕBW ratio gradually declined to ~1 or slightly less than 1 when the predator strategy was sit-and-wait or slow LW (Fig 1A and 1B). In this case, a LW (μ ≈ 2) often performs best out of all LW’s. When the predator strategy was middle or fast LW, ϕLW / ϕBW was maintained at a high value, and LW could be an efficient strategy (Fig 1C and 1D).
Likewise, in the case of a searcher with life-cycle type II, the relative fitness ϕLW/ϕBW decreased substantially as the number of predators increased when the predator strategy was sit-and-wait or slow LW (Fig 2A and 2B). Even when the strategy of predators was middle LW, ϕLW/ϕBW decreased as the number of predators increased. These results were robust to other search strategies (i.e., correlated random walk or composite Brownian walk) (S1–S4 Figs) and to Brownian walk predators (S5 Fig). The relative fitness decreased because the searcher was likely to encounter a predator. The search time was shortened by death in a manner dependent on the search efficiency, and the relative mean searching time
T¯LW/T¯BW
depended on the search strategy (Fig 3). These results indicated that the LW strategy could lead to a high predator-encounter rate; therefore, BW could potentially be a risk-averting strategy.
To investigate these results, the relationship between the relative fitness and the encounter rate with targets and predators was examined (Fig 4). This result is not limited to our simulation results or to the relative fitness of LW or BW, but it describes a general trend. The relative encounter rates with targets and predators and the expected encounter number of BW for our simulation are presented in Fig 5. When the encounter rate with predators was low (i.e., low kBW), the fitness of random search strategies clearly depended on the encounter rate with targets (Fig 4A and 4D). Hence, LW had higher fitness in our simulation (Fig 5A). On the other hand, for intermediate or high kBW, fitness also changed depending on the encounter rate with predators (Fig 4B, 4C, 4E and 4F). Furthermore, fast predators displayed the same high predator encounter rates of high kBW and
γLWγBW≈1
(Fig 5B and 5C). Thus, LW had higher fitness under the fast-predator conditions for life-cycle type I and almost equal fitness for life-cycle type II. Similarly, the fitness of other random search strategies was determined by the encounter rate with targets and predators. The degree of encounter rate improvement not only depends on the search strategy, but also on the distribution or density of the targets [5,7], suggesting that the conditions for the optimal search strategy are complex.
Our results revealed that the random search strategy affected the death rate arising from predation, and that trade-offs could occur between foraging efficiency and predation risk. In nature, animal species have different ecological traits or interactions associated with their foraging behavior [34,38,39]. Considering such ecological factors, optimal foraging theory, as it currently exists, successfully predicts various types of animal behaviors from the viewpoint of maximizing fitness through natural selection [38,39]. However, previous studies of random search movements have only focused on foraging (i.e., search efficiency for targets), which may be unrealistic when considering the diversity of ecological characteristics and biotic interactions in nature. Lima et al. [34] reported that animals performed more efficient strategies in response to ecological factors, including risks, with such trade-offs. Our simulations predicted that where predators were abundant, a searcher performing a LW might have lower fitness depending on its ecological characteristics and those of the predators. This suggests that the optimal search strategy may change. Therefore, the parameter range in which the LW is advantageous may be narrower than previously estimated (Fig 4). The mechanism explaining these dynamics was that LWs not only increased the encounter rate with targets, but also with predators, which shortened the lifespan in exchange for the capture of more targets. The rare ballistic movements of LWs led to the high encounter rate with predators (Fig 5C), and this effect has been reported as a high encounter rate of a straight line motion with randomly distributed destructive targets [7] or new targets [35]. In the presence of predators, a searcher was confronted with conditions similar to the destructive search problem, because encounters with predators resulted in the death of the searcher. Although we assume the ecological context in this paper, such searching-avoiding trade-offs in the random search problem that we revealed here may occur in other contexts such as protein-DNA interactions [2,40].
Previous studies analyzing the predation effect on search strategies focused on the predation risk within a patch [30, 31], and reported that the predation risk could alter the optimal time spent for intensive searches if the predation risk increased as the time spent within a patch increased. In contrast, we concentrated on the predation risk in a whole area and predicted the fitness ratio between LW and BW by calculating the encounter rate with targets and predators. Also, Reynolds simulated the moving preys searched by one predator [31], and the study discussed that the prey movement patterns were determined by their foraging and not by cost of predation when predators are fast. This idea is consistent with our results for life-cycle type I (especially in Fig 1D), but we defined the fitness based on life-cycle and simulated the tri-trophic system consisting of targets, searchers, and predators. Consequently, we revealed the general effect of predation risk on search strategy (Fig 4).
To disentangle the effects of density, radius, and velocity of a searcher or predators on the relative fitness, we refer to analytical results of simple situations. Hutchinson et al. [35] and Dusenbery [37] reviewed the analytical results for the encounter rate of two kinds of straight motion agents (e.g., target and searcher, or searcher and predator) in 2-D and 3-D. In this case, the encounter rate is proportional to both the density of agents and encounter distance (i.e.,
Rs+Rp′
in our model). In our results, the density of predators is an important factor that can determine the relative fitness. In Fig 5C, the left (low density) and right (high density) figures are almost identical because the effect of predator’s density in the ratio of encounter rate
γLWγBW
is cancelled out. Although the movements in our simulation are not straight motions but LW or BW, the proportionality of density effects on encounter rate could be common. Therefore, the density of predators can affect only the number of encounters to predators k. The ratio of encounter rate depends on the characteristic of movements (i.e., LW or BW, and velocity) rather than the density of predators. Additionally, the radius of searcher and predators, that is, encounter distance can be also the same effect as the density of predators in our results because the encounter rate can be proportional to the encounter distance.
The encounter rate in 2-D of a stationary searcher and straight motion predators with constant speed vp is 2ρRsvp where ρ is the density of predators [37], and that of a straight motion searcher and predators with the speed vs = vp is 8ρRsvp / π [35]. Hence, the ratio of the encounter rates is 4 / π. This is consistent to our result for the ratio of encounter rate of BW (i.e., like a stationary searcher) and LW with small μ (i.e., like a straight motion searcher) under the presence of LW predators (i.e., like straight motion predators) in the case of vp/vs = 1 (green line in Fig 5C). In the case of vp/vs = 0, 0.2, 5, the movement of the faster individuals has a large effect on the encounter rate [37]. Therefore, compared with BW, LW in vp/vs = 0, 0.2 has the high encounter rate with predators (Fig 5C). Additionally, the analytical result for 3-D conditions is similar to that for 2-D [37]. Thus, our conclusion could be applied to 3-D such as prey-predator interactions of planktons in lakes or ocean.
Moreover, a recent study proposed a framework for encounter rates that are derived from an arbitrary trajectory of a searcher and immobile targets using an encounter kernel [41]. The combination of this technique and our results for general relationship between encounter rate and fitness (Fig 4) may provide the general framework integrating movements and fitness. This could give us the information about fitness directly from the trajectory and distribution of targets.
Many empirical studies have reported that the movement patterns of animals, from insects to human beings, are expressed as LWs with μ ≈ 2 [1]. However, the power-law exponents fitted to movement patterns sometimes ranged from 2 to 3 [1], suggesting that movement patterns may be diverse. Additionally, the data best fitted to the exponential decay distribution (i.e., BWs) has also been reported [13,14,16,18,42,43]. In theoretical studies, the first attempt reported that LWs with μ ≈ 2 were optimal for targets that can be revisited (i.e., non-destructive) or those that are extremely patchy [5]. Moreover, LWs with μ→1 (i.e., straight movement) were the optimum for randomly distributed destructive targets. After the study, the results of several versions of simulations suggested that LWs with 1 < μ ≤ 2 are more efficient depending on the prey distribution and other factors [20,21,26]. For the power-law exponent μ > 3 (i.e., BW), it has been theoretically reported that the foraging efficiency is similar to LW under high-resource conditions [6]. Our results suggests that under high predation risk, animals with power-law exponents close to three have higher fitness than μ ≈ 2 or μ < 2 (Figs 1 and 2), and those under intermediate predation risk, LW with 2 < μ < 3 also benefit. Therefore, it can be an alternative explanation for the diversity of power-law exponents.
There is a question of whether movements in animals are spontaneous patterns for adaptation or a reflection of interactions with targets or complex environments [42]. de Jager et al. experimentally explained Brownian movement patterns of mussels by truncations resulted from encounters with conspecific individuals, which is the original mechanism of Einstein’s collision-induced BWs [42,44]. In contrast, our findings suggested that spontaneous BWs were beneficial, and this conclusion is supported by the fact that the pattern can spontaneously change depending on internal physiological states [45,46]. Of course, our hypothesis does not contradict the claim of de Jager et al., because the spontaneous LW pattern has higher efficiency in the absence of risk.
Furthermore, our results suggest that animals can change their search strategy according to their developmental stage or in response to predator cues. For example, a juvenile individual under high predation pressure might adopt the BW strategy to avoid predator encounters, but an adult might adopt the LW strategy to obtain more targets in the absence of predators or under low predation pressure. In smaller scale responses, when an individual receives a chemical cue (kairomone) that indicates the presence of a predator, switching the internal pattern from LW to BW may represent an adaptive searching strategy, because the stochastic or random pattern can arise from internal processes [32,46–49]. Although such switching strategies depending on the target distribution have been investigated [9,13,14,42,50], the response to predators is less understood [51] and may be a topic for further study.
We introduced fitness determined not only by search efficiency but also by predation risk into the random search scenario unlike previous studies. In our assumption, the encounter with predators leads to death of the searcher with probability 1. This means that the first encounter with the predator is crucial for the searcher, and seems to be more dangerous for the searcher than the actual situation in nature because the encounter in nature does not always lead to death. If the probability is less than 1 and the searcher survives the encounter with a predator, then the searcher starts to move from the position near the predator as the simplest assumption. In this case, the problem reduces to the difference of initial positions. The result of effects of initial distance between the searcher and the nearest predator suggests that the short distance decreases the relative encounter rate with predators
γLWγBW
when the predator’s strategy is sit-and-wait (S6 Fig). Therefore, the Lévy walk strategy can temporarily benefit from departing from the close predator [30], indicating that the switch between strategies could be more efficient.
However, considering the biological plausibility, animals would not start to move around in a random manner immediately after an encounter with a predator. Instead, the searcher must depart from sit-and-wait predators using the information about the location of the predator in a deterministic manner, or dash to a safe area (e.g., bushes) to hide from moving predators and wait for the predator to leave. The predators would leave the location after some giving-up-time. The encounter event with a predator seems to transcend the simple framework of the random search problem. However, if the searcher starts random searches after fully departing from predators, the condition should not change much. Thus, probability 1 can represent several situations of prey-predator interactions.
Although we can use the probability of the survival for simplification, more complex interactions between prey and predator should occur in nature. Some empirical studies have reported the variability in predator avoidance [52,53], and theoretical studies have solved the pursue-evasion problem [36,54]. Although the issue of how the random search problem relates to such complex interactions is an interesting one, the relationship is poorly understood at present, and awaits further study.
Tracking animal movements over a prolonged period of time (biologging) is a method developed within the last decade that can lead to the understanding of dynamic phenomena ranging from the individual level to population and community levels [55,56]. Because the differences in searching strategies influence diffusiveness and movement patterns of animals, it is crucial to identify the search strategy that animals adopt in a natural environment. The tracking of animal movements within the framework of movement ecology requires information on biotic interactions and interactions between individual animals [57–59]; therefore, the context in our model should be common to various animal species in nature, because most animals are exposed to predation pressures or to the risk of death during searching. Likewise, predators may be exposed to the risks of higher-order predators. For further investigation, it will be interesting to explore the complex dynamics via the interactions between movement and population dynamics. Thus, considering ecological factors can lead to a fruitful understanding of the dynamics at various scales.
|
10.1371/journal.ppat.1006632 | Rewiring monocyte glucose metabolism via C-type lectin signaling protects against disseminated candidiasis | Monocytes are innate immune cells that play a pivotal role in antifungal immunity, but little is known regarding the cellular metabolic events that regulate their function during infection. Using complementary transcriptomic and immunological studies in human primary monocytes, we show that activation of monocytes by Candida albicans yeast and hyphae was accompanied by metabolic rewiring induced through C-type lectin-signaling pathways. We describe that the innate immune responses against Candida yeast are energy-demanding processes that lead to the mobilization of intracellular metabolite pools and require induction of glucose metabolism, oxidative phosphorylation and glutaminolysis, while responses to hyphae primarily rely on glycolysis. Experimental models of systemic candidiasis models validated a central role for glucose metabolism in anti-Candida immunity, as the impairment of glycolysis led to increased susceptibility in mice. Collectively, these data highlight the importance of understanding the complex network of metabolic responses triggered during infections, and unveil new potential targets for therapeutic approaches against fungal diseases.
| Fungal infections are a major health concern for immunocompromised individuals due to the lack of success of the currently available antifungal therapies. Unveiling the metabolic processes involved in the immune function offers a promising opportunity for the development of new therapeutic approaches against these infections. In this report, we describe how changes in monocyte glucose metabolism are crucial for host defense against infections caused by the opportunistic pathogenic yeast Candida albicans. We report how the participation of various metabolic routes, such as glycolysis, oxidative phosphorylation and the pentose phosphate pathway, were differentially required after yeast or hyphal exposure, depending on the cellular energy requirements for each response. The proper control of metabolic reprogramming of immune cells was crucial to afford protection against fungal infections in vivo.
| The immune system is constantly challenged by pathogens, and this requires immune cells to optimize the management of metabolic resources in order to exert their crucial role in host defense. A number of studies have shown how different stimuli induce metabolic reprogramming in immune cells, required for the response against microbial infections [1–3].
The recognition of pathogen-associated molecular patterns (PAMPs) triggers substantial changes in cellular metabolism of immune cells, leading to modulation of the effector functions of these cells. Recent studies led to the understanding that the differential use of carbon and nitrogen sources can subsequently affect the immune response. In this sense, proinflammatory macrophages and neutrophils favor aerobic glycolysis over oxidative phosphorylation [4], anti-inflammatory macrophages rely more on fatty acid oxidation and TCA cycle [5], whereas full T cell activation also requires the induction of mitochondrial ROS [6].
Candida albicans is a dimorphic fungus that normally colonizes skin and mucosal surfaces in the majority of the healthy population [7], but in immunocompromised hosts can cause severe life threatening infections [8]. Although the cell wall of both Candida yeast and hyphae contains a variety of glucans, mannans and glycoproteins that can be recognized by a wide range of PRRs, the expression of these molecules greatly varies between yeast and hyphal forms, leading to substantial differences in cytokine induction [9]. Monocytes undergo metabolic and functional reprogramming after exposure to β-glucans and C. albicans yeast, leading to a ‘trained immunity’ functional status characterized by an enhanced cytokine production after secondary stimulation with related or non-related stimuli [10]. In addition, monocytes have been shown to play a crucial role against C. albicans infection, as the deficiency in this immune cell subset has been related with higher susceptibility to fungal infections both in mice and humans [11,12].
Few data are available regarding the role of cellular metabolism for the immune function of monocytes, especially the impact on antifungal host defense. This prompted us to study the metabolic pathways triggered by C. albicans in human monocytes after yeast or hyphal stimulation, analyzing the different degree of engagement of the main PRRs involved in C. albicans recognition with the cellular metabolic changes induced, and to study the influence of these alterations on the cytokine response profiles. We report an association between C. albicans-specific recognition by C-type lectin receptors (CLRs) and the enhancement of glucose metabolism and aerobic glycolysis in monocytes. These metabolic changes are connected with an enhanced proinflammatory cytokine production in a differential way after yeast or hyphal stimulation. The relevance of these immunometabolic changes was validated in vivo, showing that inhibition of glucose metabolism led to impaired cytokine production, lower fungicidal activity, and a higher susceptibility to systemic C. albicans infection.
Since the molecules expressed in the cell wall of C. albicans can be recognized by a great variety of receptors that could be involved in the upregulation of different metabolic pathways, we measured genome-wide transcriptional profiles in peripheral blood mononuclear cells (PBMCs) from healthy volunteers upon stimulation with C. albicans for 4 h and 24 h. Transcriptomic analysis of the genes involved in the main cellular metabolic pathways revealed that the only pathway whose gene expression was consistently upregulated after stimulation was glycolysis, and that this enhancement occurred at 24h, but not at 4h after stimulation (Figs 1A and S1). We validated the upregulation observed in those genes by qPCR in monocytes isolated from healthy volunteers stimulated with heat-killed yeast or hyphae for 24 h finding a significant upregulation of some of the main enzymes involved in glycolysis such as hexokinase (HK) and phosphofructokinase (PFKP). We also found an upregulation of the expression of glutaminase (GLS), an enzyme that allows the entrance of glutamine into tricarboxylic acid (TCA) cycle by converting it into glutamate that is subsequently transformed into α-ketoglutarate (Fig 1B).
The upregulation of the expression of genes involved in glycolysis has been related to a boost of glucose consumption and lactate production [13]. In agreement with these data, we observed a significant increase in the lactate concentrations released in the supernatants of C. albicans-stimulated monocytes, which was accompanied by an increase in glucose consumption both after yeast and hyphal stimulation (Fig 2A), reflecting the induction of the glycolytic pathway [14]. In line with these data, the increased basal and maximal extracellular acidification rate (ECAR) values measured in these cultures reflected an enhancement of the glycolytic activity of monocytes after C. albicans yeast stimulation (Fig 2B). Importantly, the oxygen consumption rate (OCR), which is accepted to be an indicator of the oxidative phosphorylation activity [15], was also higher in monocytes that had been stimulated with C. albicans both for 4 and 24 h (Fig 2C), also reflecting an enhancement of the oxidative mitochondrial activity in C. albicans-stimulated monocytes. Of note, the OCR/ECAR ratio did not change within the time points measured reflecting a proportional increase of glycolysis and OXPHOS (S2 Fig). In addition to this, C. albicans-stimulated monocytes showed a slightly increased mitochondrial spare respiratory capacity (SRC), especially 4 h after stimulation (Fig 2D).
Since the stimulation of monocytes with C. albicans led to an increase in glycolysis and oxidative phosphorylation, we quantified the levels of different metabolites of the TCA cycle. Interestingly, after 4 h stimulation with yeast a slight decrease was observed in the levels of glutamate, fumarate and malate, metabolites than can be synthesized from glutamine [16]. On the other hand, overnight stimulation induced a general increase in the intracellular metabolite levels (Fig 2E), suggesting that within 24 h after stimulation, cells had time to induce an extensive activation of the cellular metabolic machinery and fulfill the energy requirements needed for the functional changes induced by cell activation. Of note, these data correspond to the increases in ECR and OCAR observed 4 and 24 h after stimulation (Fig 2C and 2D), reflecting an enhancement of the cellular metabolic activity after C. albicans recognition by monocytes. We identified no overall changes in cell numbers after stimulation of cells with C. albicans, we can thus conclude that cell growth or an enhanced survival of monocytes stimulated with C. albicans is not the reason of the differences observed in ECAR or OCR.
Heat-killing alters the cell wall structure and exposes antigens and PAMPs on the surface of C. albicans yeasts and hyphae [17], and we validated the results obtained with heat-killed forms of the fungus by using live yeast and hyphae. In order to distinguish between the metabolic changes induced either by yeast or hyphae, we cultured monocytes with a yeast-locked strain of C. albicans (Δhgc1) or with the hyphae-forming wild-type corresponding strain (hgc1), as the in vivo culture conditions used stimulate hyphal development from live yeast. As described for heat-killed forms, stimulation with live C. albicans increased lactate production by human monocytes (Fig 2F). Of note, the increase in lactate measured after stimulation with live fungal forms was lower than with heat-killed forms, most likely due to the effective masking of β–glucan in the cell wall of live yeasts [18]. In line with this hypothesis, the levels of intracellular metabolites 24 h after infection were significantly higher for hyphae-forming live Candida than for yeast-locked Candida (Fig 2G).
Apart from phagocytosis and killing, one of the main effector functions of monocytes and macrophages during C. albicans infection is the production of proinflammatory cytokines, required for the development of a protective immune response [8]. Therefore, we tested how the specific inhibition of various metabolic pathways affected the proinflammatory cytokine production after stimulation with C. albicans-yeast and hyphae (Fig 3A). The inhibition of glycolysis with 2-deoxyglucose (2-DG), a competitive inhibitor of hexokinase (HK), or with dichloroacetate (DCA), a compound that skews the glycolytic flux through TCA cycle by reducing the transformation of pyruvate into lactate by enhancing the activity of pyruvate dehydrogenase (PDH), strongly downregulated C. albicans-induced IL-1β, TNFα and IL-6 production in human monocytes (Fig 3B). Monocyte training with β-glucan, a ligand from C. albicans cell wall, has been reported to cause a switch from oxidative phosphorylation to aerobic glycolysis via activation of the PI3K-Akt-mTOR axis in human monocytes [19]. We found that inhibition of the mTOR pathway with Torin1 (a direct mTOR inhibitor) or with AICAR (an indirect mTOR inhibitor via AMPK activation) caused a decrease in the cytokine production by human monocytes after stimulation with yeast, but not with hyphae (Fig 3B).
In addition, inhibition of glutaminolysis by BPTES, a selective inhibitor of glutaminase (GLS), also impaired the production of IL-1β and IL-6 in yeast-stimulated monocytes, although to a lower extent (Fig 3B). The inhibition of β-oxidation with etomoxir or the interference of the pentose phosphate pathway with 6-aminonicotinamide (6-AN) did not produce any significant differences in the production of the cytokines measured (Fig 3B). Importantly, impairment of oxidative phosphorylation with oligomycin, an ATP synthase inhibitor, caused a significant decrease in the proinflammatory cytokine production after stimulation with C. albicans yeast (Fig 3B), and this effect is consistent with the increased OCR reported after C. albicans stimulation (Fig 2B). We confirmed the importance of glycolysis in these processes as 2-DG treatment of monocytes abolished the differences observed in ECAR and OCR after C. albicans recognition (Fig 3C and 3D). As a whole, these data suggest that C. albicans yeast-induced cytokine production in monocytes relied on an mTOR-dependent enhanced glycolysis as well as on an increased oxidative phosphorylation activity of the cells and, to a lesser extent, on glutamine metabolism. In the case of hyphae-induced cytokine production, these data suggest that it mostly relies on glycolysis.
Since Th1 and Th17 responses have been reported to play a protective role in C. albicans infection [20,21] we also measured how the inhibition of the different metabolic pathways affected the production of Th1/Th17-derived cytokines after C. albicans stimulation of human PBMCs. We found that the inhibition of glucose metabolism notably impaired IL-17, IL-22, IFNγ and IL-10 production, while glutamine metabolism also played a role in the production of Th17-derived cytokines after yeast but not after hyphal stimulation (S3 Fig). In the case of the anti-inflammatory cytokine IL-10, its production after hyphal stimulation was only significantly affected by glycolysis inhibition, suggesting that the metabolic pathways leading to the production of pro- or anti-inflammatory cytokines might be regulated by different metabolic routes, as already described for macrophages [22].
C. albicans recognition by human cells is known to rely on a wide series of Pattern Recognition Receptors (PRRs) such as Toll-Like Receptors (TLRs) and C-type lectin receptors (CLRs) [8]. Since C. albicans-induced cytokine production in human monocytes seemed to be under the control of glycolysis, we wondered which receptors were responsible for triggering the metabolic changes reported. To this aim, we blocked different PRRs involved in C. albicans recognition and assessed lactate production in cell supernatants after overnight culture. Interestingly, neither the treatment of monocytes with a specific mAb against TLR2, nor the blockade of TLR4 with Bartonella quintana LPS, a natural antagonist of this receptor [23] produced any changes in the lactate production triggered by heat-killed Candida stimulation, indicating that TLR-derived signaling did not play a role in the induction of glycolysis after Candida recognition (Fig 4A). Nevertheless, blockade of C-type lectin receptors with laminarin (a dectin-1 specific antagonist) or with a specific mAb against mannose receptor (MR), caused a significant decrease in the lactate production measured upon stimulation with heat-killed Candida yeasts, but not hyphal stimulation. Of note, blockade of CR3, a receptor that has a lectin domain involved in C. albicans [24] and β-glucan [25] recognition, also led to a decrease in the extracellular lactate levels determined after C. albicans yeast stimulation (Fig 4A). These results reflect that metabolic changes induced by recognition of C. albicans by monocytes were mainly driven by CLR-mediated rather than TLR-mediated signaling, in contrast to the metabolic rewiring induced by bacteria [26]. These data also confirmed the differences in the intracellular metabolic requirements triggered after yeast or hyphal recognition.
PRR blockade after monocyte stimulation with live yeast-locked C. albicans did not produce any significant changes in lactate production except for the case of monocytes with an impaired dectin-1 signaling, which showed a discrete reduction in this readout (Fig 4B). This can be again attributed to the low degree of β-glucan exposure in the cell wall of the live wild-type C. albicans [27]. The differences seen after live dectin-1 blockade in wild-type C. albicans-stimulated cells can be explained by the fact that β-glucan, the dectin-1 ligand, is highly exposed in the cell wall of newly-formed hyphae as described by Cheng et al. [18], and further confirmed that yeast and hyphae triggered metabolic changes in human monocytes in a differential fashion.
A number of studies have related the production of Reactive Oxygen Species (ROS) with the resolution of C. albicans infection [8,28]. We hypothesized that ROS induction could be affected by monocyte metabolism after C. albicans stimulation. Glycolysis inhibition with 2-DG prior to stimulation with yeast or hyphae almost completely abolished ROS production by monocytes (Fig 5B). In line with this, the impairment of the glycolytic routes with DCA treatment also impaired ROS production strongly (Fig 5C). We did not find any other metabolic pathways involved in ROS production after C. albicans stimulation (S5 Fig), except for the case of the pentose phosphate pathway, for which the specific inhibition of its oxidative branch with 6-AN led to a strong decrease in ROS production (Fig 5D). This can be related to the drop in availability of NADPH, a key factor for the induction of ROS, as already described for LPS-activated macrophages [29]. On the other hand, phagocytosis of C. albicans yeast was not significantly affected by treatment of monocytes with 2-DG (S6 Fig).
Because the data presented above argue for a crucial role of monocyte glycolysis for antifungal host defense, we wanted to validate these results in an in vivo model of systemic C. albicans infection. In this model, C57BL/6 mice were intravenously injected with a single dose of 105 colony-forming units (CFU) of C. albicans yeast, causing a disseminated infection [30]. In order to validate the role of glucose and glutamine metabolism in vivo, we treated these mice with 2-DG or BPTES prior to systemic C. albicans intravenous challenge and evaluated the systemic response to infection. Treatment with 2-DG led to a significant increase in the fungal burden measured in the kidneys of these mice 5 days after C. albicans infection, while BPTES-treated animals had a fungal burden comparable to control individuals (Fig 6A). Of note, one of the mice treated with 2-DG had to be euthanized during the experiment due to the infectious process. We also assessed the candidacidal activity of blood neutrophils, which have been described to be the main effector cells in this model of infection [8], finding a strong impairment of their fungicidal potential in the case of 2-DG treated mice (Fig 6B). Mouse neutrophils treated in vitro with 2-DG also had a significantly lower candidacidal activity than control cells (S7 Fig). We also measured cytokine production after C. albicans restimulation of splenocytes obtained from 2-DG or BPTES-treated mice after the infection, finding a significant reduction in the production capacity of IL-1β, IL-6, IL-10, TNFα and IFNγ in mice treated with 2-DG. Thus, the impairment of glucose metabolism alters the capacity of splenocytes to respond to a secondary C. albicans stimulation (Fig 6C). In mice treated with BPTES, we found reduced levels of IL-6, which is in agreement with the data obtained from human monocytes. Therefore, while the inhibition of glutamine metabolism seems to have a relatively small effect in systemic antifungal response in vivo, these data confirm that glycolysis plays a central role in the induction of an effective anti- C. albicans host response both in vitro and in vivo (Fig 7).
C. albicans is the most important fungal pathogen, and immunotherapeutic approaches to boost antifungal host defense are urgently needed to decrease mortality in systemic candidiasis which currently reaches up to 30–40% [31]. Here we aimed to investigate for the first time the role of cellular metabolism of immune cells for the induction of an effective immune defense against C. albicans. We observed that C. albicans yeast and hyphae induce differential rewiring of cellular metabolism: while yeast-stimulated monocytes rely on glycolysis, oxidative phosphorylation and glutaminolysis to mount cytokine responses, monocytes stimulated with hyphae rely mainly on glycolysis. These processes are mediated by fungal recognition by CLRs, but not by TLRs, and glycolysis is crucial for an effective host defense in vivo against disseminated candidiasis.
The metabolic circuits triggered in immune cells following pathogen recognition are very complex. Microbial stimuli such as LPS promote glucose conversion into lactate and decrease oxidative phosphorylation in monocytes and macrophages, in a process known as the Warburg effect [4]. However, a recent study demonstrated that induction of the Warburg effect in monocytes is a specific feature of LPS engagement of TLR4, as the engagement of other TLRs by their specific ligands or by complete microorganisms led to a much more complex response, notably a strong increase in the oxidative phosphorylation activity of the cells after stimulation [26]. While several studies investigated the immunometabolic circuits involved in anti-bacterial responses, very few have addressed the role of these pathways in anti-fungal immunity. In this report, we describe differential roles for glucose metabolism, glutamine metabolism, oxidative phosphorylation, and the pentose phosphate pathway in the immune response against C. albicans.
The increase of glycolysis following microbial stimulation of myeloid cells has been reported in several studies, being usually linked to an enhanced function of mTOR-related pathways [19,32]. Our data confirmed that after C. albicans recognition, human monocytes underwent an increase not only in their glycolytic activity, as demonstrated by the higher glucose consumption, lactate production and ECAR reported after stimulation, but also in their oxidative phosphorylation capacity and OCR. This enhancement of both aerobic glycolysis and oxidative phosphorylation is reminiscent of the metabolic stimulation induced by other whole microorganisms [16,26], in contrast to purified ligands such as LPS or β-glucan [19,33], reflecting the complex nature of the cellular metabolic networks induced by the engagement of different PRRs.
An important discovery is the difference in induction of cellular metabolism between C. albicans yeasts and hyphae. The differences in the abundance and the degree of exposure of cell wall components such as β-glucans, mannans or glycoproteins between yeasts and hyphae have been described to induce distinct profiles of cytokine responses [9]. Along this line, we showed that stimulation with yeast or hyphae led to different metabolic responses in monocytes. Heat-killed C. albicans yeast induced production of proinflammatory cytokines, through a process highly demanding for the cells, which makes use of glycolysis, oxidative phosphorylation and glutaminolysis in order to fulfill the expensive energy requirements. In the case of heat-killed C. albicans hyphae, cytokine production was solely dependent on glycolysis. The transition to hyphal growth creates the opportunity for improved recognition of inner cell wall components such as β–glucan, which become more exposed, allowing their interaction with PRRs and triggering the proinflammatory cytokine production [18]. The data presented here, correlates with previously published data showing a strong induction of glycolysis by β–glucan [19] and argues for a model in which the poor β–glucan recognition in yeasts [18] requires different recognition systems for activation of multiple metabolic pathways, while the broad exposure of β–glucan by hyphae induces a much stronger induction of glycolysis in the immune cells which is sufficient to fulfill their energy requirements.
Glucose that enters the cell is not solely processed via glycolysis, but can be also transformed into fatty acids, aldoses, glycogen, or enter the pentose phosphate pathway to generate NADPH and pentoses [34]. ROS production has been historically linked to NADPH oxidase and mitochondrial activity [35], and several studies have assigned an important role for ROS production in host defense against microbial infections [29,36]. In our study, we found that glucose metabolism is crucial for the generation of ROS in human monocytes after C. albicans stimulation. Nonetheless, this could not be only attributed to glycolysis, as the impairment of the pentose phosphate pathway also induced a severe impairment of ROS production, as similarly described for LPS-activated macrophages, where ROS were generated through pentose phosphate pathway-dependent NADPH-oxidase activity [29]. These data suggest a differential role for the different metabolic pathways triggered after C. albicans stimulation in monocytes. While glycolysis appeared to play a central role in cytokine and ROS production, other routes such as oxidative phosphorylation or glutaminolysis seemed to play a preferential role in fueling cytokine induction. In contrast, the pentose phosphate pathway, which did not play a role in cytokine production, seems fundamental for ROS generation. These results further emphasize the variety and the specificity of the metabolic changes that cells undergo after making contact with a pathogen, highlighting the necessity of studying the distinctive features of the various stimuli.
The presence of suitable carbon sources in the environment is fundamental not only for the host but also for Candida cells. Regarding this, some studies have shown that the exposure of Candida to high concentrations of lactate is able to modulate its cell wall architecture, drug resistance and virulence [37,38]. In our study, the concentrations of lactate reached after stimulation are much lower than those demonstrated to alter cell wall structure of C. albicans, being glucose the major carbon source in all the experimental conditions tested. Therefore, in this case we find improbable that any of the experimental conditions might have been affected by the minor presence of lactate in the medium, compared to glucose.
TLR-mediated immunometabolic reprogramming have been reviewed by a number of authors [39–41]. C. albicans expresses a large variety of structures in its cell wall, which represent PAMPs that bind to different families of PRRs on the immune cells, among which CLRs and TLRs are the most important [8]. Especially deficiencies in CLRs lead to an impaired cytokine production and higher susceptibility to C. albicans infections [42,43]. In line with this, blockade of recognition of fungal β-glucan and mannans by C-type lectins led to a decrease in the glycolytic activity of monocytes triggered after stimulation with heat-killed yeast or with live wild-type C. albicans. Of note, the different background of the strains used can contribute to the complexity of the interpretation of the results obtained in this work. Besides this, our results suggest that the impact of metabolism in the outcome of Candida infection is not strain-specific, as experiments with heat-killed stimuli and with live fungi were carried out with strains with different backgrounds (UC820 in case of heat-killed fungi and SC5314 in case of live fungi).
We validated in vivo results on the role of cellular metabolism for antifungal host defense in a mouse model of systemic C. albicans infection, in which glycolysis-mediated mechanisms were demonstrated to be crucial for the defense against the pathogen. Inhibition of glycolysis led to hosts that were significantly more susceptible to the infection, presenting lower fungicidal activity and defective cytokine production capacity. Our data also suggest that glutamine metabolism could play a role in the production of IL-1β and its known downstream target IL-6 [44] after C. albicans recognition both in human and mouse. The relationship between IL-1β and IL-6 could be also playing a role in the effects observed on Th1 and Th17-derived cytokines in human PBMC cultures, whose decrease could be due to direct effects of the pharmacological inhibitors employed or to indirect effects of the dampening of IL-1β and IL-6 released from monocytes. Our results reflect the concept that the inhibition of glucose metabolism during C. albicans infection has an impact on the immune system at different levels, as the impairment of glycolysis decreased the ability to fight candidiasis both by direct and indirect mechanisms, as previously described for NK cells in an in vitro study [45]. In this sense, treatment with 2-DG induced a decrease in the production of monocyte-derived cytokines proved crucial to boost Candida clearance by neutrophils such as IL-1β, IL-6 and TNFα[8,46,47] and also reduced the fungicidal potential of neutrophils directly treated with this metabolic inhibitor. These results suggest that the establishment of a functional response against systemic C. albicans infections in vivo largely relies on the proper functioning of glucose metabolism in different immune subsets, demonstrating the importance of glycolysis in the development of an efficient antifungal response and highlighting the central role of metabolism as a cornerstone of the immune function.
The different pathways of the cellular metabolism are connected through a very complex network of enzymes and mediators. Our findings suggest that the immune functions of monocytes in C. albicans infection rely on the activation of glucose metabolism, but also required the participation of other additional metabolic pathways such as glutaminolysis or the pentose phosphate pathway. These results also delve into the distinctive features of CLR-mediated antifungal responses and highlight the need of studying the particular characteristics of the metabolic mechanisms underlying the immune responses against the wide variety of human pathogens. As combining effective anti-fungal treatment with adjuvant immunotherapy is proposed to improve the poor outcome in disseminated C. albicans infections, these data suggest that cellular metabolism of immune cells may represent a novel potential therapeutic target.
Buffy coats from healthy donors were obtained after written informed consent (Sanquin Blood Bank, Nijmegen, the Netherlands). Samples were anonymized to safeguard donor privacy. The use of the samples received IRB approval. Peripheral blood mononuclear cell (PBMC) isolation was performed by dilution of blood in pyrogen-free PBS and differential density centrifugation over Ficoll-Paque (GE Healthcare). Cells were washed three times in PBS. Percoll isolation of monocytes was performed as previously described (Repnik et al., 2003). Briefly, 150–200 x 106 PBMCs were layered on top of a hyper-osmotic Percoll solution (48.5% Percoll [Sigma-Aldrich], 41.5% sterile H2O, and 0.16 M filter-sterilized NaCl) and centrifuged for 15 min at 580g. The interphase layer was isolated and cells were washed with cold PBS. Cells were re-suspended in RPMI culture medium (RPMI medium Dutch modified, Invitrogen) supplemented with 50 μg/mL gentamicin, 2 mM Glutamax, and 1 mM pyruvate, and counted. An extra purification step was added by adhering Percoll-isolated monocytes to polystyrene flat bottom plates (Corning) for 1 h at 37°C; a washing step with warm PBS was then performed to yield maximal purity.
Candida albicans UC820 (ATCC MYA-3573) [48] was grown overnight to generate yeast cells in Sabouraud dextrose broth at 29°C, with shaking. Cells were harvested by centrifugation, washed twice with PBS, and re-suspended in culture medium (RPMI 1640 Dutch modification). To generate hyphae, yeast cells were inoculated and grown overnight at 37°C in culture medium adjusted to pH 6.4 with hydrochloric acid. C. albicans yeast or hyphae were heat-killed for 30 min at 95°C. The hyphae-specific G1 cyclin-null hgc1Δ mutant and hgc wild-type strains, kindly provided by Dr. Yue Wang (Institute of Molecular and Cell Biology, Singapore [49]) were grown under similar conditions.
PBMCs at 5 x 106 cells/mL were stimulated for 4 or 24 h with RPMI or 105 heat-killed C. albicans yeast. Global gene expression was profiled using Illumina Human HT-12 Expression BeadChip according to manufacturer’s instructions. Image analysis, bead-level processing and quantile normalization of array data were performed using the Illumina LIMS platform, BeadStudio. Only samples that had cells for which paired fold-changes could be calculated were used (i.e. for which cells from the same individual were used for both RPMI and Candida stimulation). At 4 h there were 19 matched samples, and at 24 h there were 29 matched samples. Log2 fold changes (gene expression after Candida divided by expression after RPMI stimulation) were calculated for each individual separately. The average of these values was used in the plot. Gene identifiers were mapped to entrez ids, and formatted to match the transcript gene identifiers as defined in “RECON1”, the model that contains the enzymes associated with each reaction. Gene expression values were always mapped to alternative transcript 1, i.e. all entrez identifiers were appended with “_AT1”.
The tool “Escher” [50] was used to generate a pathway to visualize the expression data containing the most interesting parts of it.
100 μL monocytes at 1 x 106 cells/mL or PBMCs at 5 x 106 cells/mL were added to flat-bottom or round-bottom 96-well plates (Greiner), respectively. Cells were incubated with culture medium with 10% serum only as a negative control or incubated with 11 mM 2-deoxyglucose (Sigma), 100 nM Torin1 (Tocris), 1 mM AICAR (5-aminoimidazole-4-carboxamide-1-β-D-ribofuranoside, Brunschwig Chemie, Amsterdam, The Netherlands), 50 nM BPTES (Sigma), 3 μM potassium dichloroacetate (DCA, Sigma), 500 nM 6-aminonicotinamide (6-AN, Sigma), 10 μM etomoxir (Sigma), 1 μM oligomycin (Sigma). After 1 h, cells were stimulated with 105 heat-killed Candida yeast, 105 heat-killed Candida hyphae, 104 live hgc1 Candida yeast or 104 live Δhgc1 Candida yeast. The different concentrations of Candida used, were based on optimization experiments and employed as seen in works from different groups [51–53]. Similar concentrations of yeasts and hyphae were employed at similar concentrations in order to get a better approach to in vivo situations. Supernatants from monocytes were collected 24 h after stimulation. Supernatants from PBMCs were collected 7 days after stimulation, except for IL-10 production, which was measured after 48 h. Concentrations of inhibitors were selected as being the highest non-cytotoxic concentrations (see S8 Fig). All supernatants were stored at -20°C until analyzed.
For receptor blockade experiments, before stimulation with C. albicans, monocytes were preincubated for 1 h with 100 μg/mL laminarin (Sigma), 10 μg/mL anti-CR3 antibody and control IgG (R&D), 100 ng/mL B. quintana LPS, 10 μg/mL TLR2-blocking antibody (anti-TLR2) and its control IgA1 (InvivoGen, San Diego, CA), 5 μg/mL MR-blocking antibody (anti-MR, R&D) and IgG1 isotype control (BD Biosciences).
Cytokine production from human cells was determined in supernatants using commercial ELISA kits for IL-1β, TNFα, IL-17A, IL-22 (R&D Systems, Minneapolis, MN) IL-6, IFNγ, and IL-10 (Sanquin, Amsterdam, The Netherlands), following the instructions of the manufacturer.
Lactate was measured from cell culture supernatants using a coupled enzymatic assay in which lactate was oxidized and the resulting H2O2 was coupled to the conversion of Amplex Red reagent to fluorescent resorufin by HRP (horseradish peroxidase) [54]. Glucose consumption was measured according to the manufacturer instructions using the Amplex Red Glucose/Glucose Oxidase Assay Kit (Life Technologies). Glutamate, fumarate, malate, α-ketoglutarate and succinate concentrations were determined by commercial assay kits (all Sigma) following the instructions of the manufacturer from at least one million monocytes lysed in 1 mL 0.5% Triton-X in PBS at 4 and 24 h after stimulation.
Real-time analysis of ECAR, OCR and SRC on monocytes was performed using an XF-96 Extracellular Flux Analyzer (Seahorse Bioscience) as described in Lachmandas et al., 2016. CD14+ monocytes were purified from freshly isolated PBMCs using MACS microbeads for positive selection, according to the manufacturer’s instructions (Miltenyi Biotec). These monocytes were seeded in quintuplicate in XF-96 cell culture plates (2 × 105 monocytes/well) in the presence of RPMI or C. albicans for 4h or 24 h in 10% human pooled serum. For the measurements of oxygen consumption and acidification rates it is therefore important to have a homogenous cell population, in this case monocytes. The CD14+ isolation is performed on PBMCs, which contain usually less than 5% of neutrophils. With subsequent CD14+ selection we obtain 95% of monocytes, so the amount of neutrophils and lymphocytes in the CD14+ selected cells is negligible. The metabolic rates of monocytes were analyzed in four consecutive measurements in XF Base Medium (unbuffered DMEM with 5.5 mM glucose and 2 mM L-glutamine, pH adjusted to 7.4). After three basal measurements, three consecutive measurements were taken following the addition of 1.5 μM oligomycin, 1 μM carbonyl cyanide-4-(trifluoromethoxy) phenylhydrazone (FCCP), 2 μM antimycin together with 1 μM rotenone, glucose (20 mM), pyruvate (1 mM) and/or 50 mM 2-DG in order to determine basal and maximum OCR and ECAR. SRC was determined as the absolute increase in OCR after FCCP injection compared with basal OCR. All compounds used during the Seahorse runs were acquired from Sigma-Aldrich.
Cells were cultured as described above. After 4 h and 24 h mRNA was extracted by TRIzol (Life Technologies), according to the manufacturer’s instructions, and cDNA was synthesized using iScript reverse transcriptase (Invitrogen). Relative mRNA levels were determined using the Applied Biosciences StepOne PLUS and the SYBR Green method (Invitrogen). Values are expressed as fold increases in mRNA levels, relative to those in non-stimulated cells, with HPRT as housekeeping gene. Primers are listed in S1 Table.
Oxygen radical production levels of isolated monocytes were evaluated using luminol-enhanced chemiluminescence and determined in an automated LB96V Microlumat plus luminometer (EG & G Berthold, Bald Wilberg, Germany) as previously described [55]. Briefly, monocytes (1 × 105 per well) were seeded into 96-well plates and incubated in medium containing either RPMI, phorbol 12-myristate 13-acetate (PMA; 5 μg/ml), heat-killed C. albicans yeast or heat-killed C. albicans hyphae (107 CFU/ml). Luminol was added to each well in order to start the chemiluminescence reaction. Each measurement was carried out in at least duplicate repetitions. Chemiluminescence was determined every 145 s at 37°C for 1 h. Luminescence was expressed as relative light units (RLU) per second.
Monocytes and PBMCs were isolated from blood donated by healthy volunteers after written informed consent. Ethical approval was obtained from the CMO Arnhem-Nijmegen (NL32357.091.10). Buffy coats from healthy donors were obtained after written informed consent (Sanquin Blood Bank, Nijmegen, the Netherlands). Samples were anonymized to safeguard donor privacy. The use of the samples received IRB approval.
All animal work was approved by the Animal Care and Use Committee of the Centro Nacional de Biotecnología—CSIC (protocol number 312–2014) in accordance with Spanish RD 1201/2005 and international EU guidelines 2010/63/UE about protection of animals used for experimentation and other scientific purposes and Spanish national law 32/2007 about animal welfare in their exploitation, transport and sacrifice.
8–12 week-C57BL/6J mice were randomized and treated with a daily intraperitoneal dose of 100 mg/kg 2-deoxyglucose (n = 6) or 100 μg BPTES (n = 6) every morning for 5 days starting at the same day with the C. albicans intravenous infection. PBS was injected as a control (n = 6). C. albicans SC5314 yeast were grown on YPD plates (Sigma-Aldrich, St Louis, MO) at 30°C for 48 h. Then, C. albicans cells were centrifuged, washed in PBS and counted using a hematocytometer. Mice were infected by intravenous injection of 1 x 105 C. albicans yeast via the lateral tail vein and daily monitored for health and survival following the institutional guidance. After 5 days, mice were euthanized in a CO2 rodent euthanasia chamber and kidneys were aseptically removed, weighed and homogenized in PBS using a T10 basic Ultra-Turrax homogenizer (Ika, Staufen, Germany). Fungal burden was determined by plating organ homogenates in serial dilutions on YPD plates. Colony forming units (CFUs) were counted after growth for 48 h at 30°C.
For ex vivo stimulation experiments splenocytes were obtained from mice at day 5 of i.v. infection with C. albicans and stimulated ex vivo with LPS (10 ng/mL) heat-killed Candida yeast (1 × 107/mL) or heat-killed C. albicans hyphae (1 × 106/mL). Splenocytes were obtained by gently squeezing spleens in a sterile 100 mm filter. After centrifugation and washing, splenocytes were resuspended in complete RPMI 1640 medium supplemented with 10% FCS, 2 mM L-glutamine, 100 U/mL penicillin, 100 μg/mL streptomycin and 50 μM 2-mercaptoethanol, and counted using a hematocytometer. Splenocyte concentration was adjusted to 5 × 106/mL. 200 μL of the cell suspension were cultured in round-bottom 96-well plates (Corning, Durham, NC) and stimulated with RPMI or 1 × 106 heat-killed C. albicans yeast or hyphae/mL. The measurement of cytokine concentrations was performed in supernatants collected after 48 h of incubation at 37°C in 5% CO2. Cytokine production from mouse cells was determined in supernatants using commercial ELISA kits for IL-1β, TNFα, IL-6, IFNγ and IL-10, all from BD Pharmingen (San Diego, CA).
Circulating neutrophils were isolated from blood drawn by cardiac puncture, diluted in PBS containing 5 mM EDTA and 3% FCS, overlaid over a density gradient of Histopaque 1119 and Histopaque 1077 (Sigma) and centrifuged for 30 minutes at 400 g. Neutrophil preparations had a purity >80%. To test neutrophil killing activity, 5 x 104 C. albicans yeast were exposed to 104 neutrophils for 2 h; neutrophils were then lysed with water and the number of surviving yeast cells was assessed on YPD agar. Killing activity was expressed as the percentage of C. albicans cells surviving in the presence of neutrophils compared to C. albicans cells surviving in the absence of neutrophils.
Percoll monocytes were plated in 96 flat bottom plates at 1x105 cells / well. Cells were allowed to phagocytose 1 x 106 (MOI 1:10) heat inactivated FITC-labeled C. albicans for 2h in the presence or absence of α-antitrypsin 10 mg/mL or 100 mg/mL. Subsequently, the fluorescence signal of extracellular non-phagocytosed Candida was quenched using trypan blue. The monocytes that phagocytosed one or more C. albicans yeast were enumerated by their positivity for the FITC signal, and could be divided into two populations: FITC- monocytes (those that did not engulf C. albicans) and FITC+ monocytes (those that did).
Cell viability was assessed using Annexin-V (Biovision, San Francisco, CA) and propidium iodide (Sigma) staining. Cells were stained for 15 minutes using Annexin-V-FITC using the protocol supplied by the manufacturer to detect early apoptotic cells. Subsequently cells were stained with for 5 minutes in 10 ug/mL propidium iodide. Cells were assessed for annexin-V and PI positivity using a FC500 flow cytometer (Beckman Coulter). Annexin-V+ cells were considered as early apoptotic cells and Annexin-V+ PI+ cells were considered as late apoptotic cells.
|
10.1371/journal.pntd.0000495 | Preclinical Assessment of the Treatment of Second-Stage African Trypanosomiasis with Cordycepin and Deoxycoformycin | There is an urgent need to substitute the highly toxic compounds still in use for treatment of the encephalitic stage of human African trypanosomiasis (HAT). We here assessed the treatment with the doublet cordycepin and the deaminase inhibitor deoxycoformycin for this stage of infection with Trypanosoma brucei (T.b.).
Cordycepin was selected as the most efficient drug from a direct parasite viability screening of a compound library of nucleoside analogues. The minimal number of doses and concentrations of the drugs effective for treatment of T.b. brucei infections in mice were determined. Oral, intraperitoneal or subcutaneous administrations of the compounds were successful for treatment. The doublet was effective for treatment of late stage experimental infections with human pathogenic T.b. rhodesiense and T.b. gambiense isolates. Late stage infection treatment diminished the levels of inflammatory cytokines in brains of infected mice. Incubation with cordycepin resulted in programmed cell death followed by secondary necrosis of the parasites. T.b. brucei strains developed resistance to cordycepin after culture with increasing concentrations of the compound. However, cordycepin-resistant parasites showed diminished virulence and were not cross-resistant to other drugs used for treatment of HAT, i.e. pentamidine, suramin and melarsoprol. Although resistant parasites were mutated in the gene coding for P2 nucleoside adenosine transporter, P2 knockout trypanosomes showed no altered resistance to cordycepin, indicating that absence of the P2 transporter is not sufficient to render the trypanosomes resistant to the drug.
Altogether, our data strongly support testing of treatment with a combination of cordycepin and deoxycoformycin as an alternative for treatment of second-stage and/or melarsoprol-resistant HAT.
| There is an urgent need to substitute the highly toxic arsenic compounds still in use for treatment of the encephalitic stage of African trypanosomiasis, a disease caused by infection with Trypanosoma brucei. We exploited the inability of trypanosomes to engage in de novo purine synthesis as a therapeutic target. Cordycepin was selected from a trypanocidal screen of a 2200-compound library. When administered together with the adenosine deaminase inhibitor deoxycoformycin, cordycepin cured mice inoculated with the human pathogenic subspecies T. brucei rhodesiense or T. brucei gambiense even after parasites had penetrated into the brain. Successful treatment was achieved by intraperitoneal, oral or subcutaneous administration of the compounds. Treatment with the doublet also diminished infection-induced cerebral inflammation. Cordycepin induced programmed cell death of the parasites. Although parasites grown in vitro with low doses of cordycepin gradually developed resistance, the resistant parasites lost virulence and showed no cross-resistance to trypanocidal drugs in clinical use. Our data strongly support testing cordycepin and deoxycoformycin as an alternative for treatment of second-stage and/or melarsoprol-resistant HAT.
| Human African trypanosomiasis (HAT), also known as sleeping sickness, is a neglected tropical infectious disease that has re-emerged in sub-Saharan Africa in the late 1900ties [1],[2]. The disease is caused by subspecies of the protozoan parasite Trypanosoma brucei and is transmitted by the blood-sucking tsetse fly. Following a hemo-lymphatic stage with waves of parasitemia, the nervous system is involved manifested as a leukencephalitis, invariably lethal if left untreated. Drugs, which are effective at an early stage of the disease, but poorly penetrate the blood-brain barrier, are ineffective for the second stage. For treatment of the second encephalitic stage of HAT, arsenic compounds such as melarsoprol, which are associated with severe and even lethal side-effects, are still widely used. Moreover, there has been an alarming increase in melarsoprol-refractory sleeping sickness cases [3]. Dl-α-difluoromethylornithine (DFMO) is used for treating the West African form of HAT, caused by T.b. gambiense. However, this drug is given by intravenous injections, is expensive, and is not effective against T.b. rhodesiense that causes the East African form of HAT. Since the development of new drugs for HAT is not likely to occur in the immediate future, the strategy of testing drugs approved or in clinical use for other diseases should be pursued in order to identify less toxic or alternative drugs to cure second stage HAT [4].
In contrast to most mammalian cells, trypanosomes cannot synthesize purines de novo. Instead they depend on the salvage pathway of nucleosides from the body fluids of the host [5]. The inability of trypanosomes to engage in de novo purine synthesis has been exploited as a therapeutic target. The trypanocidal potential of cordycepin (3′-deoxyadenosine), a metabolite from the fungi Cordyceps spp., was noted in experiments performed in the 1970's [6],[7]. However, administration of cordycepin did not result in a complete cure from infection [8], since cordycepin is rapidly converted to inactive 3′-deoxyinosine by adenosine deaminase (ADA) in vivo [9]. Deoxycoformycin is an ADA inhibitor that can prevent degradation of cordycepin, and have come into use in combination with cordycepin for the treatment of certain malignant tumors in humans, e.g. leukemia and melanoma [10].
Cordycepin is probably taken up by trypanosomes through transporters. Nucleoside transport systems, P1 and P2, are able to concentrate cordycepin inside the cells and has turned out to play an important role in the uptake of trypanocides [11]. Adenosine is salvaged through a two-step process in T. brucei where its intracellular cleavage to adenine is followed by phosphoribosylation to AMP, or by a high affinity adenosine kinase [12],[13]. Cordycepin is also a potent inhibitor of the mammalian poly-A polymerase, and has been shown to hamper the activity of nucleoside-stimulated protein kinase activity in Trypanosoma sp. [14].
We have previously shown that intraperitoneal (i.p.) injection of cordycepin, together with coformycin or deoxycoformycin, can cure T.b. brucei infections in mice [15]. Treatment was effective, as recorded by the absence of parasitemia up to three weeks after the end of the treatment, even when instituted after the trypanosomes had penetrated into the brain parenchyma. The aim of the present work was to further improve treatment strategies with nucleosides, and gain knowledge on the effect of these compounds on the parasite. We could show that cordycepin is the best candidate drug selected after screening a library of 2200 nucleoside analogues for parasite differential toxicity. The minimal inhibitory concentrations of cordycepin and deoxycoformycin and the minimal number of doses required for effective treatment were then determined in mice infected with T.b. brucei. Cordycepin and deoxycoformycin treatment did also cure second stage infection with human pathogenic T.b. gambiense and T.b. rhodesiense, and oral administration of the duplet when combined with a proton pump inhibitor was effective. Cultivation of T.b. brucei with low doses of cordycepin allowed resistance to cordycepin to develop, but resistant parasites lost virulence, and no cross-resistance between cordycepin and melasoprol. suramin and pentamidine was noted. Importantly treated mice showed diminished levels of pro-inflammatory cytokines in the brain.
Altogether, our findings encourage the possible use of cordycepin and deoxycoformycin as an alternative for treatment of second stage HAT or melasoprol-resistant HAT.
Cordycepin, suramin, pentamidine, omeprazole and EHNA (erythro-9-(2-hydroxy-3-nonyl)adenine) were all purchased from Sigma (St Louis, MO). Deoxycoformycin (pentostatin) was a kind gift of R. McCaffrey (Brigham & Women's Hospital, Harvard Medical School, Boston, Massachusetts). Melarsoprol (Arsobal) was a gift from P. Simarro (WHO/NTD, Geneve, Switzerland). Unless otherwise indicated, all reagents were diluted in PBS, fractioned and stored at −20°C until further use.
The library of nucleoside analogs used consisted of compounds design to be inhibitors of various viral and microbial enzymes and is property of Medivir AB.
BALB/c, C57Bl/6 and RAG1−/− [16] (backcrossed on a C57Bl/6 background) male mice, 8–10 weeks old at the beginning of the experiments, were used. Mice were kept with food and water ad libitum under specific pathogen-free conditions. All experiments were conducted following protocols that received institutional approval and authorization by the Stockholm Region's animal protection committee.
T.b. brucei (AnTat1.1E), T.b. rhodesiense (STIB 851) and T.b. gambiense (MBA) were kindly provided by P. Büscher (Institute of Tropical Medicine, Antwerp, Belgium). Tbat1 null parasites constructed by sequential homologous recombination of T.b. brucei Lister 427 [17], were kindly provided by P. Maser (University of Bern, Switzerland). Lister 427 parasites were used as controls. Parasitemia was measured every 2 or 3 day in tail vein blood. Body weight and morbidity were regularly recorded.
Bloodstream forms of T.b. brucei, T.b. gambiense and T.b. rhodesiense freshly isolated from infected C57BL/6 mice (3 days post infection with 1×107 parasites/mouse) were separated by DEAE-cellulose chromatography under sterile conditions [18]. T. brucei parasites were incubated in D-MEM containing 10% heat-inactivated calf serum, 28 mM HEPES, 0.14% glucose, 1.5% NaHCO3, 2 mM L-glutamate, 0.14 mg/ml gentamycin, 0.3 mM dithiothreitol, 1.4 mM sodium pyruvate, 0.7 mM L-cysteine, 28 µM adenosine, 14 µM guanosine at 37°C.
To measure drug sensitivity, 2.5×104 parasites were cultured in 96 flat-bottomed well culture plates with serial drug dilutions for 72 h at 37°C. Cultures (100 µl) were incubated for 2 h with 10 µl of WST-1 reagent (Roche, Mannheim, Germany). Viability was measured by the conversion of WST-1 reagent to formazan, recorded by multiwell scanning spectrophotometer at an excitation wavelength of 450 nm. A direct correlation between WST-1 reagent oxidation and parasite numbers was confirmed (data not shown).
Mice were deeply anesthetized with isoflurane, sacrificed and brains were dissected and snap frozen. To examine presence of trypanosomes within and outside the blood vessels in the brain, cryostat sections were cut and immunolabelled with anti-AnTat1.1 VSG (1∶5.000; obtained from N. van Meirvenne, Institute of Tropical Medicine, Antwerp, Belgium) and goat polyclonal anti-glucose transporter 1 (1∶40; GLUT-1, Santa Cruz Biotechnology, Santa Cruz, CA, USA) as described previously [19].
Stable recombinant Renilla luciferase expressing parasites were generated as recently described (F Claes, submitted). The use of such luciferase tagged T. brucei for real time studies of parasite dynamics was validated in vitro and ex vivo (F. Claes, PLoS NTD in press). BABLB/c mice were infected i.p. with 2×103 luciferase tagged T.b. brucei AnTat1.1. At different days after infection, mice were anesthetized with 2.3% isoflurane, injected intraperitoneally with 100 µL of coelenterazine (2 µg/µl dissolved in methanol) (Synchem) diluted with 90 µL PBS pH 7, and light emission in photons/second/cm2/steradian (p/sec/cm2/sr) was recorded in an IVIS Imaging System 100 (Xenogen LifeSciences) and Living Image 2.20.1 software (Xenogen) for 180 seconds. Measurements started 3–5 minutes after substrate injection to allow the spread of the coelenterazine.
Gene transcripts of several pro-inflammatory cytokines were quantified in brains from cordycepin-treated and PBS-treated, uninfected and infected mice by real time PCR. Total RNA was extracted from half of the fresh frozen brains, reverse-transcribed, and the transcripts levels quantified on an ABI Prism 7000 sequence detection system (Applied Biosystems) as described previously [16]. Ten-fold dilutions of a cDNA sample were amplified to control amplification efficiency for each primer pair. Thereafter, the threshold cycle value, Ct (the fractional cycle number at which the fluorescence reaches a fixed threshold) was obtained for all cDNA samples. The amount of transcripts of individual animal samples (n = 4 per group) was normalized to HPRT (ΔCt). The relative amount of target gene transcripts was calculated using the 2−ΔΔCt method as described [22]. These values were then used to calculate the mean and standard error of the relative expression of the target gene mRNA in the brain of uninfected and infected mice.
Total DNA was extracted from brains from treated and untreated mice using a Dneasy tissue kit (Qiagen, Duesseldorf, Germany). The amount of parasite DNA was quantified by real time PCR as described above using senso latu T. brucei primers [23]. The concentration of DNA in each sample was measured in a spectrophotometer. A correlation between parasite numbers and quantification of parasite DNA by real time PCR was determined.
Total DNA from parental and resistant strains was extracted using the DNeasy tissue kit (Qiagen). Candidate genes suspected to be involved in the resistance to cordycepin were PCR amplified from total DNA using forward and reverse specific primers designed over the following annotated genes by the T. brucei genome project (Table S1): Adenosine kinase, Adenine phosphoribosyltransferase (APRT), Hypoxanthine-guanine phosphoribosyltransferase (HGPRT), Poly(A) polymerase, Hexose Transporter 1 (HT1), Aminopurine Transporter (AT1 or P2), P1 family of nucleoside transporters (NT2, NT4, NT6 and NT12). PCR products were A-tailed and ligated into pGEM T-vector (Promega, Madison, WI). A set of 6 to 12 positive colonies for each gene was sequenced using the DYEnamic ET dye terminator kit and a Megabase sequencer (GE) with m13 forward and reverse primers, and when needed gene specific internal primers. Sequences were assembled and analyzed using phred-phrad and consed. Sequence analysis and comparisons between parental and resistant strains and genome-annotated versions were performed with ClustalW (www.ebi.ac.uk) and Blast2seq (www.ncbi.nlm.nih.gov). Single nucleotide polymorphisms between clones were taken into account.
We have previously suggested that cordycepin could be used for treatment of second stage African trypanosomiasis. Whether other nucleoside analogues could have a more selective toxic effect on T.b. brucei was first studied. For this purpose we screened a library of 2200 drug-like nucleosides compounds for toxicity against T.b. brucei AnTat1.1. The screening revealed 14 hits showing toxicity for T.b. brucei at less than 1 µM concentration. Ten of these compounds were found to also be toxic for mammalian cells at these concentrations. One of the remaining compounds was identified as cordycepin. A second one was tubercidin, a nucleoside previously shown to inhibit the respiratory chain in T.b. brucei, but that induced adaptation of parasites to a glucose-independent metabolism [24]. The two other compounds were diamines, N′trityl-1,2-diaminoethane hydrobromide and N-trityl-1,3-siaminopropane acetate, included in the library as intermediates used in the chemical synthesis of nucleoside analogues.
Subsequently, the trypanocidal activity of cordycepin was compared with that of adenosine deaminase resistant nucleoside analogues. Such nucleosides could potentially act as stand-alone drugs in treatment of T.b. brucei infections. Three out of 22 tested compounds showed at least 3 log higher toxicity for T.b. brucei as compared to mammalian HL-cells (Figure 1A–B and data not shown). One is 2-fluorodeoxyadenosine and the other are two synthetic cordycepin derivatives: 6-N-(phenylthio)-methyl-cordycepin (cordycepin 110) and 6-N-(4-N-acethylbenzylthio)-methyl-cordycepin (cordycepin 116). As expected, deoxycoformycin was not toxic for T.b. brucei (data not shown).
The two synthetic derivatives of cordycepin were tested in vivo. Cordycepin 116, but not cordycepin 110, showed anti-trypanosomal effect during the early stage of infection (Figure 1C). However, no curative effect was observed when 2 mg/kg/d cordycepin 116 was inoculated during 10 days in the absence of deoxycoformycin and starting 20 days after infection with T.b. brucei (Figure 1D). Thus, we decided to restrict our further evaluation of the effects of treatment to cordycepin during infection with T. brucei on the host and the parasite.
Experiments were then performed to confirm, using more extended protocols, the curative effect of cordycepin and deoxycoformycin on the second stage murine infection with T.b. brucei. In order to detect the presence of remnant parasites in the tissues of treated animals, immunodeficient RAG-1−/− mice were inoculated i.p. with blood or brain tissue homogenates of infected and treated mice. A double immuno-labelling method using antibodies against trypanosomes and cerebral endothelial cells was used to detect parasites in the brain parenchyma. Stained sections from brain of treated animals (4 per animal) were analyzed. Parasite DNA in whole blood or brain tissues was measured by real time PCR.
In a first experiment we showed that none of the C57Bl/6 mice (n = 9) treated i.p. with 2 mg/kg body weigth (kg)/day (d) cordycepin and 0.2 mg/kg/d deoxycoformycin for 7 consecutive days, starting at day 20 after infection with 2×103 T.b. brucei. displayed parasites at 85 days of infection (65 days after treatment). No parasites were detected in the blood of cordycepin- and deoxycoformycin-treated mice. On the contrary, although no parasites were detected directly after treatment with 20 mg/kg suramin, the presence of parasites was recorded at least once in the blood of all suramin-treated mice before sacrifice (65 days after treatment), although only in a fraction of mice at a given time.
Whereas all RAG1−/− mice inoculated with blood or brain homogenates from suramin-treated mice showed parasites early after infection, no parasites were detected in mice inoculated with either blood or brain homogenates from cordycepin- and deoxycoformycin-treated mice. Parasite DNA could not be amplified and parasite antigens were not detected in the brain tissues of cordycepin- and deoxycoformycin-treated mice. Infected and non-treated mice showed signs of morbidity before day 32 after infection. No mortality was recorded in the cordycepin-treated group and 1 out of 8 suramin-treated mice died 79 days after infection.
In a second experiment the minimal effective concentration of cordycepin was determined. For this purpose, we compared the treatment of mice given different doses of cordycepin and a constant dose of deoxycoformycin. Mice were inoculated with 2, 1.2 and 0.6 mg/kg/d cordycepin plus 0.2 mg/kg/d deoxycoformycin for 7 days starting at day 20 after infection. Whereas the trypanocidal effect was evident in all groups of mice (n = 9), a dose-dependent effect was observed, with doses lower than 2 mg/kg/d not sufficient to completely clear parasitemia (Figure 2A). One out of 9 mice inoculated with 2 mg/kg/d cordycepin showed parasites before 80 days after infection.
Next, groups of mice (n = 7) were dosed with 2 mg/kg/d cordycepin and 0.2 mg/kg/d deoxycoformycin during 3, 5 or 7 days starting 20 days after infection. Again, a dose-dependent trypanocidal effect was observed also in this experiment. Five or three doses were not enough to completely clear infection, whereas in the group treated with 7 doses only one animal remained infected (Figure 2B). Since, a total of 2 of out of 25 mice in the different experiments treated with 2 mg/kg/d cordycepin and 0.2 mg/kg/d deoxycoformycin for 7 days remained infected after treatment, we concluded that this treatment is on the brink of the minimal effective dose for clearance. We then increased the number of doses used for treatment. When mice (8 per group) were treated for 10 or 15 days with 2 mg/kg/d cordycepin and 0.2 mg/kg/d deoxycoformycin starting at day 20 after infection, a complete curative effect was achieved in two independent experiments (33 mice together) (Figure 2B and data not shown).
The minimal effective concentration of deoxycoformycin required for treatment was subsequently evaluated. For this purpose, mice (8 per group) were treated for 10 days with 2 mg/kg/d cordycepin and different concentrations of deoxycofomycin starting 20 days after infection with 2×103 T.b. brucei. Treatment with either 0.1 or 0.2 mg/kg/d deoxycoformycin resulted in complete cure of second stage infection, while no curative effect was observed when 0.02 mg/kg/d deoxycoformycin was used for treatment (Figure 2C). We also tested whether EHNA, another ADA inhibitor, could replace deoxycoformycin in treatment of late stage infections. Similar transient decreases in parasites levels in the blood were observed when 2 mg/kg/d cordycepin was inoculated in presence or absence of 0.2 mg/kg/d EHNA (data not shown), suggesting that EHNA did not protect cordycepin from degradation.
Starting treatment with cordycepin and deoxycoformycin at 20 days after infection resulted in loss of weight of infected mice. Surprisingly, toxicity was largely independent of the dose of cordycepin and deoxycoformycin used, since mice treated either for 15, 10, 7, 5 or 3 days with 2 mg/kg/d cordycepin and 0.2 mg/kg/d deoxycoformycin or treated with 7 doses of 2, 1.2 or 0.6 mg/kg/d cordycepin and 0.2 mg/kg/d deoxycoformycin all showed similar weight loss (Figure 2D, E and data not shown). All mice recovered weight after the treatment. No other sign of morbidity was noted. Importantly, weight loss was not observed in uninfected mice treated with 2 mg/kg/d cordycepin and 0.2 mg/kg/d deoxycoformycin (data not shown). Thus, the treatment toxicity required simultaneous presence of an established infection.
Real-time imaging of biophotonic emission provides a fast method to evaluate parasite dissemination in vivo. BALB/c mice were infected with 2×103 luciferase-tagged T.b. brucei AnTat1.1. Bioluminescent images showed a systemic spread of infection. No light emission was observed from mice treated with the doublet during late stage of infection (Figure 3), confirming the efficiency of treatment with cordycepin and deoxycoformycin.
The customary i.v. administration of trypanocidal drugs requires trained medical personnel, and specialized equipment and materials. Oral administration of these compounds would facilitate their use in patients. The effect of oral administration of cordycepin and deoxycoformycin in the outcome of murine infection with T. b. brucei was thus tested. We found that oral administration of 5 or 15 mg/kg/d cordycepin with 0.2 mg/kg/d deoxycoformycin cured approximately 50% of the treated mice when treatment was started at the day of infection and continued for 3 days. While oral administration of cordycepin has been shown to be an effective anti-tumoral agent [25], deoxycoformycin has a well described acid-lability that has hampered its use in oral formulations. Cordycepin and deoxycoformycin were then administered with omeprazole, a proton pump inhibitor, to reduce acid concentration in the stomach and thereby allow absorption of enhanced concentrations of the ADA inhibitor. Daily treatment of mice at the day of infection with cordycepin plus deoxycoformycin in presence of omeprazole for 5 days resulted in cure of the infection of all (n = 6) treated mice (Figure 4A). We then tested if oral administration of the doublet was effective for treatment of second stage infection with T. b. brucei. We found that oral administration of 15 mg/kg/d cordycepin and 0.4 mg/kg/d deoxycoformycin, but not 5 mg/kg/d cordycepin and 0.2 mg/kg/d deoxycoformycin, for 10 days starting 20 days after infection had a complete curative effect (Figure 4 B).
Subcutaneous treatment with 2 mg/kg/d and 0.2 mg/kg/d deoxycoformycin diluted in 100 µl of vegetable oil daily for 10 days starting on day 20 after infection with T.b. brucei was also effective in curing late stage infection of treated animals. Administration of 4 mg/kg/d cordycepin and 0.2 mg/kg/d deoxycoformycin s.c. every other day failed to cure all treated mice, suggesting that strategies improving pharmacokinetics of the doublet might improve treatment efficiency (Figure 4C).
We then studied whether late stage infection with human pathogenic strains of T. brucei can be treated with cordycepin and deoxycofomycin. Two weeks after infection with 104 T.b. rhodesiense STIB 851 isolate, parasites were detected in the brain parenchyma. Parasite density in the brain parenchyma increased at later times of infection (Figure 5A, B). Parasitemia levels were low or undetectable compared to T.b. brucei AnTat1.1 (Figure 5 C), but mice lost weight and showed signs of morbidity at 35–40 days after infection (data not shown). Ten days of dosing of 2 mg/kg/d cordycepin and 0.2 mg/kg/d deoxycoformycin treatment cured infection in mice as evaluated by real time PCR and inoculation of lysates of brains from mice 65 days after treatment to naïve immunodeficient mice (data not shown).
The effect of cordycepin and deoxycoformycin administration on the outcome of infection with T.b. gambiense was then evaluated. T.b.gambiense isolates were also susceptible to cordycepin in vitro (data not shown). Whereas mice inoculated i.p. with 2×105 T.b. gambiense (MBA stabilates) showed parasitemia up to at least 70 days after infection, mice inoculated 2×103 or 2×104 T.b. gambiense showed no parasites in the blood. In contrast to T.b. rhodesiense, morbidity was only observed in T.b. gambiense-infected mice at the time of sacrifice (80 days after infection) with small numbers of parasites observed in brain parenchyma, with elevated parasite densities observed only in the septum (data not shown). Treatment of T.b. gambiense infected mice with 2 mg/kg/d cordycepin and 0.2 mg/kg/d deoxycoformycin for 10 days, starting at 30 days after infection, when parasites where first detected in the brain parenchyma, completely eliminated parasitemia for at least 50 days after treatment (Figure 5 D).
Thus, treatment with cordycepin and deoxycoformycin was effective in treatment of late stage infection with both human pathogenic subspecies.
Brains from T.b. rhodesiense-infected mice treated with cordycepin and deoxycoformycin starting 20 days after infection contained diminished levels of IFN-γ, IL-1β, IL-6 and TNF-α mRNA compared to untreated infected controls when measured 10 days after treatment (Figure 6A–D). Likewise, brains from mice infected with T.b. rhodesiense and treated with cordycepin and deoxycoformycin starting 20 days after infection with T.b rhodesiense contained lower levels of pro-inflammatory cytokine transcripts when studied after 80 days after infection, as compared to non-treated infected groups. Cytokine mRNA levels in infected and cordycepin and deoxycoformycin-treated mice and in uninfected mice were similar. (Figure 6E–H).
The mechanisms accounting for the trypanocidal effect of cordycepin were subsequently explored. Several trypanocidal drugs have been shown to activate a programmed cell death of T. brucei [26]. We found that incubation with 1 µM cordycepin induced degradation of DNA by measured by propidium iodide (PI) staining of permeabilized parasites. DNA degradation was detected 1 h after cordycepin treatment and it increased with time (Figure 7A). DNA fragments were also detected by the TUNEL assay in cordycepin-treated parasites (Figure 7C). One of the earliest indications of apoptosis is the translocation of phosphatidylserine from the inner to the outer leaflet of the plasma membrane. Parasites treated with 1 µM cordycepin showed translocation of phosphatidylserine as indicated by the binding of annexin V as early as 1 h after treatment (Figure 7B). On the other hand, no alterations in mitochondrial redox potential of cordycepin-treated parasites were detected (Figure S1; Text S1). DNA degradation occurred in the absence of cell membrane disruption, which could only be detected 7 h after cordycepin treatment (Figure 7D), indicating that cordycepin induced programmed cell death of T.b. brucei which was followed by a secondary necrosis.
Whether T. b. brucei can develop resistance against cordycepin was then studied. For this purpose T. b. brucei AnTat1.1 and Lister 427 were incubated with low concentrations (1 LD50 or higher doses) of cordycepin for 72 h at 37°C. Parasites surviving the highest concentrations of cordycepin, were then further cultured for 3 days with similar or higher concentrations of cordycepin. This procedure was repeated for four months, after which a 40–60 fold increased resistance to cordycepin, compared to the parental strain was achieved (Figure 8A, F).
Resistance to the nucleoside analogue was retained after culture of the resistant parasites for three weeks in absence of cordycepin, suggesting that resistance was due to genetic alterations rather than to a metabolic adaptation of the parasite (Figure 8 B, G). Cordycepin-resistant parasites were then incubated with melarsoprol, pentamidine and suramine to determine cross-resistance. No cross-resistance with these trypanocidal drugs was observed, indicating that these trypanocidal compounds have non-overlapping molecular targets (Figure 8C–E, H–J).
Cordycepin-resistant strains displayed lower growth in vitro compared to their respective parental strains (Figure 9A, B). In vivo, mice inoculated with cordycepin-resistant parasites showed delayed and lower parasitemia levels compared to controls (Figure 9C, D). Parasitemia in mice inoculated with cordycepin-resistant AnTat1.1 became undetectable. Loss of weight was registered in mice infected with the parental strains but this effect was reduced in mice infected with cordycepin resistant strains (data not shown).
In order to determine possible cordycepin targets we amplified, cloned and sequenced 11 candidate genes involved in uptake and salvage of adenosine analogues in resistance-induced and parental parasites. The TbAT1, TbNT2, TbNT4, TbNT6, TbNT12 nucleoside transporter genes, TbHT1 hexose transporter, TbAK adenosine kinase, TbHGPRT hypoxanthine-guanine phosphoribosyltransferase from parental and resistant strains were screened for mutations that could cause an affected or deleterious protein function.
The HT1 was selected since this hexose transporter can be inhibited by tubericidin [24]. TbNT2 is predominantly expressed when compared to other genes in P1 family. The expression of TbNT3, TbNT4, TbNT5, TbNT7, is not seen in a parental T. b. brucei strtain. It has been reported that loss of primary transporter P2 leads to over expression of TbNT4 and TbNT6 [27]. TbNT2, 4, 6 and 12 were therefore selected for sequencing.
While several polymorphisms were found, no deleterious mutations were found in the genes sequenced except for TbAT1. TbAT1 is a nucleoside adenosine transporter present in the genome of T. brucei as a 1392 bp single copy gene where both alleles have been found to be almost identical. We found two mutations in the cordycepin-resistant AnTat1.1 strain TbAT1 gene that were not present in the respective parental DNA. There was a single base insertion at position 61 and a single base deletion at position 631, that could cause synthesis of a non-functional truncated gene product were found (Figure 10B). Five out of 12 sequenced colonies presented the one base insertion while the other seven carried the deletion, both groups with a read coverage of 15×, which confirmed the mutations. We assume that these two sequences detected correspond to the two alleles of TbAT1 carrying different mutations but rendering the same non-functional transporter protein.
For the cordycepin-resistant monomorphic Lister 427 strain, PCR using primers designed at the 5′ and 3′ ends of the gene and at the 5′ and 3′ UTRs were unable to amplify any product (Figure 10A), while the same set of primers amplified the TbAT1 gene from the parental DNA. We assume that in this case the mutations on this gene probably generated the deletion of a partial or the complete gene sequence, as amplification with internal primers although producing a PCR fragment, was not of the expected size (data not shown).
Since cordycepin resistant monomorphic (Lister 427) and pleomorphic (AnTat1.1) parasites displayed mutations in the TbAT1 gene, we tested whether this transporter affected the susceptibility of parasites to cordycepin by using TbAT1 knockout parasites. We found that resistance of cordycepin by knockout parasites was only slightly affected by the mutation, suggesting that the absence of the P2 transporter is not sufficient to explain the observed increased resistance of parasites to cordycepin (Figure 10C).
We here report a preclinical evaluation of the chemotherapeutic efficiency of cordycepin and deoxycoformycin for late stage infection with T. brucei. In brief, cordycepin was selected from a direct parasite viability screening of a compound library of nucleoside analogues. The minimal effective concentrations of cordycepin and deoxycoformycin for late stage treatment in vivo were determined. Of importance, the dose of deoxycoformycin used is similar or lower to that employed in leukemia patients, and cordycepin and deoxyformycin have been tested in clinical trials in humans (R McCaffrey, Harvard Medical School, Boston, personal communication) and in preclinical studies in dogs at higher concentrations to the ones here presented [28]. The compounds could be given i.p., s.c. and orally, and our data suggests that a slow release of the compounds in vivo will improve efficacy of treatment by reducing the doses, and that enteric coating might reduce the problem of acid lability of deoxycoformycin. The nucleoside analogue treatment was also effective for treatment of late stage infections with both T.b. rhodesiense and T.b. gambiense, different to some drugs in clinical use such as DFMO (eflornithine), which is only effective against T.b. gambiense. Cordycepin and deoxycoformycin treatment diminished the inflammatory cytokine accumulation in the brains of infected mice. In agreement, the number of CD45+ inflammatory cells in the brain parenchyma of T. brucei-infected mice was also reduced after treatment cordycepin and deoxycoformycin [15]. We found standard parameters associated to programmed cell death in T. brucei to be similar to those described in pluricellular organisms. Programmed cell death was followed by secondary necrosis. Interestingly, programmed cell death processes have previously been shown to be activated during the life cycle of trypanosomatids, regulating parasite number in certain host tissues, in the maintenance of clonality, as a mechanism of immunomodulation and of parasite differentiation [29]. Cordycepin thus stimulates a natural process of cell death, suggesting the presence of parasite cordycepin-sensitive targets regulating such processes.
The development of this new drug combination with reduced toxicity, high activity and oral availability gives rise to renewed hope of new and safer treatments of HAT. However, the utility and longevity of these new treatments will, to a large degree, be determined by the development of parasite resistance to the drug. We found that although T.b. brucei developed resistance to cordycepin upon prolonged culture with low doses of the compound, such resistant parasites showed diminished virulence and reduced growth in vivo. Moreover, resistant parasites showed no cross resistance to melarsoprol, suramin and pentamidine in clinical use for early or late stage treatment of HAT, suggesting that cordycepin targets are different to those of the other drugs. This is of importance since a strategy to delay or prevent the development of resistance is to combine drugs that act on different targets and therefore are unlikely to show cross-resistance. Additionally, synergistic effects may arise from a combination of two drugs that act differently. In such a situation, the concentrations of the combination partners can be reduced compared with those used in a mono-therapy, which leads to better safety.
Since resistance to cordycepin was slowly increased during the four months-long culture of the parasites with the drug (data not shown), it is likely that a number of mutations had accumulated during the process. After cloning and sequencing 11 candidate genes from cordycepin-induced resistant and parental strains, we found that parasites showed altered functional mutations only in the P2 transporter encoded by the TbAT1 gene. Although mutations found in the other examined genes were present, they did not show any predicted deleterious or altered effect for the respective protein function. Cordycepin uptake has been reported to be mediated by both the P1 and P2 transporters [27]. We found however that a TbAT1 (P2) null parasite showed only slight reduction in resistance to cordycepin, suggesting that other undefined mutated genes also participate in increased sensitivity to cordycepin of parental parasites.
Since cordycepin-resistant parasites show a genetic defect in TbAT1, it was surprising that no cross resistance to pentamidine, suramin and melarsoprol was not observed: the P2 transporter has been shown to be responsible for uptake of different trypanocides such as melaminyl arsenicals, pentamidine, diminazenes, cordycepin and tubercidin [30],[31],[32],[33]. Homozygously disrupted TbAT1 null T. brucei showed two- and three-fold increases resistance to melarsoprol and melarsen [17].
Most drugs for treatment of HAT were developed in the first half of the twentieth century, and some of them would not pass current high safety standards. A safe drug that is effective in the treatment of the encephalitic stage of HAT would significantly change the control and management of sleeping sickness. The data here presented strongly suggest that the combination of cordycepin and deoxycoformycin can be used in clinical tests for treatment of African trypanosomiasis. Besides the results from the preclinical experiments described in our manuscript, cordycepin fulfills a number of requirements for to be considered a candidate drug: It is available for other purposes and cheap, it can be given orally, clinical trials of the doublet have been performed and toxicity in man at the same concentrations used in animals seems not to pose a problem.
|
10.1371/journal.pntd.0002402 | Sex-Biased Expression of MicroRNAs in Schistosoma mansoni | Schistosomiasis is an important neglected tropical disease caused by digenean helminth parasites of the genus Schistosoma. Schistosomes are unusual in that they are dioecious and the adult worms live in the blood system. MicroRNAs play crucial roles during gene regulation and are likely to be important in sex differentiation in dioecious species. Here we characterize 112 microRNAs from adult Schistosoma mansoni individuals, including 84 novel microRNA families, and investigate the expression pattern in different sexes. By deep sequencing, we measured the relative expression levels of conserved and newly identified microRNAs between male and female samples. We observed that 13 microRNAs exhibited sex-biased expression, 10 of which are more abundant in females than in males. Sex chromosomes showed a paucity of female-biased genes, as predicted by theoretical evolutionary models. We propose that the recent emergence of separate sexes in Schistosoma had an effect on the chromosomal distribution and evolution of microRNAs, and that microRNAs are likely to participate in the sex differentiation/maintenance process.
| Schistosomiasis is the second most common disease caused by a parasite, affecting over 200 million people. The parasites involved are flatworms of the genus Schistosoma. Unlike most non-parasitic flatworms, Schistosoma species have separate sexes, and the emergence of sex has been associated with the development of a parasitic lifestyle. The identification of gene products that are expressed in a sex-biased fashion permits the study of the origin of sexual dimorphism and, in the case of the schistosomes, the evolution of a human parasite. Here we investigated the differential expression of microRNAs in male and female individuals of the species Schistosoma mansoni. MicroRNAs are crucial gene regulators. We observed that many new microRNAs emerged in the evolutionary lineage leading to the schistosomes. However, many sex-biased microRNAs were present in the hermaphrodite ancestor of the flatworms, and therefore acquired sex-biased expression later on. Our results suggest that changes in microRNA expression patterns were associated with the emergence of separate sexes in the schistosomes.
| Human schistosomiasis is a neglected tropical disease (NTD) caused by blood flukes of the genus Schistosoma. Schistosomiasis is estimated to affect over 200 million people in developing tropical and subtropical countries, with over 90% of cases being confined to Africa [1], [2]. Schistosoma mansoni is primarily responsible for intestinal and hepatic schistosomiasis in Africa, the Arabian peninsula, parts of South America and the Caribbean Islands [3]. Unlike most other flatworms (phylum Platyhelminthes), Schistosoma species are dioecious; that is, they have two differentiated sexes. The emergence of sexual dimorphism in these species is believed to be associated with adaptation to warm-blooded vertebrates from a hermaphrodite ancestor in cold-blooded vertebrates [4]. Schistosoma mansoni has seven pairs of autosomes and one pair of sexual chromosomes with a ZW system, i.e., females are the heterogametic sex [5]. Like other species with a ZW-based system of sex determination, there is no apparent global dosage compensation in females [6].
The origin of Schistosoma sexuality has attracted much attention [4], [7]–[9]. Moreover, as the eggs laid by the female worms are primarily responsible for the pathology associated with schistosomiasis, the mechanisms associated with pairing and egg-laying, including expression of sex specific genes are of great interest. In the last years, different groups have characterized, using genomic and proteomic approaches, gene products with differential expression between males and females in Schistosoma species [10]–[13]. Since both sexes are necessary for the colonization of the host, sex-biased genes are potential targets for the infection control of schistosomes.
MicroRNAs are short endogenous RNA molecules that regulate gene expression by targeting mature mRNA transcripts [14]. This mechanism of post-transcriptional regulation is conserved in animals and is likely to be involved in all aspects of cellular function [15]. The microRNA content of Schistosoma japonicum, which affects large endemic areas around the river Yangtze in China, has been studied in detail by several groups [16]–[21]. In addition, the previous characterization of microRNAs from other non dioecious species of flatworms, such as Echinococcus granulosus and the non-parasitic Schmidtea mediterranea [22], [23], provides a background against which to identify Schistosoma-specific microRNAs with a potential role in sexual development and host-parasite interactions.
Current knowledge of S. mansoni microRNAs is limited and mostly based on computational predictions [24], [25]. Moreover, S. mansoni provides an excellent model to study the evolution and function of sex-biased microRNAs. Here we use deep sequencing of RNA libraries to explore the microRNA content of S. mansoni, identify microRNAs specific to the schistosomes, and study the potential impact of sex-biased microRNAs in sexual differentiation.
We have used small RNA deep sequencing to identify a total of 112 microRNAs in adult S. mansoni (Table 1, Supplementary Files S1 and S2). Valid microRNA candidates were required to have reads mapping to both arms of the precursor sequences (representing mature microRNA and microRNA* sequences) except for those with a previously validated homolog (see Materials and Methods). Our microRNA annotation procedure was intentionally conservative: we may not have detected some bona fide microRNAs, but our predictions are of high confidence.
De Souza Gomes et al. [25] computationally identified 42 microRNA loci in S. mansoni, significantly expanding the previous set of 6 microRNAs [16], [24]. We confirmed 20 of these (Table 1), all conserved in other species. We failed to detect the remaining 23. All but two of the unconfirmed microRNAs were not conserved in other flatworms. A second recent work characterized 211 novel microRNAs in S. mansoni by cloning of small RNA sequences from adults and schistosomulas [26]. However, the majority of the candidate microRNAs map to many positions in the genome and only a few are reported to be within putative precursor hairpin structures. Indeed, only two of the reported 211 microRNAs were confirmed in our analyses (mir-71a and let-7). We identify 92 microRNAs not previously annotated in S. mansoni, eight with obvious homologs in other species (Table 1). Amongst these, we show that the deeply-conserved mir-124 locus produces microRNAs from both genomic strands in S. mansoni (Supplementary Files S2). The remaining 84 novel predictions had no detectable similarity with any known microRNA.
To characterize the microRNAs conserved in the parasitic Schistosoma genus, we compared our sequenced microRNAs with those already described for S. japonicum [17]–[19], [21]. We found that 26 out of our 112 microRNAs were conserved between these two species (Figure 1). In order to determine how many of those 26 are specific to the Schistosoma lineage, we searched the genomes of the flatworms Schmidtea mediterranea [22] and Dugesia japonica [27] for homologous sequences. Strikingly, all known microRNAs conserved between Schistosoma species were also conserved in other flatworms. We also detected a set of S. japonicum homologs for 12 of our newly identified S. mansoni sequences that were not conserved in other platyhelminthes (Figures 1A and B). Hence, these 12 microRNAs are the first instances of schistosome-specific microRNAs.
Homology searches in other animals showed that three microRNAs are specific to platyhelminthes: mir-755, mir-2162 and mir-8451 (Figure 1A). Thirteen microRNAs are protostome-specific and the remaining 10 are conserved across the animal kingdom (Figure 1A). A total of 71 microRNAs identified in this study are not detected in any other species, and are therefore likely to be S. mansoni specific.
To explore potential sex-biased expression, we compared the relative expression levels of the 112 microRNAs in males and females (Figure 2A). We found 13 microRNAs that are differentially expressed between males and females. A significant excess [10] show increased expression in females (Figure 2A, Table 2). We further quantified the relative expression level of all 3 male-biased and 4 of the female-biased microRNAs by real-time PCR (see Materials and Methods). The observed fold changes in our qPCR experiments were consistent with those observed in our RNAseq analysis (Table 3), although the two least biased microRNAs by RNAseq show small and non-significant changes by qPCR.
We next evaluated whether microRNA loci on the sex chromosomes are biased towards differential expression between sexes. The current assembly of the S. mansoni genome does not differentiate between Z and W chromosomes. At this stage, we cannot therefore evaluate the two sexual chromosomes separately. In Figure 2B, we plot the relative enrichment of sex-chromosome-linked microRNAs for male and female-biased expression. We observed that the sex chromosome has fewer female-biased microRNAs (∼3-fold change) than expected by chance, although the difference is only marginally significant (p = 0.10). The data therefore indicate that microRNA genes with female sex-biased expression may have a tendency to move out of sex chromosomes (see Discussion).
The mir-71/mir-2 microRNA cluster is highly conserved in invertebrates, and it has been shown that this cluster is duplicated in Platyhelminthes [25], [28]. Interestingly, one of the clusters (mir-71/2a/2b/2e) is on the sex chromosomes while the other (mir-71b/2f/2d/2c) is on chromosome 5 [25], [29]. This has led some authors to postulate that the mir-71/mir-2 clusters may be involved in sexual maturation in Schistosoma [25]. Our analysis reveals that all microRNAs in the autosomal cluster have female-biased expression, while the sex chromosome cluster does not show any bias (Figure 2A). Interestingly, the two clusters emerged by a duplication in the ancestral lineage leading to Schistosoma, and the multiple copies in other Platyhelminthes [29], [30] came from independent duplication events (Supplementary Files S4). This example may shed some light on how sex chromosomes evolved in dioecious species (see Discussion).
Although the computational prediction of microRNAs has been useful to understand the biology of small RNAs in S. mansoni [25], sequencing is required to validate the existence of these microRNAs as well as for detecting new sequences. Our work has confirmed the existence of 20 of the microRNAs predicted by de Souza Gomes et al. [25] and we have expanded the S. mansoni microRNA set to 112 loci. We specifically detect microRNAs expressed in sexually mature adults, and microRNAs specifically expressed in other developmental stages (such as schistosomulas) may have escaped our analysis. The use of deep sequencing also permits the characterization of microRNAs produced from both strands of the same locus, and we have identified sense and antisense microRNA production from the mir-124 locus. However, there is no evidence that this microRNA is also bidirectionally transcribed in other species. Indeed, bidirectionally transcribed microRNAs are rare and poorly conserved; only two cases of conserved bidirectional microRNAs are known in protostomes: iab-4 and mir-307 [31].
We observe an excess of microRNAs that exhibit female-biased expression (Figure 2A). This is in agreement with the overall female-bias observed for protein-coding genes in both S. mansoni [10] and S. japonicum [11], although a recent expression analysis in S. japonicum showed no gender bias [13]. Recently, a work in the parasitic nematode Ascaris suum showed that microRNAs are differentially expressed between males and females [32]. Although the differences were small and the targeting properties of male and female microRNAs similar, they reported a tendency of male microRNA to target extracellular proteins [32]. Another work in the bird Taeniopygia guttata (zebra finch), suggests that the male-biased expressed microRNA mir-2954 specifically target genes in the Z chromosome, and may be involved in sexual dimorphism in song behavior [33]. Together, these papers and our work point to a general mechanism of microRNA-modulation of sex-specific gene expression.
A recent work shows that some microRNAs are specifically expressed in Schistosoma japonicum eggs [21]. In that work, the authors also measured the microRNA expression levels in males and females. We reanalyse their data and find that two out of our three male-biased microRNAs (mir-1 and mir-61) have a consistent bias in Schistosoma japonicum, while the third is not present in their dataset. Also, five of our female-biased microRNAs also showed a female bias in their work (mir-71b, mir-2c, mir-2d, bantam and mir-31). These findings further validate our results and show that the sex-biased pattern of microRNA expression is evolutionarily conserved between these two species.
Sex-biased gene expression affects the genetic composition of chromosomes, since selection has different effects on sex-biased genes depending on whether they are located on sex chromosomes or on autosomes (reviewed in [34]). Likewise, sex chromosomes have distinctive evolutionary patterns, which affect the genes encoded within [35]. We may therefore expect to see signatures of chromosome evolution in sex-biased microRNAs. Indeed, we observed that microRNAs that are female sex-biased are depleted in the sexual chromosomes (Figure 2B). This may indicate a loss of sex-biased genes at the sex chromosomes. One interesting example is the female-biased mir-71b/2f/2c/2d autosomal microRNA cluster, which has a paralogous copy in the sex chromosome with no biased expression. If a microRNA gene that is selectively advantageous for females becomes part of a sex chromosome (by sexualisation of the chromosome, or otherwise), selection over this gene will be less efficient in the heterogametic sex (females in our case), generating a conflict between expression pattern and chromosomal location. Duplication of a gene into an autosome has been recently proposed as a way to escape such conflict (reviewed in [36]), and the mir-71/mir-2 cluster duplication appears to be an example of this. Although a more comprehensive analysis of duplicated microRNAs with sex-biased expression is required to confirm this, our analysis shows that sex-biased expressed microRNAs have an impact in shaping the genome during evolution.
The study of microRNAs in the schistosomes is of both evolutionary and biomedical interest. First, the recent evolution of a sexual reproductive system from a hermaphrodite ancestor can give clues about how sexuality emerged in other species. Our data are consistent with the acquisition of sex-biased expression of conserved microRNAs soon after the species become dioecious. Second, the characterization of Schistosoma-specific microRNAs may provide new targets for infection control. In this work, we characterize for the first time 12 microRNAs conserved between S. mansoni and S. japonicum but not in other platyhelminthes (nor in other animals). Two of these sequences also showed female-biased expression (mir-8437 and mir-8447). However, some important questions remain to be answered: Are microRNAs conserved in the two studied schistosomes also conserved in other Schistosoma species or in other trematodes? Is sex-biased expression of microRNAs associated with sex-biased expression of their targets? Do other dioecious flatworms have sex-biased expression of microRNAs? The genomic sequencing of more platyhelminthes and characterization of their microRNAs will help us to answer these questions.
For the detection of expressed microRNAs, female mice (BKW strain) were infected with Belo-Horizonte strain Schistosoma mansoni parasites by paddling in water containing 200 cercariae. Seven weeks after infection, adult schistosomes were collected from the mice post-mortem by hepatic portal perfusion. RNA was extracted from adult schistosome samples with the miRVana miRNA isolation kit (Ambion). We used two sequencing technologies, AB SOLiD and Illumina MiSeq, to sequence S. mansoni small RNA libraries. The extremely deep coverage provided by SOLiD sequencing provides high sensitivity for the discovery of novel microRNAs. We further used Illumina MiSeq sequencing of gender-specific libraries to compare the expression level of microRNAs between males and females. Library construction was performed as previous described [37] using the SOLiD Small RNA Expression Kit (Ambion). SOLiD sequencing was performed at the Center for Genomic Research at the University of Liverpool. We obtained a total of 124,341,126 SOLiD sequence reads from two libraries. For the gender-specific differential expression of microRNAs, we prepared RNA libraries with the miRVana kit from separate male and female samples (provided by Andrew MacDonald and Rinku Rajan at the University of Edinburgh), and prepared libraries for Illumina MiSeq sequencing according to the manufacturer's instructions. MiSeq samples were sequenced in the Genomics Core Facility at the University of Manchester. High-throughput datasets were deposited in Gene Expression Omnibus (GEO) at NCBI (accession number: GSE49359).
Sequencing reads from male and female libraries were separately mapped to the S. mansoni reference genome (assembly 5.1 available at http://www.genedb.org/Homepage/Smansoni; [38], [39]) with Bowtie 0.12 using the sequential trimming strategy implemented in SeqTrimMap 1.0 [40] allowing 2 mismatches. Sequences mapping to potential rRNAs or tRNAs were first removed. Putative tRNAs were predicted in the genome sequence with tRNAscan-SE using default parameters [41] and ribosomal RNAs (rRNAs) were extracted from the SILVA database (http://www.arb-silva.de/, release 108). Mapped reads with a length of 19–25 nucleotides, matching five or fewer positions in the genome (a total of 63,771,124 sequence reads), were used to detect microRNAs as previously described [37], [40]. We further used BLAST [42] to search microRNA candidates against the S. mansoni genome and discarded those with more than 5 hits (E-value <e-10, 80% query coverage) to remove potential repetitive elements: 44 candidate sequences did not pass this filter. MicroRNA candidates were manually inspected. Potential homologs of known microRNAs (detected by BLASTN against all hairpin sequences from miRBase version 17 [43]) with reads mapped from our datasets but which did not pass our criteria were also retained. Homolog of our microRNA candidates were predicted in the genome sequences of S. japonicum (version 2), S. mediterranea (v. 3.1), Caenorhabditis elegans (v. 7.1), Tribolium castaneum (v. 3.0), Drosophila melanogaster (v. 5.1), Homo sapiens (v. 37.1) and Gallus gallus (v. 2.1), with parameters: −W 4, −r +2, −q −3. Only sequences with a predicted hairpin structure and conserving at least one mature sequence were considered as putative homologs.
We mapped reads produced from MiSeq sequencing reactions to our annotated S. mansoni microRNAs and discarded all reads that map to more than one microRNA locus. Read counts were transformed with the ‘upper quartile’ normalization using the edgeR package [44], following the suggestions in [45]. Other normalization procedures (‘TMM’, ‘LOWESS’ and ‘no normalization’) did not change the results (Supplementary Files S3). Fold changes in expression levels are given in logarithms in base 2. We consider a dispersion of expression between experiments of 0.2 and a false discovery rate of 10%. With these parameters, microRNAs showing a two-fold difference in their expression levels are considered to be sex-biased, as routinely suggested [46], [47]. The expression level in males and females of seven microRNAs (mir-1b, mir-61, mir-281, mir-36b, mir-71b, bantam and mir-8437) were further validated by quantitative PCR. We used custom made TaqMan assays manufactured by Life Technologies. Fluorescent quantification was done in a Chromo 4 qPCR system (BioRad) for a log fluorescent threshold of 0.05, using mir-36a (the microRNA showing the least bias in the MiSeq experiments) as the non-sex-biased control microRNA. For each amplification, we performed three technical replicates to estimate the significance of the observed differences.
Laboratory animal use was within a designated facility regulated under the terms of the UK Animals (Scientific Procedures) Act, 1986, complying with all requirements therein. The experiments involving mice in this study were approved by the Natural History Museum Ethical Review Process and work was carried out under Home Office project licence 70/6834.
|
10.1371/journal.ppat.1006413 | Genomic fossils reveal adaptation of non-autonomous pararetroviruses driven by concerted evolution of noncoding regulatory sequences | The interplay of different virus species in a host cell after infection can affect the adaptation of each virus. Endogenous viral elements, such as endogenous pararetroviruses (PRVs), have arisen from vertical inheritance of viral sequences integrated into host germline genomes. As viral genomic fossils, these sequences can thus serve as valuable paleogenomic data to study the long-term evolutionary dynamics of virus–virus interactions, but they have rarely been applied for this purpose. All extant PRVs have been considered autonomous species in their parasitic life cycle in host cells. Here, we provide evidence for multiple non-autonomous PRV species with structural defects in viral activity that have frequently infected ancient grass hosts and adapted through interplay between viruses. Our paleogenomic analyses using endogenous PRVs in grass genomes revealed that these non-autonomous PRV species have participated in interplay with autonomous PRVs in a possible commensal partnership, or, alternatively, with one another in a possible mutualistic partnership. These partnerships, which have been established by the sharing of noncoding regulatory sequences (NRSs) in intergenic regions between two partner viruses, have been further maintained and altered by the sequence homogenization of NRSs between partners. Strikingly, we found that frequent region-specific recombination, rather than mutation selection, is the main causative mechanism of NRS homogenization. Our results, obtained from ancient DNA records of viruses, suggest that adaptation of PRVs has occurred by concerted evolution of NRSs between different virus species in the same host. Our findings further imply that evaluation of within-host NRS interactions within and between populations of viral pathogens may be important.
| This paper addresses the adaptive strategies of ancient defective viruses recorded in grass genomes. We mined numerous virus segments from various grass genomes and assembled several defective pararetrovirus (non-autonomous PRV) species. We attempted to understand how these non-autonomous PRVs can complete parasitic life cycles in host plants. We determined that these non-autonomous PRV species have participated in interplay with autonomous PRVs or different non-autonomous PRV species. This interplay between different virus genomes has involved the exchange of noncoding regulatory sequences, which consequently evolved to be extraordinarily highly similar in different viruses within the same host. In non-autonomous PRVs, adaptive strategies to compensate for a lack of functionality have consequently involved concerted evolution of noncoding sequences establishing the partnerships.
| Similar to virus–host interactions, virus–virus interactions, especially those occurring during mixed plant virus infections in nature, have complex outcomes ranging from antagonism to synergism [1, 2]. Such interactions between different virus species affect their adaptation [1, 2]. Numerous virus-derived sequences, referred to as endogenous viral elements (EVEs), have recently been discovered in various eukaryotic genomes [3–6]. In addition to EVEs derived from retroviruses, EVEs originating from viruses without active reverse-transcription or integration abilities have been identified [4, 7–10]. Because these elements are vertically inherited viral sequences integrated into the germline genome of a host, they are viral genomic fossils and hence serve as invaluable historical records [3, 11, 12]. Although EVEs may provide an unprecedented opportunity to advance our understanding of evolutionary-scale virus–virus interactions, these records have rarely been exploited to explore such interactions.
Pararetroviruses (PRVs), including Caulimoviridae and Hepadnaviridae families, are reverse-transcribing double-stranded DNA viruses that lack an integrase and a process for integration [5, 13]. PRVs also possess EVEs called endogenous PRVs that originated from the incidental integration of PRV DNA into host genomes through non-homologous end-joining [14, 15]. Endogenous PRVs have been identified in an increasing number of plant genomes and have also been recently discovered in bird and reptile genomes [4, 5, 11, 16–18].
PRVs are thought to be distantly related to long terminal repeat (LTR) retrotransposons [19]. Interestingly, many LTR retrotransposons are non-autonomous with respect to their parasitic life cycle in host cells, i.e., they have lost most or all of their coding capability but can amplify themselves by using the protein machinery of autonomous LTR retrotransposons that are functionally and structurally intact [20–22]. A hallmark of the parasitism of non-autonomous LTR retrotransposons on their autonomous partners is the substantial sequence similarity of their LTRs—the location of noncoding regulatory sequences (NRSs) [22–24]. Plant PRVs have open circular genomes and encode a movement protein (MP), a capsid protein (CP) harboring a zinc finger motif, a protease (PR), and a reverse transcriptase with RNase H activity (RT/RH) [25]. In addition to the domains encoding these essential proteins, diverse non-standard domains or open reading frames (ORFs) have frequently been found in plant PRV genomes, the protein products of which generally play roles in vector transmission or immune suppression [26, 27]. The intergenic region (IGR) of plant PRVs, a highly diverse noncoding region containing multiple NRSs, is crucial for viral transcription, translation, and replication [25, 27]. All known PRVs encode all essential proteins and are thus autonomous PRV species during their parasitic life cycle in host cells. Limited cases of non-autonomous virus species have been previously documented. One well-known example is adeno-associated virus (Dependoparvovirus, a single-stranded DNA virus), which has been applied as a gene therapy vector [28]. No non-autonomous PRV species have been reported from nature to date.
In this study, we uncovered paleogenomic evidence for non-autonomous PRVs and revealed their interplay with different PRV species through an analysis of endogenous PRVs in grass family (Poaceae) genomes (S1 Table). We discovered two examples of virus–virus interactions: a possible commensal partnership between a non-autonomous PRV and an autonomous PRV species, and a possible mutualistic partnership between two functionally complementary non-autonomous PRV species. Unexpectedly, we found that the two partners in each interplaying system have frequently exchanged (>18 estimated major recombination events) their NRSs with each other via region-specific recombination to maintain partnership and coevolution. The NRS homogenization between partner viruses led by such recombination events suggests that concerted evolution has occurred in these proposed partnerships. Our results provide paleoviral insights into the genesis and adaptation of complex virus systems.
We previously identified the first known endogenous PRV family in the genome of rice (Oryza sativa) [29]. This family, derived from a sister species of rice tungro bacilliform virus (RTBV)—an autonomous PRV that infects O. sativa—has been designated as endogenous RTBV-like (eRTBVL) [14, 29, 30]. In the present study, we observed domain reshuffling in at least 13 eRTBVL segments in the O. sativa genome, 7 of which formed a long cluster on chromosome 8 with segments of eRTBVL-X (the youngest group of eRTBVL [30]) (S1 Fig). These reshuffled sequences exhibited a consensus pattern among the 13 segments (S2 Fig), which suggests that the domain reshuffling must have occurred in the corresponding viral genome prior to integration. We named this reshuffled eRTBVL as endogenous RTBV-like 2 (eRTBVL2) and reconstructed its ancestral virus circular genome (Fig 1A). Instead of an RT/RH domain and a third ORF, this eRTBVL2 possessed a functionally unknown domain, henceforth referred to as the SFKTE domain (for the conserved five-residue SFKTE present in all homologous sequences) (Fig 1A and 1B). A BLAST search for the SFKTE domain sequence in the O. sativa genome identified 15 loci (e-value < 4.00 × 10−44) that have recently been annotated as endogenous PRVs similar to petunia vein clearing virus (PVCV) sequences; these PRVs are hereafter referred to as endogenous PVCV-like (ePVCVL) (Fig 2A; [18]). By aligning the regions around the identified sequences, we constructed the ancestral virus circular genome for these ePVCVL segments (Fig 1A; details in S3 Fig). The results of a detailed sequence comparison using consensus sequences of viral genomes imply a possible recombination event between the viruses of eRTBVL and ePVCVL that may have generated a recombinant virus responsible for eRTBVL2 (Fig 1B). Recombination analyses with multiple methods statistically validated this recombination event (P = 7.18 × 10−309; S2 Fig). Examination of presumed recombination breakpoints revealed no obvious sequence similarity between the parent sequences; instead, we detected a small microhomologous region at the left breakpoint (S2 Fig), which suggests an illegitimate recombination event. Three predicted essential domains (MP, CP, and PR) were confirmed by conserved motif alignment, but the RT/RH domain indispensable for replication was not detected in eRTBVL2 or ePVCVL (S2 Table and S4 Fig). Despite the absence of the RT/RH domain, the presence of multiple genomic fossils of these viruses (13 eRTBVL2 and 24 ePVCVL segments in the O. sativa genome; S2 Fig and S3 Table) suggests the success of their proliferation. We therefore propose that the viruses of eRTBVL2 and ePVCVL are non-autonomous PRV species.
To achieve replication, non-autonomous PRVs of eRTBVL2 and ePVCVL should require an autonomous partner virus or other related elements. Considering the high sequence similarity of IGRs carrying NRSs (Fig 1B; predicted NRSs in S5 Fig), we hypothesized that the virus of eRTBVL2 may depend on the protein machinery of the virus of eRTBVL (an autonomous PRV) for proliferation, similar to the case of parasitic interactions between non-autonomous and autonomous LTR retrotransposon pairs [20–22]. We thus tested the spatio-temporal likelihood of this proposed interplay. In a phylogenetic tree of IGR sequences of eRTBVL and eRTBVL2 (Fig 1C), most eRTBVL2 sequences were placed within or close to the eRTBVL-X clade, with three other eRTBVL2 sequences each falling into one of three older eRTBVL clades (-A1, -A2 and -B) [30]. Phylogenetic trees of other homologous regions (ORF1, MP, CP, and PR domains) between eRTBVL and eRTBVL2 had topologies similar to the IGR-based tree (see S6 Fig for these four ORF/domains). The results of these phylogenetic analyses suggest that recombination may have occurred between the viruses of eRTBVL and eRTBVL2 at IGRs and other homologous regions, implying their spatio-temporal coexistence. Detailed recombination analyses confirmed the contribution of the virus of eRTBVL to the recombination of the viruses of the three eRTBVL2 sequences phylogenetically close to eRTBVL-A1, -A2, and -B clades, and also supported recombination events between the viruses of eRTBVL-X and other eRTBVL2 sequences (P = 1.37 × 10−9 to 1.44 × 10−181; S7A Fig). We next analyzed the temporal relationship of eRTBVL2 segments based on a phylogeny of the SFKTE domain (S8 Fig). We rooted the phylogenetic tree of SFKTE amino acid sequences of eRTBVL2 and ePVCVL (S8 Fig) using the oldest ePVCVL segment, where the relative antiquity of the latter was determined by a bidirectional genome-wide orthology analysis of ePVCVL loci in Oryza species (see Materials and Methods and S4 and S5 Tables; PCR and Sanger sequencing validation in S9 Fig). In the generated SFKTE domain tree (S8 Fig), the eRTBVL2 segments related to the eRTBVL-X group (Fig 1C) were the latest branching sequences, whereas the three eRTBVL2 segments related to eRTBVL-A1, -A2, and -B groups (Fig 1C) branched earlier (S8 Fig). Because the eRTBVL-X group is the youngest eRTBVL group and eRTBVL-A1, -A2, and -B groups are older [30], the SFKTE phylogeny indicates that the evolution of the virus of eRTBVL2 is temporally consistent with that of eRTBVL. Taken together, these results strongly support the coexistence and coevolution of the viruses of eRTBVL2 and eRTBVL and provide evidence for a possible partnership between the two viruses during mixed infection. The virus of eRTBVL2 did not seem to be a parasite on the virus of eRTBVL, because we observed no higher magnitude of proliferation in the former relative to the latter (Fig 1C and S6 Fig). Taking into account the observation that the replication dependence of the virus of eRTBVL2 on the virus of eRTBVL had no recognizable deleterious effect on the latter, we suggest a possible commensal partnership between the viruses of eRTBVL2 and eRTBVL.
Although our search for the autonomous partner of the virus of ePVCVL revealed no such candidate in the genomes of O. sativa or other Oryza species, we noticed another endogenous PVCV-like family (hereafter ePVCVL2) showing defective structures (Fig 2B and 2C; [18]). We successfully reconstructed the ancestral virus circular genome of ePVCVL2; this ancestral genome possessed MP, PR, and RT/RH domains but the CP domain was absent (Fig 3A; details in S2 Table, S3 and S4 Figs). The composition of this genome suggests that the virus of ePVCVL2 is structurally and functionally complementary to the virus of ePVCVL. Given the existence of the naturally defective genome as well as multiple fossils of the virus of ePVCVL2 (11 segments in the O. sativa genome; S3 Table), we suggest that this virus is another non-autonomous PRV species. Detailed comparison of ePVCVL and ePVCVL2 consensus sequences revealed a high degree of local similarity between their IGRs as well as their MP domains (97.2% nucleotide identity: 99.3% for IGR and 95.3% for MP) (Fig 3B). Given that IGR sequence identities between eRTBVL groups (intraspecies level) ranged from 72.6% to 92.8%, this interspecies similarity of IGRs is exceptionally high. Both ePVCVL and ePVCVL2 encode a PR domain, but the nucleotide sequence of this region was very dissimilar between these two types of endogenous PRVs (Fig 3B). This dissimilarity of PR domains, extraordinarily high IGR sequence similarity (identical NRSs between IGRs; predicted NRSs in S5 Fig), and observed functional complementarity between the viruses of ePVCVL and ePVCVL2 all suggest a possible mutualistic partnership in which the two viruses mutually compensate to facilitate proliferation.
To confirm the proposed partnership, we performed a bidirectional genome-wide orthology analysis of ePVCVL2 loci in Oryza genomes (the same analysis of ePVCVL loci mentioned above). This analysis revealed that ePVCVL and ePVCVL2 segments are species-specific, except for four shared ePVCVL loci and two shared ePVCVL2 loci, and coexist in each analyzed Oryza genome (Fig 2C; details in S9 Fig, S4 and S5 Tables), thereby supporting the coexistence of the viruses of ePVCVL and ePVCVL2 during host divergence. No major ePVCVL cluster related to a major ePVCVL2 cluster was present in the phylogenetic tree of ePVCVL and ePVCVL2 IGR sequences in the O. sativa genome (Fig 3C). On the contrary, three ePVCVL IGR sequences clustered with three ePVCVL2 IGR sequences in a strongly supported clade (Fig 3C). To confirm this finding, we examined single nucleotide polymorphisms (SNPs) among the six IGR sequences, which revealed six SNP sites shared by the IGRs of ePVCVL and ePVCVL2 (Fig 3D). We further carried out recombination analyses on these ePVCVL and ePVCVL2 sequences, which resulted in the identification of significant recombination events between the IGRs of the viruses of ePVCVL and ePVCVL2 (P = 1.28 × 10−8 to 2.90 × 10−23; S7B Fig). When we extended our phylogenetic analysis of IGR sequences to segments in other Oryza genomes, we also found that the IGR sequences of ePVCVL and ePVCVL2 clustered together (S10 Fig). The recombination of IGR sequences between the viruses of ePVCVL and ePVCVL2, implied by the phylogenetic analysis, was likewise confirmed by recombination analyses of ePVCVL and ePVCVL2 sequences in these Oryza genomes (P = 4.06 × 10−4 to 6.87 × 10−23; S7C Fig). Taken together, these data thus provide strong evidence that two non-autonomous PRVs in a possible mutualistic partnership have recombined their IGR sequences to continue their coevolution during mixed infection.
By searching for homologous sequences of eRTBVL2, ePVCVL, and ePVCVL2 and reexamining reported endogenous PRVs in non-Oryza grass genomes [18], we found both ePVCVL and ePVCVL2 homologous sequences coexisting in the genomes of sorghum (Sorghum bicolor) and switchgrass (Panicum virgatum) (Fig 2 and S3 Table). These sequences formed a phylogenetic sister group (non-Oryza group) to either ePVCVL or ePVCVL2 segments of analyzed Oryza genomes (Oryza group) (Fig 2A and 2B). We constructed two ancestral virus circular genomes for these sequences (Fig 4A; details in S3 Fig). One was structurally equivalent to Oryza group ePVCVLs, that is, the RT/RH domain was absent. The other genome resembled Oryza group ePVCVL2s, but lacked both MP and CP domains (this genome contained a region slightly resembling the CP domain but without an essential zinc finger motif) (Fig 4A and 4B; details in S2 Table, S3 and S4 Figs).
IGR sequences of ePVCVL and ePVCVL2 non-Oryza groups shared extremely high nucleotide identities (97.1%; Fig 4B), whereas IGR sequence similarities between ePVCVL Oryza and non-Oryza groups and between ePVCVL2 Oryza and non-Oryza groups were low (43.6% and 44.6% nucleotide identities, respectively; S11 Fig). In a phylogenetic tree based on sequences from the S. bicolor genome, IGR sequences of non-Oryza ePVCVL and ePVCVL2 groups were mixed together (Fig 4C). (The number of IGR sequences in the P. virgatum genome was too limited for phylogenetic analysis). We also performed recombination analyses on these sequences in the S. bicolor genome, which resulted in the detection of significant recombination events occurring between the IGRs of the viruses of ePVCVL and ePVCVL2 non-Oryza sequences (P = 6.30 × 10−10 to 1.23 × 10−22; S7D Fig). A close examination of the S. bicolor sequences revealed that virus-derived small insertion/deletion (indel) variations in IGRs were shared between partial non-Oryza ePVCVL and ePVCVL2 segments (Fig 4D). The presence of these indels is direct evidence that IGRs have frequently been recombined between the virus genomes of ePVCVL and ePVCVL2. Taking all of these results into consideration, we conclude that non-autonomous PRVs have adapted to a long-term partnership via IGR homogenization mediated by frequent recombination, leading to concerted evolution of NRSs.
The discovery and analysis of various EVEs in eukaryotic genomes has contributed to our understanding of viral origin and evolution as well as long-term interactions between viruses and hosts [3, 31–33]. Endogenous PRVs in plant genomes have been frequently reported [5, 18, 34], and extreme cases of endogenous PRV reactivation under certain conditions, such as in endogenous banana streak virus, have been well documented [35–38]. Using grass endogenous PRVs as ancient DNA records of viruses, we performed paleogenomic analyses of PRVs to explore their long-term virus–virus interactions. In contrast to all previously known PRVs, which are autonomous, three non-autonomous PRV species were identified in this study, namely, the viruses of eRTBVL2, ePVCVL, and ePVCVL2. Our examination of ePVCVL and ePVCVL2 sequences, which were first described by Geering et al. [18], revealed the adaptation strategies of their corresponding non-autonomous viruses. We have proposed two adaptation strategies used by non-autonomous PRVs: a possible commensal partnership with autonomous PRVs and a possible mutualistic partnership with other non-autonomous PRVs (summarized in Fig 5A). These proposed partnerships have been enabled by the existence of shared common NRSs in their IGRs. We have also demonstrated the evolutionary dynamics of these partnerships: frequent recombination of IGRs (>18 estimated major events; see below) between two partners leading to NRS homogenization between different PRV species during host divergence. This concerted evolution of NRSs is responsible for the maintenance of such partnerships and has driven the coevolution of interacting viruses.
The consensus NRSs of two partner viruses would be expected to recruit the same virus-encoded proteins and host factors to complete their life cycles in hosts. In the possible commensal partnership suggested by this study (Fig 5A), the non-autonomous virus of eRTBVL2 should benefit from sharing the RT/RH protein of the autonomous virus of eRTBVL. With respect to the SFKTE domain of the virus of eRTBVL2, neither the RT-like motif nor its degenerate residues could be distinguished in this domain by amino acid alignment with all known types of RT-like domains (S4 Fig and S1 Dataset) or by using HHpred, a sensitive detection method based on profile hidden Markov models (S2 Table; see Materials and Methods) [39]. Although the possibility cannot be completely excluded and future biochemical verification is needed, the likelihood of RT activity in SFKTE proteins is very low. In fact, plant PRV genomes usually possess various additional non-standard domains or ORFs that often play a role in vector transmission or immune suppression [26, 27]. SFKTE proteins may have functions similar to those of well-known additional PRV proteins, such as interaction with insect vector proteins or host antiviral factors [26, 27]. Although not necessary for its replication, the virus of eRTBVL may also benefit, to some extent, from such a function of SFKTE proteins encoded by the virus of eRTBVL2 during mixed infection. Consequently, an alternative relationship may exist between the two viruses: a mutualistic partnership. In the possible mutualistic partnership suggested for the viruses of ePVCVL and ePVCVL2 (Fig 5A), the two non-autonomous viruses benefit from each other via functional complementary. The RT/RH protein from the virus of ePVCVL2 reverse transcribes its own pregenomic RNA as well as that of the virus of ePVCVL, while the CP protein from the virus of ePVCVL assembles its own viral particles as well as those of the virus of ePVCVL2. Products from additional domains/ORFs of these two viruses (the SFKTE domain of the virus of ePVCVL and ORF2 of the virus of ePVCVL2) may also contribute to the putative mutualistic partnership. In the case of the non-Oryza group, the MP protein from the virus of ePVCVL is responsible not only for its own cell-to-cell movement, but also for that of the virus of ePVCVL2; at the same time, the region in the virus of ePVCVL2 slightly similar to the CP domain but lacking a zinc finger motif may encode defective CP proteins (i.e., those lacking viral DNA binding activity because of missing zinc finger motifs) to bind host antiviral proteins to disable viral-CP-binding activities. This system of two interplaying viruses is reminiscent of extant complex viruses possessing multiple polynucleotide sequences, which suggests that functional complementarity and co-regulation may have contributed to the origin of multipartite viruses.
The interspecies recombination event that generated the virus of eRTBVL2 (Fig 1B and S2 Fig) occurred between the viruses of eRTBVL (Tungrovirus-related species [29]) and ePVCVL (Petuvirus-related species; Fig 2A), which belong to different genera and possess distinct genomic structures with very weak sequence similarities. The presence of reshuffled domain combinations in the viral genome of eRTBVL2 relative to the virus of eRTBVL (Fig 1A and 1B) supports the theory of modular evolution that has been considered to be applicable to all known virus types [40, 41]. Putative interspecies recombination events have frequently been reported in viruses [42–46]. We propose that interspecies recombination is one of the mechanisms driving viral modular evolution. We particularly note that the frequent exchange of IGRs revealed in this study implies that modular evolution applies not only to coding domains, but also possibly to NRSs. Other studies have observed that recombination between endogenous and exogenous retroviruses has occasionally occurred and produced recombinant viruses [47–52]. This recombination may occur when exogenous and endogenous retroviral RNAs are coexpressed in host cells [47]. Recombination between endogenous and exogenous PRVs has not been reported to date [12]. Although in our study we also found no evidence to support the origin of any non-autonomous PRVs from such recombination, consideration of the evolutionary influence of this type of occasional albeit hypothetical recombination event is still of interest.
Concerted evolution has been widely observed to accompany the sequence homogenization process of some duplicated genes or elements in prokaryotic and eukaryotic genomes; one notable example is the sequence homogenization of ribosomal DNA repeats within a species [53, 54]. Concerted evolution has also been reported in nanoviruses, which are single-stranded DNA viruses [55, 56]. In our study, concerted evolution was observed during the homogenization of IGRs between a pair of partner viruses. IGRs are noncoding and highly divergent across PRV genomes; for example, IGRs of RTBV and PVCV respectively share less than 44.4% and 35.1% nucleotide identities with those of other PRVs (NCBI genome database). Nevertheless, the overall set of IGRs (and neighboring regions) between the two partner viruses in this study displayed an extraordinarily high sequence similarity (Figs 1B, 3B and 4B). This finding suggests that recombination, rather than mutation selection, is the main contributor to IGR homogenization between partner viruses. The results of our detailed phylogenetic and recombination analyses support the idea that persistent recombinations have driven this IGR concerted evolution (Figs 1C, 3C and 4C; S7 and S10 Figs). When we generated consensus sequences for eRTBVL2, ePVCVL, and ePVCVL2, we found a consensus pattern for each recombination breakpoint (Figs 1B, 3B and 4B, S2 and S3 Figs). This discovery suggests that these recombinations took place between homologous localized regions of two partner viruses; in other words, the recombinations were region-specific [23]. We propose the following model to explain the process of concerted evolution of IGR sequences (Fig 5B). Once illegitimate recombination produced identical (or highly similar) IGR sequences between the viruses of eRTBVL and eRTBVL2, mutations accumulated in these IGRs over time; however, region-specific recombination within homologous IGRs (and neighboring regions) of the two viruses exchanged these mutations between virus populations during mixed infection, with subsequent recombination within a viral population able to further spread the exchanged mutations. The constant repetition of this mutation–recombination cycle caused the two viruses in the putative partnership to maintain highly similar IGRs. As one of the two partner viruses diverged into a new lineage during evolution, the other coevolved via region-specific recombination between their homologous regions; this resulted in different viruses of eRTBVL2 possessing different IGRs that were highly similar to those of each of the viral lineages of eRTBVL groups (Fig 1C). Likewise, the constant repetition of this mutation–recombination cycle during the evolution of the viruses of ePVCVL and ePVCVL2 caused each partner of the virus pair infecting the same grass species to always maintain highly similar IGRs, even as the viruses of ePVCVL/ePVCVL2 diverged into distinct lineages infecting different host species in different habitats (Figs 3C and 4C, and S10 Fig). Consequently, divergent evolution occurred in each of the four studied virus species, whereas concerted evolution took place between the IGRs of each pair of partner viruses (Fig 5B).
Although precise quantification of the recombination frequency in these viral partnerships appears to be difficult, we tried to estimate the number of major recombination events between IGRs of partner viruses based on phylogeny. Phylogenetic clustering of eRTBVL2 IGRs with those of each of four eRTBVL groups (Fig 1C) suggested the occurrence of more than four major recombination events. Similarly, a total of 10 major recombination events were suggested by phylogenetic analyses of ePVCVL and ePVCVL2 IGRs (Figs 3C and 4C, and S10 Fig). In regards to the remaining grass genomes, which were not phylogenetically analyzed because of the high truncation and limited number of sequences, the independent endogenization and IGR concerted evolution of ePVCVL and ePVCVL2 in each genome imply that more than one major recombination event has taken place in each genome (a total of four) (Fig 2C and S3 Table). We consequently detected more than 18 independent major recombination events, which supports the idea that partner viruses have frequently recombined IGRs with each other to maintain partnership and coevolution. Although recombination has probably been much more frequent than we have estimated, these major events have had significant impacts on viral phylogeny during long-term evolution.
Similar to the recombination of retroviruses, PRVs such as cauliflower mosaic virus (CaMV) have been thought to recombine mostly through intermolecular template switching during reverse transcription in the host cytoplasm [23, 57, 58]. In our study, however, locational patterns of viral strand discontinuities (primer binding sites and polypurine tracts) did not correspond well to patterns of sequence similarity between viral genomes (Figs 1, 3 and 4, S2 and S3 Figs). When present in the host nucleus, PRV DNA is organized into minichromosomes [27], and indirect evidence exists that CaMV recombinations sometimes take place between viral minichromosomes [59, 60]. Consequently, the region-specific recombinations identified in this study may have occurred mainly through homologous recombination between local homologous regions of viral minichromosomes with the help of host recombination machinery. One homologous recombination mechanism, gene conversion, has been suggested to be responsible for the concerted evolution of ribosomal DNA and other genes [53, 61, 62].
Our study has provided paleogenomic evidence for non-autonomous PRVs as well as their adaptation. Considering the abundance of diverse EVEs harbored in eukaryotic genomes and the rapid accumulation of genomic data [3], many EVEs derived from previously unknown unusual virus types may still await discovery and analysis. At the same time, plentiful remnants of ancient virus–virus interactions may have been recorded in host genomes; our study has revealed one such paleovirological case of interplay between viral NRSs. One important future research focus should be evaluation of the prevalence and dynamics of NRS interactions between viral pathogens in mixed infections in plants and humans or within a viral population, as these may have significant impacts on viral evolution and pathology.
Whole-genome sequences of 20 grass species were downloaded mainly from the Gramene database [63] (detailed data sources in S1 Table). To identify endogenous PRVs, we first performed a BLASTn search (with default settings) using the BLAST+ 2.2.27 utility and previously reported sequences [64]. The hit sites (e-values < 1 × 10−10 and lengths >100 bp) along with their 5,000-bp upstream and downstream sequences were retrieved and assembled into consensus sequences (the nucleotide with the highest frequency at each position in the alignment was selected) using the Vector NTI Advance 11.5 toolkit (Invitrogen). A second round of BLASTn searching and a BLASTp search were then performed using these consensus sequences and their translated amino acid sequences, respectively. Only hit sequences longer than 100 bp were retained. Each translated protein sequence was subjected to the HHpred server [39], with all standard HHM databases (as of 3 May 2014) chosen for homologous domain detection (using default parameters). To check unidentified domains/ORFs, their amino acid sequences were resubmitted to the HHpred server and also subjected to BLASTp and tBLASTn searches against NCBI databases. Identified domains were confirmed by conserved motif alignment. Coordinates of eRTBVL2, ePVCVL, and ePVCVL2 sequences and their genes/regions in grass genomes are available in S2 Dataset (BED format). Dot plots were generated using the EMBOSS package (word size = 10; threshold = 45) [65].
Nucleotide sequences of each dataset were aligned in ClustalW [66] followed by manual editing. After being translated from the aligned nucleotide sequences, amino acid sequences of each dataset were realigned using MUSCLE [67] followed by manual editing. Highly truncated sequences (generally shorter than 80% of the entire region) and ambiguous regions were removed from the final alignments. Best-fitting substitution models were determined for each aligned dataset according to the Akaike information criterion calculated using jModelTest version 2.1.4 [68] or ProtTest version 3.2 [69]. For eRTBVL2 datasets comprising IGR (nucleotide positions 6063–6704 of the consensus genome), MP (486–1853), CP (1854–2845), PR (2831–4090), and ORFx (48–485) sequences, the best-fitting models were HKY+G, TrN+G, GTR+G, TrN+I+G, and TrN+G, respectively, with JTT+I+F chosen for the SFKTE sequences corresponding to amino acid positions 1220–1741 of the ORF2 protein sequence. Models VT+F+G and LG+I+F+G were respectively selected for the ePVCVL CP dataset (amino acid positions 709–996/722–1010 of the protein sequence of Oryza/non-Oryza groups) and the ePVCVL2 RT/RH dataset (amino acid positions 1017–1414/945–1342 of the ORF1 protein sequence of Oryza/non-Oryza groups). Models HKY+G, GTR+G, and HKY+G were respectively chosen for the IGR datasets of ePVCVL and ePVCVL2 of O. sativa, genus Oryza, and S. bicolor genomes (nucleotide positions 5878–6415/5786–6323, 5878–6611/5786–6519, and 6008–6659/5691–6317 of the consensus genomes of ePVCVL/ePVCVL2, respectively). Maximum-likelihood (ML) phylogenetic analyses were performed in PhyML version 3.0 [70] or MEGA version 6.06 (only for Fig 1C and S6 Fig for display purposes) [71]. Branch support in all trees was calculated using 1,000 bootstrap replicates. The tree for the SFKTE domain of eRTBVL2 and ePVCVL segments was rooted using the oldest ePVCVL segment as determined by orthology analysis of ePVCVL loci in Oryza species (see below). ePVCVL was assumed to be older than eRTBVL2, as eRTBVL2 only exists in a subspecies of O. sativa, whereas ePVCVL is present in all O. sativa subspecies (see S8 Fig). All sequence alignments for phylogenetic analyses are available in S3 Dataset.
Sequences suggested as having a high probability of recombination according to the phylogenetic analyses and sequence alignments were subjected to recombination analyses using RDP version 4.72 [72]. We used six different methods (RDP [73], GENECONV [74], BootScan [75], MaxChi [76], Chimaera [77], and SiScan [78]) in this program to identify potential recombination events and perform statistical tests. Sequence alignments for the recombination analyses were generally extracted from the alignment datasets of phylogenetic analyses. In the case where no suitable phylogenetic dataset was available, sequence alignments used for recombination analyses were made in MUSCLE [67] followed by manual editing. Default parameters were used for each method, except that the reference sequence parameter of the RDP method, in accordance with the RDP manual, was adjusted to “internal references only” when many closely related sequences existed in the alignment [72]. For each method, P < 0.005 was used as a threshold value for possible recombination events. Only the recombination events independently detected by more than three methods with statistical significance were considered reliable, and the best P value for each event was chosen. These recombination events were checked and displayed in BootScan plots (window size = 300 nt; step size = 10 nt) using the RDP program. All alignments used for recombination analyses are available in S4 Dataset.
If an ePVCVL/ePVCVL2 segment in an Oryza genome was located next to or near another ePVCVL or ePVCVL2 segment (i.e., less than 5 kb away on the same chromosome or scaffold), the two (or more) segments were generally considered to be one locus for the analysis. The left and right 5-kb flanking sequences of each locus of ePVCVL and ePVCVL2 in the O. sativa genome were first mapped onto five other Oryza genomes (O. glaberrima, O. glumaepatula, O. longistaminata, O. meridionalis, and O. punctata) using BLASTn. The mapping results were rechecked using genome collinearity data (genome-wide alignments between Oryza genomes) obtained from the Gramene database [63]. Both 5-kb flanking sequences of each locus of ePVCVL and ePVCVL2 in the five above-mentioned Oryza genomes were next mapped onto the O. sativa genome and rechecked in the same manner. Some flanking sequences in O. glumaepatula and O. meridionalis genomes contained many uncharacterized (‘N’) bases; the examined length of these flanking sequences was therefore extended to 15 kb.
Genomic PCR and Sanger sequencing were used to confirm orthologous loci of ePVCVL and ePVCVL2. Loci shared among Oryza species were examined; in addition, representative O. sativa-specific ePVCVL and ePVCVL2 loci were selected and analyzed. Wild and cultivated rice plants (accession numbers in S9 Fig) were grown in a greenhouse at Hokkaido University, Sapporo, Japan. Total DNA was extracted from leaf samples using cetyltrimethylammonium bromide extraction buffer. DNA concentrations were all diluted to the same order of magnitude. PCR amplifications were performed using Ex Taq or LA Taq polymerase (Takara) on a PTC-200 thermal cycling system (GMI). PCR products were resolved on a 1–2% agarose gel, stained with ethidium bromide, and viewed using an AE-6933FXES Printgraph system (ATTO). Sanger sequencing was performed on an ABI 3730 DNA Analyzer (Applied Biosystems) using a BigDye Terminator v3.1 cycle sequencing kit (Applied Biosystems) according to the manufacturer’s protocol. Information on the primers used in this study is provided in S6 Table.
All relevant data are within the paper and its Supporting Information files except for the assembled sequences of non-autonomous PRVs, which are available from DDBJ database under accession numbers BR001403–BR001407.
|
10.1371/journal.pmed.1002617 | Heat-related mortality trends under recent climate warming in Spain: A 36-year observational study | Anthropogenic greenhouse gas emissions have increased summer temperatures in Spain by nearly one degree Celsius on average between 1980 and 2015. However, little is known about the extent to which the association between heat and human mortality has been modified. We here investigate whether the observed warming has been associated with an upward trend in excess mortality attributable to heat or, on the contrary, a decrease in the vulnerability to heat has contributed to a reduction of the mortality burden.
We analysed a dataset from 47 major cities in Spain for the summer months between 1980 and 2015, which included daily temperatures and 554,491 deaths from circulatory and respiratory causes, by sex. We applied standard quasi-Poisson regression models, controlling for seasonality and long-term trends, and estimated the temporal variation in heat-related mortality with time-varying distributed lag nonlinear models (DLNMs). Results pointed to a reduction in the relative risks of cause-specific and cause-sex mortality across the whole range of summer temperatures. These reductions in turn explained the observed downward trends in heat-attributable deaths, with the only exceptions of respiratory diseases for women and both sexes together. The heat-attributable deaths were consistently higher in women than in men for both circulatory and respiratory causes. The main limitation of our study is that we were not able to account for air pollution in the models because of data unavailability.
Despite the summer warming observed in Spain between 1980 and 2015, the decline in the vulnerability of the population has contributed to a general downward trend in overall heat-attributable mortality. This reduction occurred in parallel with a decline in the vulnerability difference between men and women for circulatory and cardiorespiratory mortality. Despite these advances, the risk of death remained high for respiratory diseases, and particularly in women.
| The Euro-Mediterranean region arises as a major climatic hot spot because of global warming.
Warmer temperatures should in principle contribute to an increase in the number of deaths because of heat.
We do not know yet if and to what extent societal adaptation and/or socioeconomic development is modifying this expected increase.
We analysed daily mortality records from 47 major cities in Spain.
There has been a general and sustained decline in the vulnerability of the population since 1980.
Despite the observed warming, the decline of the vulnerability has generally contributed to a progressive reduction in the number of deaths attributed to heat since 1980.
It is generally believed that climate change will cause an increase in heat-related mortality.
Societal adaptation and/or socioeconomic development contributed, up to now, to a general decline in heat-related mortality.
It is still uncertain if this decline in heat-related mortality will also occur at higher future levels of climate warming.
| Anthropogenic climate change represents a major threat for human health and a challenge for public health services [1]. One of its most important effects is the potential increase in heat-related mortality resulting from rising temperatures and the associated increase in the frequency, intensity, and duration of extreme heat events [2,3]. However, the extent of this impact will not only depend on the increase in the level of exposure to heat but also on any underlying change in the vulnerability of the exposed population [4].
Several factors have the potential to modify population vulnerability over time and, therefore, the eventual incidence of increasing temperatures on heat-related mortality. In ageing societies such as Europe, the rising elderly population is expected to increase vulnerability to high ambient temperatures, given that the elderly have diminished physiological capacity for the regulation of body core temperature under heat stress conditions [5]. On the contrary, general improvements in housing conditions (e.g., wider use of air conditioning systems in retirement homes) and healthcare services (e.g., improved treatment of heat-related morbidity) [6,7] as well as some planned adaptive measures to reduce the exposure and vulnerability to heat (e.g., implementation of effective heat health warning systems) [8] could all contribute to reducing the negative health consequences of temperatures and warming trends.
The Euro-Mediterranean region arises as a major climatic hot spot as a result of global warming [9], and, particularly, Spain was the country with the largest relative number of excess deaths during the record-breaking summer 2003 heat wave [10], even if interannual temperature anomalies were twice as large in France during the episode [11]. Some authors have shown that there has been a decline in heat-mortality associations during the last decades in some, albeit not all, of the countries [12–15], and especially since this event in a subset of locations [16–18]. In the case of Spain, the vulnerability associated with heat was found to decrease for the range of extreme summer temperatures [12], suggesting an adjustment response of the Spanish society to rising temperatures despite the ageing of the population. However, evidence about the risk attributable to heat, either as absolute or relative excesses or its temporal evolution, is lacking, with no clear upward or downward trend in the number of deaths. Furthermore, changes by cause of death and sex have not been described for Spain, nor for the rest of the countries.
In the present work, we assess the impact of the summer warming observed in Spain during the period 1980–2015 on cause-specific mortality by sex. The main objective was to determine whether warmer summers were associated with an upward trend in excess mortality attributable to heat. Addressing the early impacts of increased ambient temperatures on human health is a relevant question that could translate into more effective public health adaptation strategies to current and future climate change conditions.
The Spanish National Statistics Institute (INE) provided daily death counts from circulatory (ICD-9: 390–459, ICD-10: I00–I99) and respiratory (ICD-9: 460–519, ICD-10: J00–J99) diseases, disaggregated by sex and covering the period from 1 January 1980 to 31 December 2015 in 47 major Spanish cities. Daily mortality data had no missing value. We derived daily mean 2-meter temperature observations from the European Climate Assessment and Dataset (ECA&D), which were computed as the average between daily maximum and minimum values from meteorological stations. About 1% of the temperature time series was missing data.
The analysis was restricted to the warm season from June to September, and it was performed in two stages. In the first part, standard quasi-Poisson regression models allowing for overdispersion were individually applied in each of the 47 cities included in the analysis in order to estimate location-specific temperature-mortality associations, summarised in terms of relative risk (RR) values by cause of death and sex. The seasonal trend was controlled for in the models by using a natural cubic B-spline of day of the season, with 2 degrees of freedom per year/summer and equally spaced knots, and it was allowed to vary from one year to another through the specification of an interaction between the B-spline and indicators of summer/year. The models also included a natural cubic B-spline of time, with 1 degree of freedom per decade and equally spaced knots to control for the long-term trend, as well as a categorical variable to control for the day of the week.
The complex nonlinear and delayed dependencies found for temperature and mortality were captured by using a distributed lag nonlinear model (DLNM), a flexible methodological framework widely used to investigate the health effects of air pollution and temperature. This model is based on the definition of a cross-basis function, obtained by the combination of two functions describing the exposure-response association and the lag-response association [19]. Specifically, the exposure-response curve was modelled through a natural cubic B-spline, with one internal knot placed at the 15th percentile of the daily temperature distribution, and the lag-response curve was modelled through a natural cubic B-spline, with an intercept and two internal knots placed at equally spaced values in the log scale, with a lag period extended up to 10 days to account for the lagged effects of heat and short-term harvesting. In this way, the overall effect of a given summer day temperature on mortality was defined as the sum of the effect on that day and the 10 subsequent days. These modelling choices were thoroughly tested in sensitivity analyses, which are shown in the Supporting information (S1 Fig). The Poisson regression model for the whole study period was given as follows:
LogE(Yt)=intercept+cb+dow+S1(dos,df=2):factor(year)+S2(time,df=1perdecade)
where Yt denotes the series of daily mortality counts, cb the cross-basis matrix produced by DLNM, dow the day of the week, S1 the natural cubic B-spline of the day of the season, and S2 the natural cubic B-spline of time.
To model temporal variations in the bidimensional exposure-lag-response associations between temperature and mortality, the DLNM model described above was extended to a time-varying DLNM, specified through a linear interaction between the cross-basis and time variables [20]. This extension allowed us to predict the temperature-mortality relationship for each summer in the series, by centring the time variable in the central day of the respective summer season.
In the second stage, a simple multivariate random-effects meta-analysis [21] was used to estimate the average temperature-mortality association across cities for the whole study period (estimates provided by the model without interaction) and for each year in the series (estimates provided by the model with interaction). The fitted meta-analytical model was also used to derive the best linear unbiased predictions (BLUPs) of the temperature-mortality relationships and the related point of minimum mortality temperature (MMT) in each location. Uncertainty in estimated MMTs was investigated through the method described in Tobías et al. [22].
To assess the temporal evolution in the effect of heat on mortality, the pooled RR curves from the time-varying DLNMs with interaction terms were compared between years. The temporal variations in the RR curves were analysed by means of a multivariate Wald test on the pooled coefficients of the interaction terms, which represent the change in the average temperature mortality curves. The null hypothesis of the test is that no change in the temperature-mortality association ocurred throughout the study period. We also summarised the results by calculating the pooled RR at the 90th and 99th temperature percentiles from year-specific curves between 1980 and 2015.
The mortality burden attributable to heat across the whole study period, reported as relative excess (i.e., attributable fraction) of deaths, was estimated by using the methodology developed by Gasparrini and Leone [23]. First, the RR of mortality corresponding to each day and city was used to calculate the attributable fraction of deaths on that day and the next 10 days. Then, the daily attributable number of deaths was computed by multiplying the daily attributable fraction by the daily number of deaths. The overall number of attributable deaths caused by heat was given by the sum of the contributions from all days of the series with temperatures higher than the value of MMT derived from the BLUP of the model with no interaction in each city, and its ratio with the total number of deaths provided the total heat-attributable fraction. This attributable component was separated further into the contributions of moderate and extreme heat. Moderate heat was defined as the range of temperatures between the city-specific MMT and the city-specific 97.5th daily temperature percentile, and extreme heat as the range of temperatures warmer than this threshold. We also computed the attributable risk of heat for each summer from time-varying DLNMs (model with interaction) to investigate temporal variations in heat-attributable deaths. Confidence intervals (CIs) of attributable risk were obtained empirically through Monte Carlo simulations.
All statistical analyses were performed with R software (version 3.4.3) using functions from the packages dlnm (first-stage regression) and mvmeta (second-stage meta-analysis).
No data-driven changes were done during the analyses. The development of the statistical analysis plan, including changes inspired by referees, is described in S1 Analysis Plan. This study is reported as per the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines (S1 STROBE Checklist).
We analysed data from 47 major cities representing about 32% of the total Spanish population, which includes 544,491 summer deaths, corresponding to the period 1980–2015. Circulatory counts represented 78.9% of the total cardiorespiratory mortality (here the word ‘cardiorespiratory’ is used to refer to deaths from circulatory and respiratory deaths together), while respiratory deaths accounted for the remaining 21.1%. The temporal pattern of each cause of death was similar in men and women (S2 Fig), with a decline in the number of deaths from circulatory diseases and an increase in the number of deaths from respiratory diseases. Nevertheless, mortality decreased at a slower pace in women for circulatory diseases, therefore increasing the magnitude of the difference between women and men. Moreover, for respiratory diseases, mortality increased at a faster pace in women, therefore reducing the magnitude of the difference between women and men. The geographic distribution of the cities included in the analysis, along with the corresponding evolution in overall summer mean temperature, are displayed in Fig 1. As expected, summer temperatures have been increasing, on average, at a rate of 0.32°C per decade.
Fig 2 depicts the pooled RR values associated with the temperature-mortality relationships by cause of death and sex from the model with no interaction, interpreted as the average relation throughout the whole study period (1980–2015). The temperature-mortality relationships computed as BLUPs for the 47 cities and the corresponding MMTs are provided in the Supporting information, S3 Fig and S4 Fig. All these curves are J shaped, indicating a monotonically increasing mortality risk for temperatures above and below the MMT. The slope of the curve above this point varied greatly among cases, being generally larger for respiratory than circulatory diseases, and for women than for men. Thus, in each of these two groups of causes, taken either individually or together, women showed systematically higher values of RR for the whole range of warm temperatures, with generally lower MMT than men. For the ensemble of cities, the values of the MMT and the summer mean temperature are largely correlated, with a spatial dependency of between 0.81 and 0.96°C in MMT per 1°C in summer mean temperature (Student t test/p-value < 0.001, S5 Fig).
Results from the analysis of the temporal evolution of the temperature-mortality associations predicted from the model with interaction (time-varying DLNM) are summarised by cause of death and sex in Figs 3 and 4. On the one hand, Fig 3 displays the comparison of the pooled RR curves for representative years (i.e., every 5 years; see also S6 Fig for estimates in each of the 47 cities for years 1980 and 2015). On the other hand, Fig 4 shows the temporal evolution of the pooled RR corresponding to the 99th temperature percentile of the whole time period (left panels) and corresponding to the time-varying 99th temperature percentile computed from data of each individual summer (right panels; find equivalent results for the 90th percentile in S7 Fig). Temporal changes suggest a strong and progressive reduction in the RR of heat-related mortality during the whole study period for the pooled analysis and generally for the ensemble of cities, but despite these advances, the RR to extreme warm temperatures remained high for respiratory diseases, and particularly in women. The significance test (S1 Table) indicates strong evidence for significant temporal changes in the RR curves, with the only exception being respiratory mortality for women. As a result of all these results, differences in mortality risk between men and women from both circulatory and cardiorespiratory diseases generally declined during the study period.
Table 1 summarises the estimates of mortality deaths attributable to heat during the whole study period. The overall fraction of deaths caused by summer heat was 10.65% (95% CI 9.93–11.33), with moderate heat being responsible for almost nine times as many deaths as extreme heat. This was explained by the fact that the days with moderate temperatures occur more frequently in the series, and, therefore, they have a higher absolute impact on the total number of deaths. The attributable fraction due to respiratory causes (17.60%, 95% CI 16.00–18.95) was twice as high as that from circulatory causes (8.94%, 95% CI 8.22–9.61), although in absolute terms circulatory deaths accounted for most of the mortality burden. When considering these two groups of causes by sex, both the attributable fractions and numbers were substantially higher for women. The same pattern was separately observed for moderate and extreme heat.
The black curves in Fig 5 depict the temporal evolution of the attributable fraction by cause of death and sex (see S8 Fig for the relative contribution of moderate and extreme heat). Regardless of the interannual variability and warming trend associated with the year-to-year changes in summer temperature, the heat-attributable fraction shows no clear and generalised upward trend during the study period. The only exception is found for respiratory diseases (for men and women together, and for women only), in which case the rather small reduction in RR only partially contributed to reducing the negative impact of increasing summer temperatures. Apart from these few exceptions, however, the general evolution shows that, regardless of the summer warming trend shown in Fig 1, the general lack of increasing trends in the attributable fraction is explained by the large decrease in the mortality RR shown in Figs 3 and 4.
The red and blue curves in Fig 5 show the temporal evolution of the attributable fraction to heat if the effect of long-term warming temperatures were removed. This hypothetical evolution is constructed by replacing the temperature time series by the daily time series of a given summer (i.e., 122 values corresponding to the four summer months here considered) that is repeated by construction for all the years in the study period. Thus, the blue and red lines result from predicting the attributable fraction by applying the time-varying annual RR curves (1980–2015) to the daily time series of temperatures for summers 1984 (i.e., the coldest during the study period) and 2003 (the warmest), respectively. In this way, given that more moderate summer temperatures were observed in 1984 and more extreme temperatures in 2003, the magnitude of the corresponding trends in the attributable fraction results to a large extent from the decreasing magnitude in the RR for the moderate or extreme temperature percentiles, respectively (cf. with Fig 4). In this way, the range between these lines as well as differences with regard to the central curves show the relative contribution of interannual temperature anomalies and year-to-year climate variability to the evolution of the attributable fraction; they also show a rough estimation of the potential impact range of near-future summer temperatures. For example, compared with the attributable fraction observed in 2003, the estimated impact of a 2003-like summer with the RR of 2015 would be 21.13% smaller for men, 28.16% for women, 38.31% for cardiovascular diseases, and 12.34% for respiratory diseases (i.e., decrease of the red curves between years 2003 and 2015).
To the best of our knowledge, this is the first study that comprehensively addresses the eventual impact of recent climate warming on summer mortality in Spain by cause of death and sex. The study pointed to a strong reduction in cause-specific and cause-sex mortality RR associated with summer temperatures for the last three and a half decades and, with the exception of respiratory diseases (for men and women together, and for women only), downward trends in heat-attributable deaths. These results strongly support the hypothesis that the observed warming trend in summer temperatures in Spain has not been paralleled by a general increase in the mortality fraction attributable to heat, as a result of substantial decline in population vulnerability to warm temperatures [11].
In this study, the effect of heat on mortality largely varied by cause of death. We observed a greater impact of heat on respiratory rather than circulatory mortality. This is in agreement with previous studies [24–27] and probably reflects the large health vulnerability among people with pre-existing or chronic respiratory diseases during hot periods [28]. It should be remembered that respiratory mortality accounted for only 21.1% of cardiorespiratory deaths during the study period, but the observed rise in the relative prevalence of respiratory diseases over recent decades might continue in the near future and keep increasing the relative percentage of the population susceptible to heat-related respiratory mortality. The underlying physiological mechanisms behind the effect of heat on mortality from circulatory and respiratory causes are not well known yet, but they seem to be largely mediated by a thermoregulatory pathway.
We assessed circulatory- and respiratory-specific mortality by sex, and we found that women were systematically more at risk of dying from heat. This finding has been reported in many previous articles analysing the effect through the whole range of hot temperatures [27,29–31] and in many studies based on extreme heat events [16,32–35]. These quantitative differences between men and women may partially arise from physiological characteristics in body temperature regulation between males and females [36]. However, most of these differences could simply be attributed to existing sociodemographic characteristics of the society (e.g., differences between men and women in age pyramid, life expectancy, or social isolation). Previous studies have actually shown that this type of difference can in some cases result in gender being an important variable to predict the mortality risk; i.e., 64% of the deaths were women during the 2003 heat wave in France [33]. The gender gap was, however, found to be reduced by the group of planned adaptation measures implemented just after this record-breaking episode [16]: while 60% of the deaths were expected to be women during the following major heat wave in 2006, gender differences in observed mortality were significantly lower; i.e., 53% of the deaths were women.
The heat attributable fractions reported in our study are higher than those presented, for example, in Gasparrini et al. [37] and somewhat different, for example, from those reported by Carmona et al. [38]. However, these results are not comparable, because in our study, mortality data are analysed only for circulatory and respiratory causes, which are two of the main groups associated with ambient temperatures (see, for example, Basagaña et al. [39]). Instead, in Gasparrini et al. [37], mortality data for Spain are analysed for all causes of death together, and in Carmona et al. [38] for natural causes, including many causes of death that are not largely associated with ambient temperatures. As a result, the attributable fraction (i.e., the ratio between attributable deaths and total deaths) is understandably higher in our study than in Gasparrini et al. [37] and Carmona et al. [38]. We note that other methodological differences might explain (to a lower extent) these differences; e.g., the time period of data is different (taking into account that the relative risk decreases with time).
The temporal evolution of heat-related mortality risks here found is, in general, consistent with those reported by previous studies in some other countries [12–15], which provide evidence for a decrease in vulnerability to climate warming despite the ageing of societies. For example, in Spain, the proportion of people aged over 64 years increased from 11.6% to 15.0% in men and from 15.9% to 19.6% in women between 1991 and 2011 [40]. The general downward trend in mortality risks has been attributed by some investigators to socioeconomic development and structural transformations, such as improvements in housing and healthcare services [12–15], or even to specific public health interventions [16–18]. The large socioeconomic advances that occurred in Spain during the last decades might have also contributed to this response, thus reducing the effect of mortality risks over time. For example, the gross domestic product (from €8,798 per capita in 1991 to €22,813 in 2009), the life expectancy at birth (from 77.08 years to 81.58), the expenditure in healthcare (from €605 per capita to €2,182) and social protection (from €1,845 per capita to €5,746), and the number of doctors (from 3,930 per million inhabitants to 4,760 per million inhabitants) have all largely increased in Spain [41]. In addition, the use of air conditioning, which has been postulated as a major contributor to the reduction in heat-related mortality in the United States [13], has also experienced a strong increase in Spanish households within the analysed period (from 5.3% to 35.5%) [42].
Another potential contributing factor to the reduction of the mortality risks might have been the ‘National plan for preventive actions against the effects of excess temperatures on health’ from the Spanish Ministry of Health [43], which was implemented in 2004, just after the 2003 summer heat wave, in order to minimise the negative effects of summer extreme temperatures on the population’s health, particularly among vulnerable groups such as the elderly, children, people with chronic disease, and disadvantaged persons. Nonetheless, we do not see a change in the slope of the generally linear declining trend of RR and attributable fraction in the mid-2000s, which seems to indicate that this measure had at most a minor beneficial impact (specific analyses will be performed elsewhere). The cause-specific and cause-sex attributable numbers all showed a rather linear downward trend during the study period as a result of the strong decrease in RR, which was largest for the warmest temperatures. Although the range corresponding to moderate hot temperatures had a comparatively lower RR, it included the majority of days in the series, accounting for most of the overall deaths caused by heat. In that regard, public health interventions should also be directed towards non-extreme temperatures, to include the whole range of moderate conditions above the MMT [37].
Finally, some limitations of the study deserve to be mentioned. First, we were not able to control for air pollution because of data unavailability. However, previous studies showed that effects of hot temperatures on mortality in the US were only slightly reduced after adjusting for air pollution [24]. Secondly, the study did not take into account the long-term changes in the age structure of the population, which will be analysed when a more complete dataset by age is available.
|
10.1371/journal.pbio.0050230 | Insights into the Genome of Large Sulfur Bacteria Revealed by Analysis of Single Filaments | Marine sediments are frequently covered by mats of the filamentous Beggiatoa and other large nitrate-storing bacteria that oxidize hydrogen sulfide using either oxygen or nitrate, which they store in intracellular vacuoles. Despite their conspicuous metabolic properties and their biogeochemical importance, little is known about their genetic repertoire because of the lack of pure cultures. Here, we present a unique approach to access the genome of single filaments of Beggiatoa by combining whole genome amplification, pyrosequencing, and optical genome mapping. Sequence assemblies were incomplete and yielded average contig sizes of approximately 1 kb. Pathways for sulfur oxidation, nitrate and oxygen respiration, and CO2 fixation confirm the chemolithoautotrophic physiology of Beggiatoa. In addition, Beggiatoa potentially utilize inorganic sulfur compounds and dimethyl sulfoxide as electron acceptors. We propose a mechanism of vacuolar nitrate accumulation that is linked to proton translocation by vacuolar-type ATPases. Comparative genomics indicates substantial horizontal gene transfer of storage, metabolic, and gliding capabilities between Beggiatoa and cyanobacteria. These capabilities enable Beggiatoa to overcome non-overlapping availabilities of electron donors and acceptors while gliding between oxic and sulfidic zones. The first look into the genome of these filamentous sulfur-oxidizing bacteria substantially deepens the understanding of their evolution and their contribution to sulfur and nitrogen cycling in marine sediments.
| In 1888 Winogradsky proposed the concept of chemolithotrophy—growth using inorganic compounds as an energy source—after studying the sulfur bacterium Beggiatoa. These filamentous bacteria and related organisms inhabit the surface of marine and freshwater sediments, where they oxidize hydrogen sulfide using either oxygen or nitrate. In particular, conspicuously large marine representatives accumulate nitrate in vacuoles to survive anoxia, a unique feature among prokaryotes. Since nitrate-storing Beggiatoa are not available in pure culture, we amplified and sequenced the genomic DNA of single multicellular filaments. We comprehensively tested the incomplete sequence assemblies for foreign DNA. We show that the Beggiatoa genome encodes the pathways of chemolithoautotrophy but also appears to support the use of alternative electron donors and acceptors. We propose that vacuolar-type ATPases generate an electrochemical gradient to drive nitrate transport over the vacuole membrane, a mechanism similar to eukaryotic solute accumulation. Intriguingly, we found evidence for substantial gene exchange between Beggiatoa and cyanobacteria. In both phyla, hemagglutinins are possibly involved in filament formation. The breadth of storage and metabolic capabilities encoded in its genome enables Beggiatoa to act as a “rechargeable battery,” which glides between oxic and sulfidic zones to overcome non-overlapping availabilities of electron donors and acceptors.
| Mats of conspicuously large sulfur-oxidizing bacteria often cover the seafloor in organicly rich coastal areas, at hydrate ridge methane seeps, at hydrothermal vents, on whale falls, and in coastal upwelling regions [1–5]. The closely related genera Beggiatoa, Thioploca, and Thiomargarita are among the largest prokaryotes known, and they usually contain a vacuole that can account for up to 90% of the cell volume [6]. On the seafloor these large sulfur-oxidizing bacteria fulfill an important ecological function by preventing the release of toxic hydrogen sulfide from the sediment into the water column. Studying Beggiatoa, Winogradsky [7] demonstrated the principle of chemolithotrophy, a process in which the oxidation of inorganic sulfur is coupled to oxygen respiration. By their gliding motility Beggiatoa aggregate at the oxic–anoxic transition zone, where oxygen and sulfide occur in opposed diffusion gradients [3,8]. Beggiatoa compete using chemical sulfide oxidation [8,9], mainly by Fe(III), and can significantly contribute to biological sulfur oxidation [10,11]. Oxygen has been regarded as the major electron acceptor coupled to sulfur oxidation; however, there is growing evidence that when experiencing anoxia these large vacuolated Beggiatoa, Thioploca, and Thiomargarita respire nitrate, which they concentrate up to 10,000-fold (∼500 mM) within their intracellular vacuoles [5,12,13]. Their nitrate and sulfur storage capacities allow them to bridge the suboxic zone, where neither sulfide nor oxygen is detectable, which gives them an advantage over other sulfide-oxidizing bacteria. In addition, these large sulfur-oxidizing bacteria may release phosphate from accumulated polyphosphate (polyP), which has been hypothesized to account for the large phosphorite deposits on the seafloor [14,15].
None of these large nitrate-storing bacteria are available in pure culture. Thus, little is known about the gene content associated with their chemolithotrophic properties, their conspicuous morphology, or their exceptional nitrate storage abilities. Previous physiological and genetic studies were mainly performed on the small, readily culturable, non-vacuolated B. alba, a species that is phylogenetically distant from the large sulfur-oxidizing bacteria [16]. Because of phenotypic similarities such as gliding motility and filamentous shape, Beggiatoa spp. were regarded as colorless cyanobacteria (discussed in [17]) before they were reclassified as Gammaproteobacteria based on 16S rRNA gene sequences.
It is now standard to study large genomic fragments of uncultured microbes by shotgun cloning and sequencing of bulk DNA extracted from mixed communities [18–20]; however, assembly of genomes for discrete species is problematic. Alternatively, DNA can be exponentially amplified (up to 109-fold) from single cells [21] by multiple displacement amplification (MDA) [22–25], enabling sequencing from uncultured microorganisms isolated from the environment [25–27]. Despite background amplification and chimera formation [28], this method amplifies complex DNA much more faithfully than earlier whole genome amplification (WGA) strategies. Recently, more than 60% of the genome of single cultured Prochlorococcus cells were amplified and sequenced with improved methods that greatly reduced background amplification and chimera formation [29]. Here, the cloning of amplified, hyperbranched DNA was suspected to facilitate the formation of chimeric sequences. However, chimeric sequences can occur to a similar extent in pyrosequenced datasets [28], indicating that MDA is the causative agent in chimera formation. Non-electrophoretic sequencing methods such as pyrosequencing [30] offer the advantage of massively parallel sequencing of large numbers of DNA fragments without cloning and hence less chimera formation. They also obviate the problems of cloning bias and of sequencing GC-rich DNA.
The combination of the low representational bias of MDA-amplified genomic DNA with the advantages of clone-free pyrosequencing augers well for the great potential to rapidly analyze the genomes of unculturable microbes. Here, we report what is to our knowledge the first large-scale genomic analysis of an uncultured, environmental bacterium based on WGA and pyrosequencing. Using MDA the genomic DNA of two individual multicellular (>600 cells) filaments of uncultured Beggiatoa (∼30 μm in diameter) from a Baltic Sea harbor sediment were separately amplified. One of these amplification products was sequenced using a clone-free pyrosequencing method developed by 454 Life Sciences [31]; the other was sequenced using electrophoretic (Sanger) sequencing of clone libraries. To estimate the heterogeneity among individual Beggiatoa filaments and the proportion of the Beggiatoa genome covered by our sequences, the genome size was independently determined by optical mapping [32] using filaments of co-occurring Beggiatoa.
Here, we present the draft genome sequences of two individual filaments of Beggiatoa sp. recovered from the surface of a marine sediment. The sediment–water interface in marine and freshwater habitats is characterized by steep gradients of electron donors and acceptors such as sulfide, oxygen, and nitrate. Since the zones of availabilities of electron donors and acceptors usually do not overlap, nitrate-storing Beggiatoa move between the oxic and sulfidic sediment layers to overcome this limitation. In the following, the general genome features and genome-encoded adaptations for this lifestyle in two individual Beggiatoa filaments are illustrated. In particular, we focus on the chemolithotrophy and the unique storage capabilities of the vacuolated Beggiatoa. Furthermore, we provide evidence of horizontal gene transfer with cyanobacteria, which likely reflects the long-term coexistence of these two phyla at sediment surfaces.
Comprehensive genomic analysis of specific environmental microorganisms is hampered by a high microdiversity of co-occurring and closely related organisms [33]. Hence, accurate estimates of sequence heterogeneity and genome size are required. To estimate the heterogeneity and the genome size of large, uncultured Beggiatoa, we performed optical mapping of single DNA molecules. Unamplified, high-molecular-weight DNA molecules were isolated from five co-occurring, 35-μm-diameter filaments, each composed of more than 600 putatively clonal cells descended from the filament's progenitor cell. We used a small number of morphologically identical Beggiatoa filaments to reduce the risk of obtaining mapping data compromised by co-occurring and closely related organisms. The DNA from the Beggiatoa yielded a consensus optical map of a single circular chromosome of approximately 7.4 Mb (Figures 1A and S1). This is over twice the estimated size (3 Mb) of the genome of the non-vacuolated species B. alba [34]. Consensus maps were also obtained for four linear contigs, with sizes ranging from 0.9 to 3.4 Mb (Figure 1B–1E). In some regions the restriction patterns of the consensus maps of these smaller linear contigs were similar to regions of the consensus map of the larger circular chromosome, whereas other regions were highly dissimilar. The diverging DNA restriction patterns of the five contig maps are likely not attributable to an unusually high genome plasticity but rather reflect the high microdiversity among the five Beggiatoa filaments, as has been reported for marine Vibrio spp. [33]. This led us to sequence the genome of a single filament rather than the metagenome of a mixed community of closely related species (Figure 2B).
WGA using MDA from a single or a few cells is highly sensitive to random DNA synthesis. It is also compromised by the presence of non-target DNA, which is a major concern particularly in environmental projects. To minimize these problems we obtained Beggiatoa DNA from a well-purified multicellular filament consisting of more than 600 cells to provide a large number of putatively clonal chromosome copies as a template for WGA. Consistently, the data analysis strongly suggests successful amplification and assembly of genomic DNA from Beggiatoa filaments cells exclusively, even though the filaments had been obtained directly (without prior cultivation) from a marine sediment.
The whole genomic DNA of a single filament was amplified using MDA. From the amplified DNA a clone library was constructed that was Sanger sequenced (SS dataset). This approach yielded a low-coverage (3×) partial assembly of 1,091 contigs with a total length of 1.3 Mb (Table 1). In a separate experiment the DNA from a second filament was amplified and subsequently pyrosequenced (PS dataset). The PS assembly achieved a high coverage depth (17×) and a total length of 7.6 Mb. A detailed overview of the sequencing results and preliminary genome features are given in Tables 1, S1, and S2. The maximum contig size was 18.6 kb for the PS genome and 5.5 kb in the SS genome.
For open reading frame (ORF) prediction, only contigs larger than 2 kb were considered. The average ORF length was 594 bp (SS2) and 827 bp (PS2). The high number of short, non-overlapping contigs (Table 1) suggests a genome larger than 7.6 Mb. Reconciliation of the optical genome map of Beggiatoa (Figures 1 and S1) with the Beggiatoa PS genome sequence was impractical because of the incomplete sequence assembly.
The low level of sequence assembly is not attributed to high genome plasticity among cells in a single filament. We assume that a multicellular filament is derived from one progenitor cell and thus is clonal. It is highly unlikely that massive genome rearrangements occur within approximately ten generations (29–210 cell divisions = 512–1,024 cells/filament). Thus, the sequence dataset of each filament represents the genome of a single strain rather than a population of slightly different genomes or even a metagenome of mixed organisms.
Several tests at different stages of this study were conducted to determine if there was any significant contribution of potential non-Beggiatoa DNA to the PS sequence assembly: (1) an analysis of the PS sequence read metadata; (2) an analysis of intrinsic DNA signatures of the assembled sequences, and (3) genome annotation and phylogenetic reconstruction of different marker molecules and analysis of single-copy genes. The results of these analyses are highly consistent with the claim that the assembled sequences are derived from Beggiatoa only.
Reads from repeat regions (excluded from an assembly) were an unusually high percentage (11.3%) of the PS reads (Tables S1 and S3). It is unclear if this reflects the repetitive DNA content of the Beggiatoa genome, or if this is an artifact of WGA. The assembled and repeat reads had similar properties (Table S2), and there were multiple examples of repeat reads with more than ten copies, indicating they were not randomly amplified DNA. These data provide a possible explanation for the large number of contigs in the PS assembly, as repeat reads typically result in gaps that an assembler cannot resolve.
In addition, approximately 10 Mb of PS reads were singletons (5.1% of the total reads), which had a significantly different GC content (42.5%) than the assembled reads (Tables S1 and S2). Singleton reads may originate from randomly amplified DNA, from Beggiatoa DNA sequences that amplify poorly, or from non-Beggiatoa DNA. They could represent a random sampling of trace contaminating DNA that has a potentially very large complexity but very low copy number per discrete contaminating genome. Although contamination cannot be completely ruled out, there was probably not enough non-Beggiatoa DNA present in the MDA reaction to yield sufficient read coverage depth to significantly affect the sequence assembly. Moreover, our analyses of nucleotide composition, single-copy genes, and 16S rRNA genes (see below) also do not support significant contribution of non-Beggiatoa DNA.
To identify potentially contaminating DNA sequences in our assembled data, all contigs of the PS dataset (7.6 Mb) were analyzed in a binning approach based on intrinsic DNA signatures. Relative abundance of dinucleotides, Markov-model-based statistical evaluations of tri- and tetramer over- and underrepresentation, and normalized chaos game representations for tri- and tetramers were investigated. This approach has been shown to enable a highly sensitive clustering of DNA sequences even among closely related gammaproteobacteria [35]. In the Beggiatoa PS dataset no outliers were identified that would indicate potentially contaminating DNA (data not shown).
Beggiatoa is a representative of the large sulfur-oxidizing bacteria that form a monophyletic cluster within the Gammaproteobacteria [36]. In both genomes we identified partial 16S rRNA gene sequences that were highly similar to sequences of marine Beggiatoa (Figure 2A). The gammaproteobacterial affiliation is supported by the phylogeny of a set of 41 concatenated proteins (Figure S2). Comparative sequence analysis revealed that the two Beggiatoa filaments (PS and SS datasets) are phylogenetically different despite their similar diameter of approximately 30 μm. This result is consistent with the potential genomic microdiversity among filaments indicated by the optical mapping results. The distinct phylogenetic origin is also reflected in the GC contents of both sequence datasets, which differ by 4% (Table 1). No additional 16S rRNA gene sequences were found.
Based on 16S rRNA sequence similarity, Nitrosococcus oceani and Methylococcus capsulatus are the closest relatives of Beggiatoa for which whole genome sequences are available. An analysis of the conserved ORFs for best BLAST hits against a local genome database was largely consistent with this affiliation (Table S3).
As both filaments are closely related at the 16S rRNA gene level, a large fraction of genes in both datasets were expected to be likewise highly similar. Therefore, the ORFs of the PS2 and SS2 datasets were compared for reciprocal best match (RBM) hits. In both datasets 378 ORFs mutually display the highest similarity (cut-off of e−05, 65% minimum sequence coverage). Because of only partially covered genes, many ORFs present in both genomes were not apparent despite showing their highest sequence similarities to the other sequenced Beggiatoa genome after manual reinvestigation. Thus, the observed number of ORFs with RBM hits constitutes only the minimum.
Interestingly, many ORFs showed their highest similarity to genes from the filamentous Nostoc sp. and gliding Anabaena variabilis. Furthermore, some gene fragments are exclusively shared with cyanobacteria, among them Nostoc sp., Gloeobacter violaceus, and A. variabilis. Most of these ORFs encode conserved hypothetical genes, of which many show similarities to putative transposases (e.g., BgP0160 and BgP1020ff), reverse transcriptase, and fdxN element excision controlling factor proteins. ORF BgP4037 encodes a conserved hypothetical protein (196 aa) with the highest sequence similarity (58%) to predicted proteins of Trichodesmium sp. (Figure S3A), of which at least 30 paralogs are present in the PS dataset. Moreover, BgP4037 co-localizes with “authentic” Beggiatoa genes such as nitrate reductase subunit genes (Figure S3B). The phylogenetic reconstructions of proteins containing either adenylation domains (AMP-A) or hemagglutinin domains (Figures S3 and S4; see below) confirm the hypothesis of horizontal gene transfer. Furthermore, contigs carrying cyanobacterial-like genes did not group in the cluster analysis, which indicates an already Beggiatoa-adapted codon usage pattern. In conclusion, these findings suggest extensive gene exchange between (filamentous) cyanobacteria and Beggiatoa. This apparent gene sharing is particularly interesting since Beggiatoa was formerly classified as a colorless cyanobacterium because of many shared phenotypic characteristics (for review see [17]).
To estimate the extent of putative contaminating DNA, in particular of cyanobacterial origin, we searched for duplicate genes that usually occur only once per prokaryotic genome. We identified 47 ribosomal proteins in the PS dataset that exclusively affiliated with Gammaproteobacteria (Table S4). The gammaproteobacterial affiliation is well confirmed by the phylogenetic reconstruction of a set of 41 concatenated proteins comprising 39 ribosomal proteins, recombinase A (recA), and RNA polymerase subunit B (Figure S2). Recently, a novel approach for the prediction of the number of genome equivalents in metagenomic samples was proposed [37] that is based on the occurrence of 35 widely conserved, single-copy marker genes present in most prokaryotic genomes. Out of these 35 we identified 30 genes (Table S5) in the PS dataset, none of which were found more than once. In addition, we found 40 genes of an extended set of 55 single-copy genes that are not as widely distributed (Table S6). Consistent with these findings 18 out of 24 amino-acyl tRNA synthetase genes were observed as single-copy genes in the PS dataset (Table S7). In conclusion, the single occurrence of proposed single-copy genes, ribosomal proteins, and amino-acyl tRNA synthetases is indicative of the presence of a single dominant genome in the assembled DNA sequence. Alternative phylogenetic markers such as recA, ATP synthase subunits, elongation factor Tu, RNA polymerase, and DNA gyrase AB were most similar to the Gammaproteobacteria based on BLASTP analysis. The only exception was a heat shock protein, Hsp70 (dnaK), that affiliated with Hsp70 of Firmicutes. However, it is known that Hsp70 genes are horizontally exchanged [38,39].
The genome size of Beggiatoa was estimated based on the ratio of single-copy marker genes, amino-acyl tRNA synthetase genes, and tRNA genes to their expected values. This suggests a genome coverage of more than 70% by the PS data, or a genome size of up to 11 Mb.
In 1888 Winogradsky [7] demonstrated the concept of chemolithotrophy studying a freshwater Beggiatoa. He showed that Beggiatoa gain electrons from oxidization of hydrogen sulfide to elemental, intracellularly stored sulfur and further to sulfate. However, the detailed pathways of sulfur species oxidation in these bacteria have not been elucidated.
Recent studies on nitrate-respiring Beggiatoa pointed to a two-step oxidation of sulfide [11,40]. In the anoxic zone sulfide is oxidized to elemental sulfur and sulfate at the expense of (stored) nitrate. Then Beggiatoa moves upwards into the oxic zone, where the stored elemental sulfur is further oxidized to sulfate using oxygen. When shuttling between sediment layers Beggiatoa experiences variable sulfide concentrations [41]. The initial oxidation of hydrogen sulfide to elemental sulfur is probably catalyzed via either of two alternative pathways: (1) a sulfide quinone oxidoreductase (Sqr) or (2) a flavocytochrome c/sulfide dehydrogenase (FccAB) (Figure 3A). Sqr is widespread among prokaryotes and appears to be critical for sulfide oxidation in Allochromatium vinosum [42]. FccAB was hypothesized to be more prevalent at low sulfide concentrations [43] and may be more important in the upper, oxidized sediment layers.
The genomes of both Beggiatoa filaments encode proteins of the “reverse dissimilatory sulfate reductase (rDsr) pathway” [44,45] (Figure 3A). We identified gene fragments encoding the cytoplasmic rDsrABC and also the membrane proteins DsrMKJOP that channel electrons to rDsrAB. Similar to in the betaproteobacterium Thiobacillus denitrificans [46], at least five paralogs of the DsrC-like subunit are present in the Beggiatoa genome (PS2). After formation of sulfite by DsrABC, it is oxidized and phosphorylized by an adenosin-phosphosulfate (APS) reductase to APS [47]. Finally, APS is dephosphorylized via an ATP sulfurylase to yield sulfate and ATP [47]. In Beggiatoa the AprAB is functionally linked to heterodisulfide reductases (HdrABC) that are likely responsible for electron transport to AprAB, as suggested for sulfate-reducing prokaryotes [48,49].
In Beggiatoa the oxidation of thiosulfate is catalyzed by the identified SoxABXYZ subunits of the Sox pathway [50]. However, so far all investigated organisms encoding the rDsr pathway lack the Sox(C)D subunits [51]. Simultaneously these organisms form sulfur globules while oxidizing reduced sulfur compounds. This is consistent with the observed sulfur globule formation and the missing SoxCD genes in Beggiatoa, but their presence in the unsequenced part of the genome cannot be excluded yet. In these organisms and most likely also in Beggiatoa rDsrAB is crucially involved in further oxidizing transiently stored elemental sulfur to sulfite [52]. Thus, the rDsr pathway is likely essential for Beggiatoa to perform an energetically more favorable two-step oxidation of sulfide and sulfur using nitrate and oxygen, respectively [11], when the zones of oxygen and sulfide do not overlap.
In organic-rich surface sediments oxygen is rapidly consumed. In typical Beggiatoa habitats oxygen penetrates only the upper few millimeters. Culturable Beggiatoa and their relatives commonly exhibit a negative chemotactic response to high oxygen concentrations [53], and preferentially oxidize inorganic sulfur compounds under microoxic conditions. The presence of high- and low-affinity terminal oxidases in both Beggiatoa datasets reflects the flexibility to respond to different oxygen regimes (Figure 3B). Under high oxygen concentrations a low-affinity cytochrome c aa3-oxidase is predicted to be used, whereas under microoxic conditions a high-affinity cytochrome c bb3-oxidase may be more prevalent. The differential expression of cytochrome oxidases under oxic and microoxic conditions has been reported for the freshwater relative B. leptomitiformis [54].
Vacuolated marine Beggiatoa and their relatives most likely respire nitrate under anoxic conditions [11,12,55]. The PS dataset encodes both membrane-bound (NarGH) and periplasmic (NapAB) nitrate reductases (Figure 3C). Because of the incomplete assembly, three non-overlapping fragments of a NarG gene were found (BgP3372, BgP5024, and sequences downstream of BgP4047) that were concatenated and phylogenetically affiliated with Proteobacteria (Figure S6). In addition to these proteobacterial NarGH, we surprisingly identified a second nitrate reductase, NarGH (BgP0139 and BgP4784), displaying by far the highest sequence similarities (NarG: 57% similarity at 98% coverage) to a putative nitrate reductase/nitrite oxidoreductase of the anaerobically ammonia-oxidizing planctomycete Kuenenia stuttgartiensis [56]. The phylogenetic reconstruction of both sequences revealed a novel lineage of putative nitrate reductases (Figure S6). However, nitrate reductases can also operate in the reverse direction in nitrite-oxidizing bacteria, where they are considered nitrite oxidoreductases (Nxr) [57]. Since there is physiological evidence for nitrite oxidation in K. stuttgartiensis with the NarG as candidate enzyme (M. Strous, personal communication), we speculate that Beggiatoa also utilize nitrite as an electron donor. In general, the function of NapAB (BgP1197ff) is unclear, but it may allow Beggiatoa to support nitrate respiration at low nitrate concentrations [58] or may enable Beggiatoa to respire nitrate even under aerobic conditions [59].
The preferred pathway of nitrate respiration in Beggiatoa and relatives and its regulations are of major ecological importance [60]. It is assumed that the main product of nitrate respiration in marine Beggiatoa and relatives is ammonia [16]. Although we could not identify the enzymes catalyzing the final reduction steps to ammonium ion or molecular nitrogen, they may be encoded on the not-yet-sequenced part of the genome. In Beggiatoa, a nitrite reductase (nirS; BgP1272) and two nitric oxide reductases (norB; BgP5178 and BgP3622) reduce nitrite and nitric oxide, respectively, to nitrous oxide (Figure 3C). To experimentally test the capability of Beggiatoa to denitrify, we measured nitrous oxide formation in acetylene-inhibited natural mats of nitrate-storing Beggiatoa in arctic marine sediments. The natural mat of Beggiatoa dissimilatorily reduced nitrate to nitrous oxide, while the adhering Beggiatoa-free sediment did not (Figure S7). In summary, the genomic and experimental data presented here provide a first clear indication of the significant denitrification potential of large marine sulfur bacteria.
The large, vacuolated Beggiatoa and relatives are unique among prokaryotes in their exceptional nitrate storage capabilities. They accumulate nitrate internally to high concentrations of up to 500 mM [16], which allows them to monopolize nitrate and therefore to outcompete other denitrifying bacteria [11]. The underlying physiological and genetic mechanisms of nitrate accumulation are still unknown. Plants store up to 50 mM nitrate in their vacuoles [61]. Here, the uptake of nitrate across the cytoplasmic membrane is usually driven by a transmembrane electrochemical gradient (Δp) followed by a transport of nitrate [62]. In plants, typically vacuolar-type H+-ATPases and H+-pyrophosphatases (HPPases) catalyze a proton translocation over endomembranes to generate a Δp for solute transport and likely also nitrate transport [63]. Vacuolar-type ATPases also occur in plasma membranes of some Archaea, but they are rarely encountered in Bacteria [64,65]. We propose that the accumulation of nitrate in Beggiatoa may be driven by a ΔpH generated by vacuolar-type ATPases and PPases. This energy is used by probable H+/Cl− exchanger-like proteins to exchange the accumulated protons in the vacuole and nitrate in the cytoplasm (Figure 4A). In support of this hypothesis we identified six of the nine putative subunits of vacuolar-type H+/Na+-translocating ATPase (atpABCDEI) (Figure 4A), which show their highest similarity to homologs in Nitrosococcus oceani, a related organism also containing intracellular membrane vesicles. Furthermore, a vacuolar H+-pyrophophatase (hppA) and an uncommon Ca2+-translocating ATPase were identified in the PS dataset that may also contribute to generation of a Δp/ΔpH (Figure 4A). To check for the presence of an electric potential (inside positive) over the vacuolar membrane, filaments were stained with fluorescent lipophilic cation rhodamine 123. The fact that rhodamine 123 was excluded from the vacuole of Beggiatoa cells is consistent with our hypothesis (Figure 4B). Considering the presumed ΔpH and the measured high nitrate concentrations in Beggiatoa, a corresponding acidic pH of the vacuole content similar to that observed in plants [66,67] would be predicted. In fact, preliminary pH measurements of the vacuole content of Beggiatoa sp. and Thiomargarita namibiensis (data not shown) give additional evidence of an acidic vacuole content. Nitrate accumulation in Arabidopsis thaliana vacuoles is mediated by a 2-NO3−/H+ antiporter (AtCLCa) that is similar to widely distributed H+/Cl− exchangers [66]. In the Beggiatoa genome we identified proteins (BgP0076 and BgP4800) related to H+/Cl− exchangers (clcA), and chloride channels that display weak similarities to the AtCLCa antiporter.
Flexibility in respiratory pathways is highly beneficial for organisms living under fluctuating environmental conditions such as can occur at sediment surfaces. As an alternative to nitrate and oxygen, Beggiatoa may also respire dimethyl sulfoxide (DMSO) to form the important anti-greenhouse gas dimethyl sulfide, as indicated by the presence of DMSO reductase genes (dmsABC) in the PS dataset. DMSO is frequently formed by eukaryotic plankton [68] and by photochemical oxidation of dimethyl sulfide [69]. Because DMSO is dissolved in sea water, Beggiatoa could access this alternative electron acceptor at the sediment surface. Additionally, the Beggiatoa genome encodes a thiosulfate reductase (phsABC), which is probably also involved in the reduction of elemental sulfur and tetrathionate [70]. Moreover, a thiosulfate reductase is also involved in disproportionation of thiosulfate [71], which is a significant intermediate in marine sulfur cycling [72]. The hypothesized inorganic sulfur reduction is in accordance with previous results in B. alba that have reported reduction of stored elemental sulfur under short-term anoxic conditions [73,74].
Apart from one strain, all freshwater Beggiatoa require organic substrates for growth, in contrast to autotrophic marine Beggiatoa [16]. In our Beggiatoa the ability to fix carbon dioxide for autotrophic growth is encoded as a form I ribulose-bisphosphate carboxylase oxygenase (RubisCO), first reported for a non-vacuolated strain [75]. In addition, a phosphoribulokinase and a carbonic anhydrase gene are predicted. However, the non-vacuolated B. alba and B. leptomitiformis also grow heterotrophically using acetate and other organic compounds [76–78]. Earlier studies on marine, non-vacuolated strains have shown a broad spectrum of utilized organic compounds [79]. Similarly, our data suggest that the large vacuolated Beggiatoa and their relatives are also not obligate lithoautotrophs. Both genomes harbor acetate/cation symporters, acetate kinase, and putative acetyl-coenzyme A synthetase to channel acetate into the general metabolism. Accordingly, in the related Thiomargarita, sulfur oxidation was stimulated upon acetate amendment [80]. During growth on acetate, the glyoxylate cycle is probably employed for gluconeogenesis, as observed in other Beggiatoa [54,77]. However, the key enzymes malate synthase and isocitrate lyase were not identified in the incomplete genomic sequences. Several enzymes of the tricarbonic acid cycle were identified, such as isocitrate and succinate dehydrogenase. In contrast to the free-living gammaproteobacterial sulfur-oxidizer Thiomicrospira crunogena [81], Beggiatoa encodes a 2-oxoglurate dehydrogenase and a malate dehydrogenase, whereas fumarate dehydratase, PEP carboxylase, and succinyl-coenzyme A synthase are possibly encoded on the unsequenced part of the genome. In general, these findings are consistent with experimental results [54] and suggest the presence of a complete set of tricarbonic acid cycle enzymes.
Furthermore, the presence of three subunits of a glycolate oxidase (glcDEF) suggests a utilization of glycolate, which originates from photosynthetic organisms, e.g., co-occurring cyanobacteria. The presence of genes encoding poly-β-hydroxybutyric acid synthase, acetyl-coenzyme A acetyltransferase, and acetoacetyl-coenzyme A reductase is consistent with the observation of large, visible granules of poly-β-hydroxybutyric acid in Beggiatoa and relatives [16]. The synthesis of polyglucoses in Beggiatoa has not been previously reported, but both genome datasets point to the capability to synthesize glycogen preferentially under oxic conditions, as in Thiomargarita [14], as illustrated by genes encoding glycogen synthase and glycogen-debranching enzymes. Beggiatoa could also synthesize ATP via substrate-level phosphorylation from pyruvate via a probable fermentative lactate dehydrogenase (ldh). Fermentation of storage compounds and pyruvate enables Beggiatoa to persist during periods of oxygen, sulfur, and nitrate depletion, e.g., when the oxic–anoxic interface is located above the sediment surface.
Under nutritional imbalance many bacteria accumulate phosphate, which is intracellularly stored as polyP. Thiomargarita and Thioploca exhibit an efficient phosphate uptake and storage system and contain large polyP granules. Recently, these organisms were hypothesized to account for large phosphorite deposits at the sea floor [14]. In Beggiatoa the ability for polyP storage has not been unambiguously proven [16]. Here, we provide genetic evidence for polyP storage in Beggiatoa. Interestingly, both Beggiatoa datasets encode phytases. Phytate is an important inorganic phosphate storage compound in plants and adsorbs to particles in sediments and soils. The phytases likely enable Beggiatoa to access inorganic phosphate more efficiently. In addition, Beggiatoa takes up polyP and orthophosphate via selective porins O/P and high-affinity phoBRU-regulated ABC phosphate transporters. After uptake, a polyP kinase catalyzes the synthesis of polyP granules. In analogy to phosphorus removal from activated sludge, Beggiatoa and relatives may accumulate polyP at the sediment surface under aerobic conditions and degrade polyP under anaerobic conditions at the depth where they uptake acetate [14] (Figure 5).
Unexpectedly, Beggiatoa appears to harbor the potential to synthesize secondary metabolites. We identified numerous genes of presumably cyanobacterial origin that encode non-ribosomal peptide synthetases and polyketide synthetases (PKS) (Table 2). Several functional domains are required for NRP and also for PK synthesis, respectively. Adenylation (AMP-A), acyltransferase (phosphopantetheine-binding), condensation, and thioesterase domains are present in the PS dataset (Table 2) and to a lesser extent in the SS dataset. The phylogenetic analysis of selected AMP A-type domains in Beggiatoa supports a mostly cyanobacterial origin of non-ribosomal peptide synthetases (Figure S4). The derived polypeptides show high similarities to proteins involved in synthesis of toxins and antibiotics rather than to fatty acid synthases. ORF BgP2814ff and downstream sequences (3,576 bp) display their highest similarities to anabaenopeptilide and nostopeptolide synthetases of Anabaena sp. and Nostoc sp., respectively, which are polyketide–non-ribosomal peptide hybrids of the microcystin family [82,83]. Other derived polypeptides of Beggiatoa (e.g., BgP5597 and BgP1194) exhibit significant similarities to modules of polyketide synthetases in Nostoc punctiforme. Since the presence of AMP-A domains in cyanobacteria is correlated with the synthesis of natural bio-active products [84], we hypothesize a similar capability to form secondary metabolites in Beggiatoa. These genetic findings have been corroborated by a HPLC-MS-based analysis of a methanol extract from a Beggiatoa mat from the sampling site that indicated a significant fraction of compounds of a molecular weight comparable to polyketides (S. Rachid, unpublished data).
We identified numerous ORFs that are homologous to large putative exoproteins, several of which contain a hemagglutination activity domain. Generally these glycoproteins are associated with cell adhesion and cell aggregation in biofilms of pathogenic bacteria [85]. Intriguingly, in Beggiatoa the derived proteins phylogenetically affiliate with the cyanobacterial genera Nostoc, Anabaena, and Trichodesmium, and also Hahella chejuensis, an exopolymer-producing gammaproteobacterium (Table S8; Figure S5). Similar to in cyanobacteria, several paralogs are encoded in the Beggiatoa genome, which may point to a functional relevance of the respective proteins. The striking similarity to filamentous, gliding cyanobacteria suggests a function of these proteins in gliding motility, and for sheath or filament formation. Indeed, glycoconjugates were recently detected in high amounts at the outer surface of Beggiatoa filaments using fluorescently labeled lectins (S. Hinck, unpublished data). Hence, the identified exocellular glycoproteins likely play a role in slime production, S-layer formation, or cell–cell adhesion.
We have shown that the combination of optical mapping, WGA, and pyrosequencing offers great potential for genomic analysis of individual, uncultured bacteria. However, the incomplete sequence assemblies limited the accurate determination of the genome size and an in-depth analysis of the Beggiatoa genome. Generally, the contribution of non-target DNA cannot be completely ruled out in environmental WGA projects; thus, polyphasic approaches are indispensable to test for the purity of the assembled sequences. Keeping these methodological issues in mind, the genomic analysis of single Beggiatoa filaments has generated numerous novel hypotheses with regard to their ecophysiology and evolution that can now be experimentally tested. Breadth of storage capabilities and a highly flexible energy metabolism, together with gliding motility, optimally equip these large marine Beggiatoa to thrive under spatially and temporally fluctuating conditions at sediment surfaces. The striking similarity between numerous genes of Beggiatoa and cyanobacteria, along with their obvious shared phenotypic characteristics, points to pronounced horizontal gene transfer between these organisms, likely facilitated by the long-term coexistence of Beggiatoa and cyanobacteria in surface sediments and microbial mats [86].
The Beggiatoa spp. filaments were obtained in Eckernförde Bay (Germany, Baltic Sea, 54° 47′ N/9° 83′ E). The surface of the Beggiatoa-covered sediment (∼4 m water depth) was sampled in August 2004 and December 2005 using polyacryl tubes. The sediment was kept in the dark at 4 °C until further processing. Two single Beggiatoa filaments with a diameter of 30 μm and length of ∼1 cm were transferred from the sediment surface to a Petri dish filled with artificial sea water medium containing agar. While gliding through the agar the Beggiatoa filaments were cleaned of particles and adhering bacteria.
Purified filaments of Beggiatoa were individually lysed as follows. A filament was placed in 27 μl of TE (10 mM Tris-HCl [pH 7.2], 1 mM Na2 EDTA) and subjected to ten alternating cycles of freezing/thawing in a dry ice–ethanol bath for 1 min and thawing at room temperature to enhance cell lysis. The DNA was denatured by the addition of 3 μl of KOH (0.4 M) and EDTA (10 mM). The lysate was incubated at 65 °C in a water bath for 3 min, and neutralized with 3 μl of Tris-HCl (pH 4) according to [21].
We employed MDA as a means of WGA to prepare sufficient DNA for genomic library construction and cloneless pyrosequencing. The REPLI-g kit (Qiagen; http://www.qiagen.com/) was used for MDA according to the manufacturer's instructions. Reactions contained 33 μl of the neutralized cell lysate and 25 μl of 4× MDA reaction mix, and were adjusted with water to a final volume of 100 μl. The reactions were incubated at 30 °C for 16 h and stopped by shifting to 65 °C for 3 min. The DNA concentration in the MDA product accumulated to a concentration of ∼1.4 mg/ml in all treatments.
MDA-amplified genomic DNA of one filament was sheared using a Hydroshear instrument (Genomic Solutions; http://www.genomicsolutions.com/) with speed code set to two for 30 cycles to yield DNA fragments of a size mainly between 4 and 6 kb. The gel-purified MDA products were then cloned into the pCR4 TOPO vector (Invitrogen; http://www.invitrogen.com/). The ligation products were used to transform TOP10 Escherichia coli using the pCR4 Blunt-TOPO vector cloning kit (Invitrogen) according to the manufacturer's instructions. Transformants were plated on 22-cm2 Q-trays (Genetix; http://www.genetix.com/) containing 100 μg/ml kanamycin. Kanamycin-resistant colonies were then picked using a Q-bot (Genetix) and arrayed in 96-well microtiter plates.
Plasmids for sequencing were robotically extracted from overnight cultures using a RevPrep Orbit (Genomic Solutions) or a Biomek FX Liquid Handling Robot (Beckman Coulter; http://www.beckmancoulter.com/). DNA sequencing setups, cycle sequencing, and sequencing reaction clean-ups were all performed using a Parallab Nanoliter Pipetting Robot (Parallab; http://www.parallab.uib.no/). The labeling reactions were performed in a volume of 50 nl using ABI BigDye Cycle Sequencing kits (Applied Biosystems; http://www.appliedbiosystems.com/), the thermal cycling was performed in an integral air cycler, and the clean-ups were conducted in capillaries using magnetic beads. The sequencing reactions were then loaded onto an ABI 3730xl DNA Analyzer (Applied Biosystems) for capillary electrophoretic separation and calling of the sequencing products. Both ends of each clone were sequenced using vector-based primers to provide mate-pair information. Approximately 8,800 sequence reads were obtained, of which 4,700 were usable for assembly.
The genomic DNA of a second, morphologically identical Beggiatoa filament was amplified using the MDA technique described above. The amplified DNA served as a template for sequencing using the clone-free pryosequencing technology developed by 454 Life Sciences (http://www.454.com/) [31]. Raw images from all regions of six-picotiter sequencing plates (one 60 × 60 and five 70 × 75) were processed with the three components (image processing, signal processing, and the Newbler de novo assembler) of the latest available version (1.0.51.03) of the 454 Life Sciences off-instrument data processing software to yield the PS assembly (Tables 1, S1, and S2). Additional sequencing was halted when the length of the all-contigs dataset did not increase with additional 454 Life Sciences sequencing runs, and we attribute the limited convergence of the length of the large-contig dataset (i.e., to the length of the all-contig dataset) to the unusually high percentage of repeat sequences in the MDA reaction product used for pyrosequencing. A subset (0.9 Mb; 3,448 fragments; length range 81–643 bases, each supported by at least ten reads) of the 22,858 small contigs produced by the 454 Life Sciences assembler was added to the large-contig dataset the assembler produced (6.7 Mb; 3,321 contigs, each >500 bases) to yield the 7.6-Mb PS assembly (Figure S8).
For optical genome mapping, five Beggiatoa filaments 35 μm in diameter and >1 cm in length were purified as described above and immediately transferred into an agar drop containing cell suspension buffer (10 mM Tris-HCl [pH 7.2], 20 mM NaCl, 100 mM EDTA, 5 mg/ml freshly prepared lysozyme, 1% LMP agarose kept at 70 °C). After solidification at 4 °C the agar drop was incubated in cell lysis buffer (0.5 M EDTA, 1% laurosyl sarcosine, 2 mg/ml proteinase K [pH 9.5]) at 50 °C overnight. The determination of the chromosome size was performed as reported earlier [32].
The DNA sequence data of the PS and SS approaches were each divided into two sub-databases. These sub-databases were used for the analysis of scaffolds of length <2 kb (PS1 and SS1) since ORF prediction on short fragments is not possible with standard ORF-finding tools, because of missing information. All scaffolds in these sub-databases were translated into all six reading frames and treated as artificial ORFs in the ongoing analysis to perform similarity searches. The second set of sub-databases consisted of all sequenced scaffolds longer than 2 kb for each approach (PS2 and SS2). All scaffolds in these databases were used for ORF prediction using the metagene prediction software MORFind (J. Waldmann and H. Teeling, unpublished data) developed at the Max Planck Institute for Marine Microbiology, Bremen. This system analyzes and combines the output of the three commonly used gene finders CRITICA, GLIMMER, and ZCURVE to enhance sensitivity and specificity. To resolve conflicts, an iterative post-processing algorithm is used, taking into account signal peptide and transmembrane predictions, ORF length, and the number of gene finders by which an ORF has been predicted.
Annotation was performed by a refined version of the GenDB v2.2 system [87], supplemented by the comparative analysis tool JCoast (http://www.megx.net/jcoast/) developed at the Max Planck Institute for Marine Microbiology, Bremen. For each predicted ORF the system retrieves observations from similarity searches against sequence databases NCBI-nr, Swiss-Prot, and KEGG GENES (release April 2006) and protein family databases Pfam (release 20.0) and InterPro (release 12.0, InterProScan v4.2), and from predictive signal peptide analysis (SignalP v3.0 [88]) and transmembrane helix analysis (TMHMM v2.0 [89]). tRNA genes were identified using tRNAScan-SE [90]. Predicted protein coding sequences were automatically annotated with the software MicHanThi [91] developed at the Max Planck Institute for Marine Microbiology, Bremen. The system simulates the reasoning in the human annotation process using fuzzy logic. The annotations of all ORFs described in this publication were manually refined.
To evaluate the phylogenetic consistency of the conserved ORFs in the databases PS2 and SSI2, all conserved ORFs were tested by BLAST analysis for the phylogenetic distribution of best hits against a local genome database (genomesDB; M. Richter, unpublished data). Only hits with an e-value below e−05 were considered significant. The local genome database (genomesDB) provides a computationally well-defined environment of 311 published whole genome sequences of bacterial and archaeal origin, with all ORFs of each genome carrying a unique ID. To allow genome comparisons between specific user-defined groups, all ORFs are assigned to the respective organism and metabolic group. In contrast to the general purpose database NCBI-nr, which contains every sequence ever submitted, the focus of genomesDB is the association of every protein to their phylogenetic affiliation in a refined environment.
For all sequences of the PS dataset the following intrinsic DNA signatures were calculated: (1) dinucleotide relative abundances [92], (2) Markov-model-based statistical evaluations of tri- and tetramer over- and underrepresentation [93], and (3) normalized chaos game representations for tri- and tetramers [94]. Values for (2) and (3) were computed by ocount and cgr, respectively, two self-written C-programs that are publicly available ( http://www.megx.net/tetra_new/html/download.html). The self-written Java program MetaClust [95] was used to automatically trigger the individual calculations and subsequently store them in a MySQL database. After that, MetaClust was also applied to build different combinations of subsets of the individual methods for all sequences exceeding 5 kb and trigger a hierarchical clustering of them using Cluster 3.0 [96]. For the clustering, complete linkage was used as the clustering algorithm, and the Euclidean distance was used as the distance measure. The corresponding result files were analyzed using Java TreeView (http://jtreeview.sourceforge.net/) and checked for outliers. This procedure was repeated for all sequences exceeding 4 kb, 3 kb, 2 kb, and 1kb and for all sequences of the dataset.
To compare the two datasets for shared genes we performed a “BLAST all against all” analysis between all predicted ORFs in the datasets PS2 and SS2. RBMs were counted only if the e-value was below the cut-off of e−05.
All phylogenetic analyses were performed with the ARB/Silva software package ([97]; http://www.arb-silva.de/). The partial 16S rRNA gene sequences were inserted into a phylogenetic tree based on nearly complete sequences. The alignment was corrected manually. Phylogenetic trees were calculated by maximum parsimony, neighbor joining, and maximum likelihood analysis with different sets of filters. Topologies were evaluated to elaborate a consensus tree. Branching orders that were not supported by all methods are shown as multifurcations. Subsequently, partial sequences were inserted into the reconstructed tree by applying the parsimony criteria without allowing changes in the overall tree topology. Multiple alignments of protein sequences of nitrate reductase alpha subunits (NarG), AMP domains of non-ribosomal peptide synthetases, hemagglutination-domain-containing proteins (Hgg) were established with the ClustalW program package using the BLOSUM62 substitution matrix. For the phylogenetic analysis of NarG and Hgg maximum likelihood trees (Molphy, http://plone.jcu.edu.au/hpc/software-installation/molphy) were reconstructed using JTT amino acid substitution matrix for evolutionary distance. Distance matrix trees were calculated using the neighbor joining function of ARB with the Kimura correction for proteins. Different base frequency filters were applied. For phylogenetic reconstruction of AMP-A domains of non-ribosomal peptide synthetases, nearly full-length sequences were extracted. Maximum parsimony, neighbor joining, and PHYLIP distance matrix trees were calculated using different correction factors (see above). For calculations, 219 amino acid positions were considered, excluding major deletions and insertions. A set of 41 concatenated protein sequences were considered to determine the phylogenetic position of Beggiatoa. The following protein sequences were used for maximum parsimony, neighbor joining, and maximum likelihood trees: RNA polymerase (rpoC), recA, and ribosomal proteins L1–L5, L7/L12, L9–L11, L13–L24, L27–L29, L35, S2–S8, S11–S13, and S15–S20. A 30% positional conservation filter was used (5,857 positions) to exclude variable positions.
Single Beggiatoa filaments were incubated for 40 s in filter-sterilized seawater containing 200 μM of the lipophilic cation rhodamine 123 (Molecular Probes http://probes.invitrogen.com/). After loading, filaments were thoroughly washed with seawater, placed in an incubation chamber, and mounted on the stage of an Oz confocal microscope (Noran Instruments, http://www.thermo.com/). The light from an argon ion laser (488 nm; Omnichrome, http://www.mellesgriot.com/) was delivered to the cells via a 40× oil immersion plan apochromat objective (NA 1.4; Nikon Instruments, http://www.nikoninstruments.com/). Fluorescence emission light was directed through a 500-nm LP barrier filter (Chroma Technology, http://www.chroma.com/) and quantified using a photomultiplier tube at eight-bit resolution (Hamamatsu Photonics, http://www.hamamatsu.com/). Hardware and image acquisition were controlled by Intervision software (v1.5; Noran Instruments) running under IRIX 6.2 on an Indy workstation (SGI, http://www.sgi.com/). Images (512 × 480 pixels) were collected at 30 Hz with a pixel dwell time of 100 ns and averaged using a window of 32 ns in real time.
This whole genome shotgun project has been deposited at DNA Data Bank of Japan (http://www.ddbj.nig.ac.jp/), the EMBL Nucleotide Sequence Database (http://www.ebi.ac.uk/embl/), and GenBank (http://www.ncbi.nlm.nih.gov/Genbank/) under the project accessions ABBY00000000 (Beggiatoa sp. SS dataset) and ABBZ00000000 (Beggiatoa sp. PS dataset), respectively. The version described in this paper is the first version.
|
10.1371/journal.pcbi.1006180 | Atomic resolution mechanism of ligand binding to a solvent inaccessible cavity in T4 lysozyme | Ligand binding sites in proteins are often localized to deeply buried cavities, inaccessible to bulk solvent. Yet, in many cases binding of cognate ligands occurs rapidly. An intriguing system is presented by the L99A cavity mutant of T4 Lysozyme (T4L L99A) that rapidly binds benzene (~106 M-1s-1). Although the protein has long served as a model system for protein thermodynamics and crystal structures of both free and benzene-bound T4L L99A are available, the kinetic pathways by which benzene reaches its solvent-inaccessible binding cavity remain elusive. The current work, using extensive molecular dynamics simulation, achieves this by capturing the complete process of spontaneous recognition of benzene by T4L L99A at atomistic resolution. A series of multi-microsecond unbiased molecular dynamics simulation trajectories unequivocally reveal how benzene, starting in bulk solvent, diffuses to the protein and spontaneously reaches the solvent inaccessible cavity of T4L L99A. The simulated and high-resolution X-ray derived bound structures are in excellent agreement. A robust four-state Markov model, developed using cumulative 60 μs trajectories, identifies and quantifies multiple ligand binding pathways with low activation barriers. Interestingly, none of these identified binding pathways required large conformational changes for ligand access to the buried cavity. Rather, these involve transient but crucial opening of a channel to the cavity via subtle displacements in the positions of key helices (helix4/helix6, helix7/helix9) leading to rapid binding. Free energy simulations further elucidate that these channel-opening events would have been unfavorable in wild type T4L. Taken together and via integrating with results from experiments, these simulations provide unprecedented mechanistic insights into the complete ligand recognition process in a buried cavity. By illustrating the power of subtle helix movements in opening up multiple pathways for ligand access, this work offers an alternate view of ligand recognition in a solvent-inaccessible cavity, contrary to the common perception of a single dominant pathway for ligand binding.
| Proteins often bind ligands in buried cavities that appear to be inaccessible based on static structures. The mechanisms and pathways by which ligands reach their binding sites in such cases are, thus, often unknown. Yet, ligand recognition by occluded cavities can happen rapidly. A central question remains: How does such a process occur? Experiments that provide insight at atomic resolution are currently lacking. In the current work, we have used a computational approach to capture the process by which a ligand, benzene, binds to a buried cavity in the L99A cavity mutant of T4 Lysozyme. Using multiple long, unbiased atomistic simulations, we have discovered how benzene, starting from bulk solvent, finds and binds the solvent-inaccessible cavity. We find that there is no single dominant pathway. Rather, simulated trajectories discover multiple binding pathways with low activation barriers, facilitating a rapid recognition process. We highlight the role of subtle movements in helix positions in opening up multiple crucial paths for benzene to reach its binding cavity without the need for large-scale distortions of the protein structure, explaining the small activation energies.
| Measurements of affinities of small molecules to specific binding sites in proteins have become routine, especially in the context of drug discovery studies where such experiments are often among the first to be performed. These thermodynamic measurements are routinely supplemented by kinetic studies of drug-receptor interactions [1–3]. However, in contrast to these standard kinetic and thermodynamic measurements, experiments that provide atomic level insights into the kinetic pathways by which small cognate molecules find their binding sites and the roles that protein conformational dynamics play in this process are lacking. This is particularly the case for ligand binding sites that are deeply buried in the receptor protein core, precluding access, in the absence of structural rearrangements, even for bulk solvent molecules. Yet, many ligands recognize these deeply buried, solvent-inaccessible cavities very efficiently, and often at rates that approach those which are diffusion limited. Effective methodologies for the study of such binding processes in atomistic detail would have significant implications for the rational design of pharmaceuticals at a practical level and would provide important insights into molecular recognition and the role of protein dynamics in ligand binding, in general.
The current work provides an atomistic view of the complete kinetic processes by which a hydrophobic ligand binds rapidly to the occluded and solvent-inaccessible cavity of a well-known system, namely the L99A mutant of lysozyme from the T4 bacteriophage (T4L L99A). Substitution of Ala for Leu at position 99 results in a 150 Å3 cavity [5] that can accommodate a range of ligands, including benzene, (Fig 1) [4–7]. Over the past several decades this protein has been used as a model system for understanding how buried cavities affect protein stability and structure, in general, and how ligands might navigate the protein landscape to bind rapidly to their proper sites. To this end, high-resolution X-ray structures of both the apo form of the protein and the benzene-bound complex have been solved by Mathews and coworkers [4,5] and pioneering studies have been undertaken that relate cavity size to protein stability [6]. NMR measurements have shown that the binding of benzene to T4L L99A occurs with an on rate constant of approximately 106 M−1s−1 [8]. While certainly slower than the diffusion-limited rate (~109 M−1s−1), binding is more rapid than what is generally perceived for ligand-recognition in solvent-inaccessible buried protein cavities. Additional NMR studies, based on relaxation dispersion approaches [9], established that T4L L99A exchanges in solution between a major conformer similar to the crystal structure of the protein and a minor state that is populated to 3% with a millisecond lifetime at room temperature [10,11]. However, the structure of this rare state shows that the cavity is occluded by the aromatic side-chain of Phe114 that prevents the binding of benzene so that the dynamic process that has been characterized is not relevant to ligand binding [11,12]. Oxygen and xenon binding studies have also been performed that have helped to identify hydrophobic cavities in the protein [13–15]. Further, this system has served as a prototype for many binding free energy calculations [16,17].
Despite the availability of macroscopic kinetic rate constants for aromatic ligand binding, high resolution structures of the bound and free forms of T4L L99A, and NMR spin relaxation data that indicate conformational flexibility in the region surrounding the ligand binding site, a detailed description of how ligands bind to the protein remains elusive. Recent innovations in experimental and computational techniques, notwithstanding, the mechanism by which ligands bind to buried sites in a protein has remained difficult to address, because the process often involves the formation of transient metastable states that are challenging to characterize at high resolution and because the conformational fluctuations leading to binding can be rapid, complicating detailed characterization by most biophysical techniques. To this end, molecular dynamics simulations are emerging as a complementary tool to experiment for the study of molecular recognition processes because of recent innovations in GPU-based technologies, access to distributed computing, and the development of special-purpose computers [18–21], that have facilitated studies of many pertinent biochemical processes at atomic resolution.
In this work, we elucidate the mechanism by which benzene binds to the solvent-inaccessible cavity of T4L L99A by capturing the entire binding process using unbiased molecular dynamics (MD) simulations. Despite the fact that the simulations contained no a priori knowledge of the T4L L99A binding site, the resulting bound conformations that were obtained in a series of microsecond-long simulations and independent trajectories matched that of the crystal structure. Underlying Markov state model analyses [22–24] of our simulation trajectories established multiple binding pathways, each with transient opening of a channel involving a distinct pair of helices in the C-terminal domain of T4L L99A. Using an enhanced sampling technique, infrequently biased metadynamics simulations [25,26], the ligand unbinding pathways from the bound state have also been established. The resulting estimates for ligand binding on- and off-rates as well as binding free energies are in reasonable agreement with the experimentally reported values, lending support to the conclusions of this study. Our work establishes that the fast binding of benzene to T4L L99A derives from its access to multiple pathways that are formed from low barrier concerted protein fluctuations that only occur in the mutated protein. It also establishes the utility of molecular dynamics simulations, in general, in providing detailed descriptions of binding processes that are subsequently validated by comparison of the resulting calculated kinetic and thermodynamic parameters with those obtained via experiment.
We performed six independent all-atom unbiased MD simulations with lengths varying between 2 and 8 μs, resulting in a total simulation time of 29 μs. In each of the simulations a benzene molecule, starting from a random position in solution, was correctly placed in the target-binding site, corresponding to the hydrophobic and solvent-inaccessible cavity created by the L99A mutation in T4L. Ligands were initially positioned at least 4 nm away from the binding pocket in random orientations. As shown in three representative trajectories (Fig 2 and S1–S3 Movies), the ligand diffused extensively in the solvent, occasionally contacting different parts of the protein surface, before entering the binding pocket. Fig 2A quantifies the time-evolution of pocket-ligand distances en route to binding in three representative binding trajectories, where the distance between the binding cavity and benzene gradually decreases as the ligand finds its target. As depicted in Fig 2B, the MD derived structure converges to within 1–2 Å of the X-ray based model for the holo-form of the protein [4], as quantified by the root-mean-squared deviation (RMSD) of heavy atoms from residues defining the cavity and including the ligand (detailed in Materials and methods). The excellent agreement between the simulated and crystallographic binding regions, including the orientation of benzene (Fig 2C), provides confidence in the underlying force field used to model the T4L L99A benzene binding process. Unlike standard docking approaches which search for the best ligand orientation within a predefined binding site, our long and unguided atomistic MD simulations reproducibly identified and maintained the correct ligand-bound pose without user intervention or incorporation of any prior knowledge of the binding site. As shown in Fig A in S1 Supporting Information in each of the trajectories substantial dehydration accompanies benzene entry into the hydrophobic and solvent-inaccessible binding pocket.
One of the key findings of the current work is that there are multiple distinct pathways by which benzene binds the solvent-inaccessible cavity in T4L L99A. Contrary to the belief of a single dominant pathway in protein-ligand binding events, as exemplified by Shaw and coworkers in studies of GPCRs and kinases [18,19], our post simulation analyses of the six successful binding trajectories specifically identified three distinct pathways, each involving opening of channels between pairs of helices near the C-terminus of the protein. Fig 3A and S1–S3 Movies illustrate the three representative binding pathways via the time evolution of a benzene molecule as it reaches the binding pocket. The three pathways involve crucial conformational fluctuations of the protein and specifically, creation of crevices across certain inter-helical interfaces in the C-terminal domain. Notably, these helices were different in each of the three identified pathways as illustrated in Fig 3. For example, in trajectory 1, helices 4 and 6 temporarily move away from one another, creating a pathway to the binding site as seen in the snapshot at 7.388 μs where Ala 99 colored in orange is visible from the surface (Fig 3B). This is in contrast to the snapshot at 7.375 μs where the distance between the two helices is considerably shorter, prohibiting the entry of benzene. In trajectory 2 (Fig 3C) benzene goes through a path created between helices 7 and 9. Benzene first associates with the surface as seen in the snapshot at 2.08875 μs. The subsequent transient displacement between helices 7 and 9 creates a tunnel through which benzene enters, as shown in the snapshot at 2.090 μs. After residing at an alternate site for ~2 ns the ligand reaches the final binding site (Fig 3C). In trajectory 3, benzene enters through an opening created at the junction between helices 5, 6, 7 and 8 (compare snapshots at 3.874 and 3.876 μs, Fig 3D). After a period of approximately 10 ns the ligand localizes to the correct binding site, as illustrated in Fig 3D.
In combination, these results show that key conformational fluctuations of T4L L99A lead to the formation of transient conformers where the distance between sets of helices becomes sufficiently large for benzene to reach the solvent-inaccessible cavity. As depicted in Fig B in S1 Supporting Information, the spatial density profile of benzene around the protein, obtained by combining the long trajectories, provides a cumulative picture of the different pathways that lead to binding. The localization of benzene near the cavity and near the C-terminal domain helices is evident from the time-averaged density profile. However, we also observe certain locations near the N terminus where the ligand resides for significant periods of time (Fig B in S1 Supporting Information). Whether these highly visited locations are ultimately important for ligand binding is not clear.
We have used Markov State Model (MSM)-based analysis[24] of the MD trajectories to obtain mechanistic insights into the ligand binding process and to calculate binding on- and off-rates and thermodynamic parameters. Prior to building a meaningful Markov state model, an additional three hundred, 100 ns trajectories were obtained by initiating independent simulations from different intermediates in the long trajectories. Analysis of the cumulative 59 μs of simulated data (29 μs from long trajectories and another 30 μs from 300 short trajectories, see Materials and methods) using the MSM approach yielded estimates of the thermodynamic and kinetic parameters for the binding process which are in reasonable agreement with experiment (Table 1). The calculated kinetic on and off rate constants (kon and koff,) obtained by MSM-derived mean first passage time (MFPT) values (see caption of Table 1 and Material and methods for equations), are, respectively, kon = 21 ± 9 x 106 M-1 s-1 (MFPTon = 5163 ns) and koff = 311 ± 130 s-1 (MFPToff = 3.2 x 106 ns). As compared in Table 1, the MSM-derived kon is larger than that measured experimentally, kexpon = 0.8–1 x 106 M-1 s-1. We also note that the kon value calculated from the average binding times from six long unbiased multi-microsecond MD trajectories yields an on-rate constant of 25 x 106 M-1 s-1, very close to the MSM-derived kon rate (21 ± 9 x 106 M-1 s-1). This implies that the underlying deviation of the simulated kon value from the experimental kexpon rate does not result from MSM analysis but is due to limitations in the force field. For example, the TIP3P [27] water model used in our MD simulations leads to water diffusion rates that are two to three times higher than those based on experiment. The ligand unbinding rate constant, i.e. the so-called off-rate constant koff, computed using both MSM (koff = 310 ± 130 s-1) and Metadynamics approaches (koff = 270 ± 100 s-1) (detailed later) is in reasonable agreement with experiment (kexpoff = 950 s-1). In this context, we note that there are several reports highlighting systematic deviations of MSM-derived off-rate constants from those determined experimentally [20,21]. However, it should also be emphasized that the MD estimate of koff is less precise than for kon because spontaneous unbinding events were not observed. Finally, we have also computed the standard binding free energy (ΔG0binding) of benzene to T4L L99A using the MSM-derived stationary population of the unbound and bound conformations (see Table 2). The computed standard binding free energy of -6.9 ± 0.8 kcal/mol predicts slightly higher binding affinity than the experimentally measured ΔG0binding = -5.2 ± 0.1 kcal/mol.
A simple four state MSM was constructed from our simulated data, as described in Materials and Methods. The four macrostates are illustrated in Fig 4, where MS0 and MS3 are identified with the unbound solvated benzene state and the final bound conformation, respectively, and MS1 and MS2 are two intermediates, with benzene localized near different entry points. Populations and lifetimes of each of the states, along with the committor values that measure the progress of the binding process, are summarized in Table 2. As shown in Table 2, the stationary population of the bound state, MS3, is highest and as described above, the binding free energy of -6.9 ± 0.8 kcal/mol, derived from the stationary state populations, is in reasonable agreement with the experimental measurement. The location of the benzene ligand is distinct in the intermediate structures, MS1 and MS2, highlighting again the different pathways of entry (see above). Notably, benzene is bound to different positions of the protein in the MS1 state, potentially interconverting rapidly between the different sites. These include locations near a pair of entry points that are comprised of helices 7 and 9 or the junction between helices 5, 6, 7 and 8 that place the ligand close to the cavity. The positions of benzene in MS1 are, thus, as found in MD trajectories 2 and 3 (Fig 3C and 3D), so that MS1 is a crucial intermediate. In contrast, benzene is further from the cavity in MS2, positioned instead on the surface of helices 4 and 6. MS2 is thus an intermediate of trajectory 1 that is elucidated from MD simulations (Fig 3B). The computed committor probabilities from transition-path-theory based analysis suggest a higher (0.8) commitment of MS1 towards the bound state than that of MS2 (0.2) (Table 2). The conversion from unbound (MS0) to bound (MS3) macrostates proceeds via multiple pathways involving states MS1 and MS2. Transition path theory was used to identify paths connecting unbound and bound states and to calculate the flux through the different pathways. The kinetic pathways are illustrated in Fig 4. The contributions of the four binding pathways [MS0 → MS1 → MS3], [MS0 → MS3], [MS0 → MS2 → MS1 → MS3] and [MS0 → MS2 → MS3] to the total flux from MS0 to MS3 are 52%, 39%, 5% and 4% respectively, showing that there is no single dominant pathway. The direct conversion MS0 → MS3 reflects the very rapid transformation of ligand from unbound to bound macrostates with no detectable intermediates. For the MS0 → MS3 path there have been occurrences where the ligand traverses rapidly through the helix7/helix9 or helix4/helix6 gateways or via the helix5/helix6/helix7/helix8 interface. The direct transitions from unbound to bound conformations thus often involve ligand passage via helix gateways. Out of four pathways derived by the MSM analysis two are dominant, involving MS0, MS1 and MS3, and jointly account for 90% of the flux. However, long simulation trajectories also showed a distinct ligand-binding pathway via MS2, even though its contribution to the overall flux is less. The observation of multiple pathways for this system argues that binding of ligands to proteins in general is likely to be more complex than via a single pathway which is often the model used in the analysis of binding data. As a final note, the [MS0 → MS2 → MS3] path can be identified with trajectory 1 (S1 Movie), while the [MS0 → MS1 → MS3] pathway is a composite of trajectories 2 and 3 (S2 and S3 Movies).
The fact that benzene is able to rapidly bind to T4L L99A is consistent with a low activation barrier. The activation free energy ΔG* is not readily available from experimental data, as the temperature dependencies of measured rate constants provide only an estimate of the activation enthalpy ΔH* [28]. By assuming a diffusion controlled ligand-protein binding rate of ~109 s-1M-1, and comparing this value with the experimentally measured kon ~ 106 s-1M-1 rate, Feher et. al. estimated the free energy barrier for binding to be relatively small, 4–5 kcal/mol (6.5 to 8.5 RT) [8]. Consequently, the barrier for the unbinding process, ΔGunbinding*, can be calculated to be ~8–9 kcal/mol (13–15 RT), based on ΔG0binding = -4.2 kcal/mol (Table 1). The MD simulations are also consistent with a small barrier as the observed binding rates are faster than the experimentally measured ones, even after accounting for the faster diffusion constant of the TIP3P-modeled water used here. The unbinding activation barrier can be estimated from the MD simulations using the relation koff=(1/τTPT)e-ΔGunbinding*/RT [29], where the transition path time τTPT is the time required for benzene to transition from the cavity to the unbound form. Note that τTPT is obtained from simulation while koff can be measured experimentally or estimated by simulation. Values of τTPT have been quantified directly from the six binding trajectories. In this approach, a benzene molecule was considered to be bound to T4L L99A if any of its carbon atoms was less than 0.38 nm from the Ala 99 Cβ position and, conversely, unbound when every carbon atom was greater than 0.6 nm from the Ala 99 Cβ. Based on this criterion, τTPT values for the six trajectories vary from 0.2 to 24 ns with an average of 7 ns. Using this average value along with an experimental value for koff of 0.95x103 s-1 [8], ΔGunbinding* = 11.9 RT was estimated, in agreement with the barrier obtained from the analysis of Feher et al. The small value of ΔGbinding* ~5 RT, that is estimated from ΔGbinding=ΔGbinding*-ΔGunbinding*, further emphasizes the small activation barrier to binding. In the above analysis barriers were estimated assuming a single binding pathway but the conclusion regarding small barriers will hold for the multiple paths as each of them occurred spontaneously during the simulations.
The MD trajectories described above provide insights into the origin of the small activation barrier for ligand binding. In all three binding routes identified in this study (Fig 3) the ligand reaches the cavity through the transient formation of pathways that are linked to the displacement of specific helices by 2–3 Å, without large scale changes in secondary structure or unfolding. We hypothesized that the L99A mutation not only creates the binding cavity but might also increase the flexibility of the cavity-containing C terminal domain in general, leading to an increased probability for the adoption of conformations that allow ligand binding. To test this hypothesis we calculated backbone order parameters squared, S2, from chemical shifts for both T4L L99A and T4L WT (referred to as T4L WT* in what follows, see Material and methods) [30]. S2 values report on the amplitudes of ps-ns motions of amide bond vectors [31], range between 0 (isotropic motion)– 1 (rigid), and are routinely used to identify flexible regions of proteins [32,33]. The similar S2 vs residue profiles for both T4L L99A and WT* (Fig 5A) indicate that fast timescale backbone dynamics are not affected by the L99A mutation, while the high S2 values (0.8 to 0.9) show that both forms of T4L are rigid and that rapid backbone motions are not linked directly to ligand binding. In order to ascertain whether the L99A mutation increases the population of transiently formed high energy conformers that would allow benzene to move past helices and enter the cavity we have computed free energy profiles for both T4L L99A and WT* as a function of the distances between helices 4 and 6 (Fig 5B) and between helices 7 and 9 (Fig 5C). Interestingly, we observe that the free energy penalty for larger inter-helical distances in T4L WT* is greater than for T4L L99A, and thus the conformations that facilitate binding are less probable for T4L WT*. For example, the probability of a Helix 4/6 separation of 1.30 nm, that is required for benzene entry (Fig 2B and 2C) increases by ∼40 fold in the L99A mutant, as it is 2.21 kcal/mol more unfavorable to separate this helix pair by 1.30 nm for the WT* sequence (eΔG/RT = 39 for ΔG = 2.21 kcal/mol). It is thus unlikely that a molecule of benzene could ‘squeeze’ between the helices in T4L WT* before reaching the occluded binding site due to the steric effect of residue L99. As the helices forming the C terminal domain of T4L are packed against one another, creating openings between helices that are required for binding (Fig 3) results in subtle positional changes to the other helices as well. A closer inspection of the trajectories showed that movement of helices 4 and 6 or 7 and 9 that is required to accommodate binding reduces the distance between side-chains of Ile78 and Ala99. A similar change in these inter-helical distances in the WT* protein would lead to a steric clash between Ile78 and the larger Leu99, as illustrated in Fig 5D in the case of Trajectory 1 where the helix 4-helix 6 distance transiently increases for only a few nanoseconds. We would like to emphasize that the population of the conformer where helices 4 and 6 are separated by 1.30 nm is extremely low, even for T4L L99A, (e−ΔG/RT ∼0.0046% with ΔG = 3.23 kcal/mol, Fig 5B) relative to the equilibrium conformer corresponding to a helix4-helix6 distance of ~1 nm. Thus, this ligand-accessible state cannot be quantified through experiment, further emphasizing the importance of large scale MD simulations like those performed here. A similar scenario is also found for the separation of helices 7 and 9 (Trajectory 2, Fig 3) where increasing the inter-helical distance from 1.05 nm to 1.25 nm, that is required for ligand entry, is more unfavorable by 5.9 kcal/mol for WT* than for T4L L99A.
The computational expense of unbiased binding simulations prohibits the exhaustive survey of all of the possible kinetic pathways. Hence we explored a complementary approach of validating the identified pathways and of enriching the ensemble of binding trajectories using a recent implementation of metadynamics simulation [25] to enable the extraction of the kinetics of benzene unbinding from the cavity. Using the radial distance between the binding pocket and the ligand as the collective variable, this simulation technique infrequently deposits repulsive, history-dependent Gaussian bias along the pocket-ligand distance so as to efficiently accelerate the unbinding process. The corresponding acceleration factor, multiplied by the simulation time, provides an estimate of the time for unbinding. As depicted in Fig D in S1 Supporting Information, the p-value analysis [26], as suggested by Tiwary and Parrinello, provided a good Poisson fit with a high p-value of 0.73, when averaged over numerous independent trajectories. The computed off rate, 369 s−1, is in reasonable agreement with the experimentally measured off-rate, Table 1, showing that the ligand-pocket distance is a relevant collective variable for exploring benzene-T4L L99A unbinding pathways. Significantly, as depicted in Fig D in S1 Supporting Information, the unbinding trajectories revealed two unbinding pathways of the ligand from the pocket. These involved benzene exiting through small openings between helices 7 and 9 and the juncture of helices 5, 6 7 and 8, pathways that are the reverse of the subset of those previously identified for binding.
Although ligand binding to receptors is critically important for many biological processes an atomistic understanding of how this might occur in the context of occluded binding sites is lacking. This reflects the fact that while the main biophysical tools that are used to study molecular interactions are powerful for characterizing and providing atomic resolution structures of the end points of the binding process (free and bound states) they are often much less robust in generating a description of the binding mechanism. Improvements in both computer technology and MD simulation methodology have made it possible to obtain μs-ms MD trajectories of proteins in explicit solvent so that it is possible to study processes such as ligand binding, conformational exchange, and protein folding using MD simulations [12,18,19,34]. Here, using unbiased MD simulations, we have addressed the longstanding question of how hydrophobic molecules rapidly reach buried cavities in proteins by working with a model system in which benzene binds to a 150 Å3 cavity mutant in T4L L99A. Central to every MD-based approach is that the force field used should be able to model the underlying free energy surface accurately. We chose to use the CHARMM force field [35] as previous MD studies had shown that the Phe114 buried to exposed conformational transition in T4L L99A and in the related protein, T4L L99A/G113A/R119P, is well modeled with it [12,36]. The structures of the benzene bound conformers obtained from the unbiased simulations are in excellent agreement with those determined by crystallography (backbone RMSD < 2 Å, Fig 2C), with the position of the benzene within 2 Å of the binding site residues established experimentally [4]. Further, kinetic and thermodynamic parameters obtained from the simulations are in reasonable agreement with the experimental values (Table 1), providing confidence in the results of the MD study.
Our unbiased μs timescale MD simulations establish that benzene, initially fully hydrated, reaches the internal cavity created by the L99A mutation via at least three different trajectories. Notably, the process does not involve a large-scale deformation to the T4L L99A structure, nor significant changes to secondary structure, but rather small perturbations to the positions of two or more helices in the C terminal domain of the protein that create pathways for binding (Fig 3). Similar tunnels from the protein surface to the cavity were also observed in a 30 μs MD simulation of T4L L99A using the AMBER force field [37]. The existence of multiple pathways coupled with a small net activation barrier explains how benzene can bind the occluded cavity rapidly.
Previously we had used combined MD simulations and NMR relaxation experiments to study how the side-chain of residue Phe114 fills the cavity created by the L99A mutation, that in some sense serves as a surrogate to hydrophobic ligands such as the benzene molecule considered here. It might be expected, therefore, that the binding pathways of Phe114 and benzene would share some similar features. Notably, in one of the trajectories (Trajectory 3) benzene enters the cavity in a manner similar to Phe114 [12] from its solvent exposed conformation.
The binding of ligands to occluded sites in other proteins, such as oxygen binding to the buried heme group in myoglobin, has been studied using a variety of experimental and computational techniques [38,39] In the case of myoglobin, the mechanism of binding also involves small barriers, multiple pathways and secondary binding sites in the protein [38,39]. Multiple pathways have also been observed by Dickson and Lotz in recent studies of an epoxide hydrolase inhibitor [40] and a trypsin-benzamidine system [41] that exploited a weighted ensemble based algorithm. The plasticity of T4L L99A allows the larger benzene molecule to reach the cavity in a manner similar to how oxygen connects with the heme group in myoglobin. The origin of the conformational plasticity in T4L L99A that facilitates this process is, in fact, the L99A mutation itself (Fig 5). Replacing the Leu99 residue with the smaller Ala not only creates the cavity but also allows helices to move with respect to one another to generate the pathways that are required for binding. In contrast, the presence of Leu99 destabilizes the binding competent conformations by clashing with Ile78 (Fig 5). It remains of interest to investigate if the benzene binding rate can be modulated by mutating Ile78 to smaller hydrophobic residues. Similar to the binding of benzene, protein plasticity and cavities have also been implicated in oxygen binding to T4L L99A [14].
Preliminary observations into the mechanism of ligand unbinding from T4L L99A have recently been published and are in agreement with the results of this study. Kitahara et. al. [14] developed a simple and elegant 15N NMR based method to detect O2 binding sites in proteins using T4L L99A as a model system. In order to quantify the 15N chemical shift changes that were observed upon oxygen binding MD simulations were performed to understand the rotational and translational diffusion properties of oxygen molecules in the T4L L99A cavities. Notably, an oxygen molecule escaped from the cavity via a pathway that is the reverse of Trajectory 1 (Fig 3). In MD simulations performed to understand how water interacts with the cavity at high pressures, water molecules were found to escape from the cavity due to transient openings formed at the junction of helices 5, 6 and 7 [42], as observed in the creation of a pathway to the cavity in Trajectory 3. In adiabatic-biased MD simulations of benzene unbinding from T4L L99A a pathway that is the reverse of Trajectory 3 was observed [36]. In a separate 27 μs MD trajectory of T4L L99A performed to understand the Phe114 exposed to buried conformational transition, tunnels leading to the cavity from solvent were transiently formed [37], as observed in the present work. A recent implementation of a biased simulation approach by Lindorf-Larson and coworkers also led to the extraction of the kinetics of the benzene recognition process [43]. McCammon and coworkers [44] have employed an accelerated MD-based biased simulation approach to explore the thermodynamics of benzene binding to T4L L99A cavity. Very recently, Arantes and coworkers [45] have reported weighted ensemble based biased simulations of unbinding of benzene from the T4L cavity and revealed multiple benzene exit pathways, in agreement with the current study. However, none of these simulation-based studies address the mechanism by which ligands spontaneously bind to buried protein cavities and consequently insights into multiple pathways and low free energy barriers were not obtained.
Recent studies suggest that diverse processes such as protein folding and protein interconversion between different compact conformers proceed via multiple pathways and/or small barriers [12,46,47]. More studies are required to determine if small activation barriers and multiple paths are common features of biomolecular recognition processes and to characterize at atomic resolution what these pathways are.
Here, as in most biophysical studies involving T4 lysozyme, the cysteine free version of the protein [48] was used, T4L WT*, where Cys54 and Cys97 are replaced by Thr and Ala, respectively. In T4L L99A the L99A mutation is introduced into the T4L WT* background.
|
10.1371/journal.pcbi.1005135 | Large-Scale Off-Target Identification Using Fast and Accurate Dual Regularized One-Class Collaborative Filtering and Its Application to Drug Repurposing | Target-based screening is one of the major approaches in drug discovery. Besides the intended target, unexpected drug off-target interactions often occur, and many of them have not been recognized and characterized. The off-target interactions can be responsible for either therapeutic or side effects. Thus, identifying the genome-wide off-targets of lead compounds or existing drugs will be critical for designing effective and safe drugs, and providing new opportunities for drug repurposing. Although many computational methods have been developed to predict drug-target interactions, they are either less accurate than the one that we are proposing here or computationally too intensive, thereby limiting their capability for large-scale off-target identification. In addition, the performances of most machine learning based algorithms have been mainly evaluated to predict off-target interactions in the same gene family for hundreds of chemicals. It is not clear how these algorithms perform in terms of detecting off-targets across gene families on a proteome scale. Here, we are presenting a fast and accurate off-target prediction method, REMAP, which is based on a dual regularized one-class collaborative filtering algorithm, to explore continuous chemical space, protein space, and their interactome on a large scale. When tested in a reliable, extensive, and cross-gene family benchmark, REMAP outperforms the state-of-the-art methods. Furthermore, REMAP is highly scalable. It can screen a dataset of 200 thousands chemicals against 20 thousands proteins within 2 hours. Using the reconstructed genome-wide target profile as the fingerprint of a chemical compound, we predicted that seven FDA-approved drugs can be repurposed as novel anti-cancer therapies. The anti-cancer activity of six of them is supported by experimental evidences. Thus, REMAP is a valuable addition to the existing in silico toolbox for drug target identification, drug repurposing, phenotypic screening, and side effect prediction. The software and benchmark are available at https://github.com/hansaimlim/REMAP.
| High-throughput techniques have generated vast amounts of diverse omics and phenotypic data. However, these sets of data have not yet been fully explored to improve the effectiveness and efficiency of drug discovery, a process which has traditionally adopted a one-drug-one-gene paradigm. Consequently, the cost of bringing a drug to market is astounding and the failure rate is daunting. The failure of the target-based drug discovery is in large part due to the fact that a drug rarely interacts only with its intended receptor, but also generally binds to other receptors. To rationally design potent and safe therapeutics, we need to identify all the possible cellular proteins interacting with a drug in an organism. Existing experimental techniques are not sufficient to address this problem, and will benefit from computational modeling. However, it is a daunting task to reliably screen millions of chemicals against hundreds of thousands of proteins. Here, we introduce a fast and accurate method REMAP for large-scale predictions of drug-target interactions. REMAP outperforms state-of-the-art algorithms in terms of both speed and accuracy, and has been successfully applied to drug repurposing. Thus, REMAP may have broad applications in drug discovery.
| Conventional one-drug-one-gene drug discovery and drug development is a time-consuming and expensive process. It suffers from high attrition rate and possible unexpected post-market withdrawal [1]. It has been recognized that a drug rarely only binds to its intended target, and off-target interactions (i.e. interactions between the drug and unintended targets) are common [2]. The off-target interaction may lead to adverse drug reactions (ADRs) [3], as demonstrated by the deadly side effect of a Fatty Acid Amide Hydrolase (FAAH) inhibitor in a recent clinical trial [4]. On the other hand, the off-target interaction may be therapeutically useful, thus providing opportunities for drug repurposing and polypharmacology [2]. Therefore, identifying off-target interactions is an important step in drug discovery and development in order to reduce the drug attrition rate and to accelerate the drug discovery and development process, and ultimately to make safer and more affordable drugs.
Many efforts have been devoted to developing statistical machine learning methods for the prediction of unknown drug-target associations by screening large chemical and protein data sets [5]. One of the fundamental assumptions in applying statistical machine learning methods to drug-target interaction prediction is that similar chemicals bind to similar protein targets, and vice versa. Based on this similarity principle, both semi-supervised and supervised machine learning techniques have been applied. The semi-supervised learning methods either build statistical models for the k nearest neighbors (k-NN) of the query compound with similar compounds in the database (e.g. Parzen-Rosenblatt Window (PRW) [6] and Set Ensemble Analysis (SEA) [7] are examples). Although a large number of 2D and 3D fingerprint representations of chemical structures have been developed, chemical structure similarity that is measured by Tanimoto coefficient (TC) or other similarity metrics of fingerprints is not continuously correlated with the binding activity. Activity cliff exists in the chemical space, where a small modification of a chemical structure can lead to a dramatic change in binding activity [8]. Thus, the chemical structural similarity alone is not sufficient to capture genome-wide target binding profile, as protein-chemical interaction is determined by both protein structures and chemical structures. New deep learning techniques that can learn non-linear, hierarchical relationships may provide new solutions for representing chemical space [9–12]. However, few work has been done to incorporate protein relationships into the deep learning framework. It remains to be seen whether the deep learning is applicable to genome-wide target prediction.
A number of techniques such as Gaussian Interaction Profile (GIP), Weighted Nearest Neighbor (WNN), Regularized Least Squares (RLS) classifier [13, 14], and matrix factorization [15–17] have been developed to integrate chemical and genomic space. Among them, Neighborhood Regularized Logistic Matrix Factorization (NRLMF) [17] and Kernelized Bayesian Matrix Factorization (KBMF) [16] are two of the most successful methods. However, several drawbacks in these algorithms hinder their applications in genome-wide off-target predictions. First, several algorithms with high performance such as KBMF are extremely time and memory-consuming. Second, these algorithms depend on a supervised learning framework that requires negative cases. While publicly available biological and/or chemical databases (e.g. ZINC [18], ChEMBL [19], DrugBank [20], PubChem [21], and UniProt [22]) have enabled large-scale screening of drug-target associations, the known chemical-protein associations are sparse, and the number of reported negative cases (i.e. chemical-protein pairs not associated) is too small to optimally train a prediction algorithm [23]. Using randomly generated negative cases will adversely impact the performance of these algorithms, and algorithmically derived negative cases are often based on unrealistic assumptions [23]. Finally, these algorithms have been mainly evaluated for the prediction of off-targets within the same gene family (e.g. GPCR) using a small benchmark with hundreds of drugs and targets. Their performances in predicting off-target across gene families on a large scale are uncertain. Indeed, drug cross-reactivity often occurs across fold spaces [2]. Thus, the development of in silico prediction methods that are fast as well as accurate enough to explore the available data is urgent.
Here, we make several contributions to address the aforementioned problems. First, we present an efficient method, REMAP, which formulates the off-target predictions as a dual-regularized One Class Collaborative Filtering (OCCF) problem. Thus, negative data are not needed for the training, but can be used if available. Secondly, REMAP is highly scalable with promising accuracy, thus can be applied to large-scale off-target predictions. Thirdly, we introduce a new benchmark set to evaluate the performance of drug-target interactions across gene families. Finally, we apply REMAP to repurposing existing drugs for new diseases. We identified seven drugs that have anti-cancer activity. Six of them are supported by experimental evidence.
The problem we try to solve here is to predict how likely it is that a chemical interacts with a target protein, using a chemical-protein association network, chemical-chemical similarity, and protein-protein similarity information. We start by preparing a bipartite network for chemical-protein associations as a sparse n × m matrix R, where n is the number of chemicals and m is the number of proteins. Ri,j = 1 if the ith chemical is associated with the jth protein, and Ri,j = 0, otherwise. The chemical-chemical similarity scores are in an n × n square matrix C, with Ci,j representing the chemical-chemical similarity score between the ith and jth chemicals (0 ≤ Ci,j ≤ 1) for total n chemicals. The protein-protein similarity scores are in the same format for total m proteins (0 ≤ Ti,j ≤ 1). We consider this problem an analog of user-item preferences such that users and items represent chemicals and proteins, respectively. Therefore, the problem is to provide an n × m matrix P in which Pi,j is the prediction score for the interaction between the ith chemical and the jth protein.
Our prediction method REMAP is based on a one-class collaborative filtering algorithm that recommends the users’ preferences to the listed items [24]. It assumes that similar users will prefer similar items, unobserved associations are not necessarily negative, and user-item preferences can be analogous to drug-target associations. Assuming that a fairly low number of factors (i.e. smaller number of features than the number of total chemicals or protein targets) may capture the characteristics determining the chemical-protein associations, two low-rank matrices, U (chemical side) and V (protein side), were approximated such that ∑in∑jm{R−(U⋅VT)} is minimized where R is the matrix for known chemical-protein associations and VT is the transposition of the protein side low-rank matrix V. The two low rank matrices, Un×r and Vm×r are obtained by iteratively minimizing the objective function,
minU,V≥0∑(i,j)pwt(R(i,j)+pimp−U(i,:)⋅V(j,:)T)2+preg(‖U‖2+‖V‖2)+pchemtr(UT(DC−C)U)+pprottr(VT(DT−T)V)
(1)
All symbols used in the paper are summarized in Table 1, and the overall process of REMAP is in Fig 1. Here, pwt is the penalty weight on the observed and unobserved associations which indicate the reliability of the assigned probability of true association, pimp is the imputed value (i.e. the probability of unobserved associations as real associations), preg is the regularization parameter to prevent overfitting, pchem is the importance parameter for chemical-chemical similarity, pprot is the importance parameter for protein-protein similarity, and tr(A) is the trace of matrix A (Table 1). In this study, we use global weight and imputation. However, the weight and imputation values may be determined by a priori knowledge or from the prediction of other machine learning algorithms (i.e. pwt and pimp can be matrices with the same dimension as the matrix R). The raw predicted score for the ith chemical to bind the jth protein can be calculated by P(i,j)=UUP(i,:)⋅VUP(j,:)T. The raw scores were adjusted based on the ratio of observed positive and negative cases when the negative data are available (explained in the prediction score adjustment section). Also, the matrix Un×r is referred to as a low-rank drug profile since its ith row represents the ith drug’s behavior in the drug-target interaction network as well as drug-drug similarity spaces compressed to r number of features. The REMAP code was originally written in Matlab and modified for drug-target predictions.
Chemical-chemical similarity scores are one of the required inputs of REMAP. Although there are a number of metrics developed for chemical-chemical similarity, a recent study showed that Tanimoto coefficient-based similarity is highly efficient for fingerprint-based similarity measurement [25]. The fingerprint of choice in this study is the Extended Connectivity Fingerprint (ECFP), which has been successfully applied to chemical structure-based target prediction method, PRW [6]. Thus, it allows for a fair comparison of REMAP with PRW. It is interesting to compare the different fingerprints in the future study.
To calculate a similarity score between two chemicals, c1 and c2, the Tanimoto dissimilarity coefficient dTani (c1,c2) was obtained using JChem with the Tanimoto metric for the ECFP descriptor type using the command in the Unix environment, “ChemAxon/JChem/bin/screenmd target_smi query_smi -k ECFP -g -c -M Tanimoto” [26]. The chemical-chemical similarity score, C (c1,c2) is defined as C (c1,c2) = 1-dTani (c1,c2). Briefly, two chemicals have a higher similarity score if they have more of the same chemical moieties (e.g. functional groups) at more similar relative positions. Chemical similarity scores below 0.5 were treated as noise and set to 0.
Protein-protein similarity scores are also one of the required inputs for REMAP. The similarity between two proteins was calculated based on their sequence similarity using NCBI BLAST [27] with an e-value threshold of 1 × 10−5 and its default options (e.g. 11 for gap open penalty and 1 for its extension, BLOSUM62 for the scoring matrix, and so on). Based on our 10-fold cross validation (see below), e-value thresholds from 1 to 1 × 10−20 did not significantly affect the performance (S1 Fig). Therefore, we decided to use a moderately stringent threshold (BLAST default is 1 × 10−3). A similarity score for query protein p1 to target protein p2 was calculated by the ratio of a bit score for the pair compared to the bit score of a self-query. To be specific, for the query protein p1 to the target protein p2, protein-protein the similarity score was defined such that T(p1,p2) = dbit(p1,p2)/dbit(p1,p1).
For benchmark tests, ZINC data was filtered by IC50 ≤ 10 μM, which yielded 31,735 unique chemical-protein associations for 12,384 chemicals and 3,500 proteins (ZINC dataset [18]). Targets that are protein complexes or cell-based tests were excluded. Proteins whose primary sequence is unavailable were also excluded. Protein sequences were obtained from UniProt [22], and the whole protein sequences were used to calculate protein-protein similarity scores.
To assess the predictive power of our algorithm, we performed a 10-fold cross validation on the ZINC dataset described above. We set the parameters as follows: pwt = pimp = preg = 0.1, r = 300, pchem = 0.75, pprot = 0.1, and piter = 400. The optimized values determined by the 10-fold cross validation of benchmark are shown in S2 Fig. It is noted that the best performance is achieved when pchem = 0.25 and pprot = 0.25. To further evaluate REMAP, we compared its performance on the ZINC dataset with several methods: a chemical similarity-based method (PRW [6]), the best performed matrix factorization methods so far (NRLMF [17] and KBMF with twin kernels (KBMF2K) [16]), combination of WNN and GIP (WNNGIP [14]), and another type of matrix factorization method (Collaborative Matrix Factorization (CMF) [15]) for different types of chemicals and proteins.
To obtain a detailed view of the performance of the methods, we divided the ZINC dataset into 3 categories with 2 subcategories for each, based on the connectivity of known chemical-protein associations and the degree of uniqueness of the chemicals. First, all the chemicals in the dataset were classified into the chemicals having only one known target (NT1), two known targets (NT2), or three or more known targets (NT3). Then, for the chemicals in each category, they were further divided based on either the number of known chemicals (ligands) the target proteins are associated with (number of ligands in increments of 5) or the maximum chemical-chemical similarity score for the chemical in the dataset (the similarity score range increment is 0.1). The label used in this paper for the dataset are NTaLb, or NTaMaxTcd, where ‘NT’ stands for the Number of known Target, ‘L’ for the number of known Ligand, and ‘Tc’ for the maximum (Tanimoto coefficient-based) chemical-chemical similarity score for the given chemical in the dataset, with NT = a, b ≤ L ≤ b +4, and d − 0.1 < Tc ≤ d. For instance, NT2L1 is the data set label for chemicals having two known targets and proteins having 1 to 5 ligands in the dataset, and NT1Tc0.9 is for chemicals with the most similar chemicals between 0.8 and 0.9 of similarity scores and having one known target. Chemicals having more than three known targets are included in the NT3 class, and proteins having more than twenty-one known ligands were included in L21 (not limited to 25). The categories of the ZINC dataset were then used to evaluate the performance of off-target prediction, and their labels mean the number of known ligands (L) or the maximum structural similarity (Tc) with their corresponding ranges. For example, ‘L21more’ stands for the dataset for proteins having 21 or more known targets, and ‘Tc0.9to1.0’ stands for maximum structural similarity greater than 0.9 and up to 1.0 (Tc0.5to0.6 is inclusive of 0.5). Note that NT1 is equivalent to chemicals without any known target when they are tested for cross validation. Therefore, performances on NT1 datasets reflect the ability to address the cold start problem. In other words, when one known drug-target association is intentionally hidden for the chemicals in the NT1 dataset, the tested chemicals will not have any known target in the training data, and they are less likely to be given a good recommendation of targets. This is analogous to the new user or new item problem reviewed by Su et al. [28].
A typical measure of prediction performance is the Receiver Operating Characteristic (ROC) curve by which one can assess the reliability of the positively predicted results. However, it is difficult to apply the ROC curve on our chemical-protein association datasets since the vast majority of the chemical-protein pairs have not been tested, and thus it is unclear whether the missing entries are actually unassociated or just not yet observed.
In order to assess how reliable the positively predicted results from REMAP are, we needed to define a performance measurement that is analogous to ROC curve but not dependent on the true negatives. Our primary measure of performance is the true positive rate (∑True Positives∑Condition Positives; Recall or Recovery) at the top 1% of predictions for each chemical. To be specific, the top 1% of predictions includes up to the 35th-ranked predicted target protein for a chemical for our datasets (3,500 possible target proteins for each chemical). Thus, for instance, a TPR of 0.965 at the 35th cutoff rank (top 1%) means that 96.5% of the total tested positive pairs were ranked 35th or better for the tested chemicals.
In order to assess the speed of REMAP for practical uses, we measured its running time by varying the rank parameter or the size of dataset. On the ZINC dataset (12,384 chemicals and 3,500 proteins), up to r = 2,000 was tested, and at fixed r = 200, dataset sizes up to 200,000 chemicals and 20,000 proteins were tested. The number of iterations (piter) was fixed to 400. A single node of CPU with 2.88 GB of memory in the City University of New York High Performance Computing Center (CUNY HPCC) was used for REMAP running time tests. We also compared the running times of different matrix factorization methods with ours. Due to the large time complexity and memory requirement for other algorithms, a multi-core node with up to 700 GB of shared memory system in CUNY HPCC was used for them on the ZINC dataset.
Chemical-protein associations were obtained from the ZINC [18], ChEMBL [19] and DrugBank [20] databases. To obtain reliable chemical-protein association pairs, binding assays records with IC50 information were extracted from the databases, and the cutoff IC50 value of 10 μM was used where applicable. Two chemicals were considered the same if their InChI Keys are identical, and two proteins were considered so if their UniProt Accessions are identical. For records with IC50 in μg/L (found in ChEMBL), the full molecular weights of the compounds listed on ChEMBL were used to convert μg/L to μM. Chemical-protein pairs were considered associated if IC50≤10 μM (active pairs), unassociated if IC50>10 μM (inactive pairs), ambiguous if records exist in both ranges (ambiguous pairs), and unobserved otherwise (unknown pairs). A total of 198,712 unique chemicals and 3,549 unique target proteins were obtained from the combination of ChEMBL and ZINC with 228,725 unique chemical-protein active pairs, 76,643 inactive pairs, and 4,068 ambiguous pairs. Of the 198,712 chemicals, 722 were found to be FDA-approved drugs. Furthermore, drug-target relationships were extracted from the DrugBank and integrated into the ZINC_ChEMBL dataset above. A total of 199,338 unique chemicals and 6,277 unique proteins were obtained from the combination of ZINC, ChEMBL, and DrugBank with 233,378 unique chemical-protein active pairs.
Since REMAP showed promising performances on predicting off-targets for chemicals with at least one known target, it is possible to use REMAP to suggest new purposes for some FDA approved drugs. As the matrix product of UUP (chemical-side low-rank matrix) and VUP (protein side low-rank matrix) is the predicted drug-target interaction matrix P, the ith row of UUP contains the target interaction profile for the ith drug. Therefore, we analyzed the drug-drug similarities based on the low-rank matrix UUP. We ran REMAP with the data combination of three databases explained above, with the parameters used in the benchmark evaluations. Then, we calculated drug-drug cosine similarities based on the matrix UUP. For each row of UUP for FDA approved drugs, the cosine similarity of drug c1 and drug c2 can be calculated by, Scos,(c1,c2) = Uc1→∙Uc2→Uc1Uc2. To search for possibly undiscovered uses of the drugs, we focused on drugs that are found to have high cosine similarity but low Tanimoto similarity (< 0.5). Markov Cluster (MCL) Algorithm [29, 30] was used to cluster drugs based on their cosine similarity of a low-rank target profile. Drug-disease associations were obtained from the Comparative Toxicogenomics Database (CTD) [31].
The raw prediction score (P(i,j) = UUP(i,:)∙VUP(j,:)T) can be adjusted to better reflect the real data as well as to statistically discriminate the positive and negative predictions. We used the active, inactive and ambiguous pairs obtained from the ChEMBL database to adjust the score. REMAP prediction on the ZINC_ChEMBL dataset showed a clear division between the active and inactive pairs, suggesting that predictions scored around 1.0 are highly likely to be positive (Fig 2A). As mentioned above, however, there is a large difference between the number of active and inactive pairs, which is not likely to reflect the ratio of the actual positive and negative chemical-protein pairs. Greater accuracy is expected by adjusting the prediction scores to reflect such a positive/negative ratio. To estimate the ratio, we first normalized the counts in each bin in the histogram (Fig 2A) and calculated the weights that minimize the sum of error, Esum. Esum(w1) = Σi[Ai − {w1pi + (1 − w1)Ni}]2, where w1 and w2 are the weights on active and inactive pairs, respectively (w1 + w2 = 1.0), and Ai, pi and Ni are the normalized counts in ith bin of ambiguous, active and inactive pairs, respectively. The optimum adjustment weights were approximately w1 = 0.16, w2 = 0.84 (Fig 2B). This implies that approximately 16% of total observations are positive. Since the ratio of negative/positive is about 5.25 (w2w1 = 5.25), we increased the number of observations for inactive pairs in each bin by 5.25 times and rounded down. The adjusted prediction score for each bin (Bi) was calculated using the increased negative counts.
It is noted that the prediction score adjustment was not used in the benchmark study, where no negative data were used.
Drug-drug clustered network was visualized using Cytoscape [32].
We evaluated the performances of algorithms for chemicals having one, two, or more than three known targets with varying maximum chemical-chemical similarity ranges or with proteins having a certain number of known ligands (dataset prepared as explained in the methods and materials section). In general, the performances of both algorithms improve as the number of known ligands per protein or the maximum chemical-chemical similarity value increases.
It was noticeable that REMAP performed significantly better than PRW when there was at least one known target for a chemical whose targets are predicted (Figs 3 and 4). REMAP showed greater than 90% recovery at the top 1% when the tested chemicals have at least one known target. All algorithms are sensitive to the number of ligands per target. The more ligands, the higher accuracy. While PRW also reached reasonably high recovery for some categories (e.g. more than 11 known ligands per proteins, or C(c1,c2)>0.6 of the most similar trained chemicals), REMAP showed that it is reliable for testing chemicals without high similarity to the trained chemicals (Figs 3B and 4B). In other words, REMAP is applicable to chemicals that are structurally distant to the chemicals already in the dataset. Except where the target proteins have 1 to 5 known ligands, REMAP performed best among the three algorithms in all cases with at least one known target for the tested chemicals (Figs 3 and 4). In the most of cases, the differences in the performance between REMAP and other two algorithms are statistically significant. Therefore, in practice, REMAP can predict potential drug targets for chemicals with at least one known target as training data, even when the chemicals are structurally dissimilar to the training chemicals. With the optimized parameters (see below), ROC-like curves shows the general trend of performances of the three algorithms up to the top 10% of predictions (S3 and S4 Figs).
As shown in Figs 3 and 4, REMAP outperforms the state-of-the-art NRLFM algorithm in most of the tested cases. As NRLMF is sensitive to the rank parameter, we carried out optimizations to determine optimal rank and iterations for NRLMF (S5 Fig). The optimal rank and iterations used in the evaluation were 100 and 300, respectively. Moreover, in the current implementation, REMAP is approximately 10 times faster and uses 50% less memory than NRLMF. Consistent with the results by Liu et al. [17], the accuracies of NRLFM are significantly higher than KBMF2K, CMF, and WNNGIP in all of ZINC benchmarks. Overall, REMAP is one of the best-performing methods for the genome-wide off-target predictions.
To test whether the chemical-chemical similarity matrix helps prediction, we performed 10-fold cross validation on the ZINC dataset with the contents of the chemical-chemical or the protein-protein similarity matrix controlled. In other words, about half of the non-zero chemical-chemical similarity scores were randomly chosen and removed (set to 0) for the “half-filled chemical similarity” matrix, and all entries are set to 0 for the “zero-filled chemical similarity” matrix. The predictive power of REMAP showed noticeable improvement when all available chemical-chemical similarity pairs were used, compared to the half-filled or the zero-filled similarity matrix (Fig 5A). Similarly, the contents of the protein-protein similarity matrix were controlled (e.g. half-filled protein similarity, and zero-filled protein similarity) while the full chemical similarity matrix was used. Unlike the chemical-chemical similarity, the protein-protein similarity information did not necessarily improve REMAP’s predictive power. The performance was best when a half of the protein-protein similarity information was used together with the full chemical-chemical similarity matrix (Fig 5B). This suggests that there is significant noise in the protein-protein sequence similarity matrix although the information does help prediction. A careful examination of the BLAST-based protein-protein similarity matrix may give an insight into the design of a novel protein-protein similarity metric for drug-target binding activities (see discussion section).
We also performed optimization tests for pchem and pprot on ZINC dataset. Although the performance was slightly better when the chemical-chemical similarity importance was maximum (Fig 6A), the difference was too small to conclude that it is best to fix pchem = 1. Instead, the prediction may rely too much on the chemical-chemical similarity scores. Therefore, to allow flexibility on chemical-chemical similarity information, we set pchem = 0.75 at which the performance was almost as accurate as pchem = 1. On the other hand, the performance was best when the protein-protein sequence similarity importance, pprot, was 0.1 (Fig 6B), further supporting our claim that protein-protein sequence similarity is not an optimal choice for the prediction of a drug-target interaction. When jointly optimizing pchem and pprot, their optimal value is 0.25 and 0.25, respectively, in the 10-fold cross validation benchmark evaluation (S2B Fig).
Our result supports a recent study [25] which showed that Tanimoto coefficient is efficient for the chemical similarity calculation. Chemical fingerprint-based chemical-protein association prediction has been studied by Koutsoukas et al [6]. By defining bins (target proteins) that can contain certain chemical features based on the chemical fingerprints, Koutsoukas et al. successfully demonstrated that their algorithm, PRW, can efficiently predict unknown chemical-protein associations [6]. While the basic idea of dissecting chemical compounds into functional groups is the same, it should be noted that PRW does not consider the information obtained from proteins, as well as interactome.
For all our tests, REMAP showed great speed without losing its accuracy. On our benchmark dataset (ZINC; 12,384 chemicals and 3,500 proteins), it took approximately 120 seconds to run 400 iterations at the rank of 200 (r = 200, piter = 400). The time complexity is linearly dependent on the rank (Fig 7A). The scalability of REMAP is superior when compared to KBMF2K, a state-of the art matrix factorization algorithm that is implemented in Matlab and has been extensively studied for predicting drug-target interactions [16]. KBMF2K took more than 10 days for the same size matrix using the same computer system in the ZINC benchmark. Moreover, REMAP was capable of higher rank factorization while KBMF2K was limited to rank 200 in our system due to the memory requirement (over 100 GB of memory). At a much higher rank (r = 2,000), less than one hour was required for REMAP on the same dataset (Fig 7A). Time complexity experiments on larger dataset showed that REMAP completed predictions on a dataset with 200,000 rows and 20,000 columns within 2 hours on a single core computing system with 2.88 GB of memory, demonstrating its ability to screen the whole human genome of approximately 20,000 proteins in two hours (Fig 7B).
Since REMAP is scalable and shows superior accuracy based on our benchmark tests, we performed large scale prediction of drug-target interactions on the ZCD dataset (explained in the Materials and Methods section). As explained in the prediction score adjustment section, prediction scores for the active pairs were mostly located between 0.75 and 1.0 (Fig 2A).
As expected, the percentage of pairs of chemicals that share common targets decreases with the decrease of the chemical structural similarity measured by the Tc of ECFP fingerprints (C(c1,c2)). The percentage of target-sharing chemical pairs drops below 50% and 0.5% when the Tc is between 0.5 and 0.6, and less than 0.5, respectively (S6 Fig). Thus, it is less likely that the chemical structural similarity alone can reliably detect novel binding relations between two chemicals when the Tc is less than 0.5. It is interesting to see how REMAP performs when the chemical structural similarity fails.
We analyzed the low-rank drug profile (matrix UUP) to check whether it represented the target-binding behavior of the drugs. When filtered by low chemical structure similarity (C(c1,c2)<0.5)), there are 899,871 drug-drug pairs. Among them, the profile similarity score (Scos,(c1,c2)) of 91,888 pairs is higher than 0.3. With high profile similarity (0.99≤Scos,(c1,c2)≤1)), a total of 1,327 drug-drug pairs were found of which 1,033 pairs shared at least one common known target. S7 Fig shows the percentage of pairs that share the common target in different profile similarity bucket for FDA-approved drugs. This result suggests that REMAP is able to provide a chemical-protein binding profile that cannot be captured by chemical structure similarity alone.
When Scos,(c1,c2)≤0.3, the percentage of two drugs that share a common target drops below 50% (S7 Fig). We constructed a drug-drug similarity network by filtering out drug pairs with Scos,(c1,c2)≤0.3, then applied the MCL algorithm on the drug-drug network to find clusters of similar drugs. The largest cluster of drugs contained a total of 313 drugs, and their relationships to diseases were examined based on the known associations annotated in CTD [31]. As a result, we found that the drugs are mostly related to mental disorders, including hyperkinesis, dystonia, catalepsy, schizophrenia and basal ganglia diseases as the mostly related diseases. The most frequent known protein targets by the drugs were GPCRs (S1 Table). It is comparable that GPCRs were 1,924 times targeted while kinases were targeted only 55 times. While it is interesting to further examine the cluster, validating all of the possible drug-target pairs in the largest cluster may be inefficient.
A smaller cluster of drugs contained a total of thirty-one FDA approved drugs twenty-six of which are known to target kinases or interact with microtubule (Table 2). Seven drugs in the cluster have not been used for cancer treatment and were found to be closely linked to the anti-cancer drugs (Fig 8 and Table 2). Interestingly, several of them have been tested for their anti-cancer activity. For example, colchicine (also known as colchine), an FDA approved drug for gout treatment, has been shown to have anti-proliferative effects on several human liver cancer cell lines at clinically acceptable concentrations [33]. Griseofulvin, an antifungal antibiotic drug, appears to be effective as an anti-cancer drug when used together with other anti-cancer drugs [34]. The three anthelmintic drugs, albendazole, mebendazole and niclosamide, have been studied and repurposed for their anti-cancer effects on different types of cancers. Albendazole has been shown to be effective in suppressing liver cancer cells both in vitro and in vivo [35], and recently has been repurposed for ovarian cancer treatment with a bovine serum albumin-based nanoparticle drug delivery system [36]. Mebendazole showed anti-cancer activities in human lung cancer cell lines [37] and human adrenocortical cell lines [38], and it has been repurposed for colon cancer treatment [39]. Both niclosamide and mebendazole showed beneficial effects in glioblastoma in different studies [40, 41]. It has been proposed to use aprepitant in combination with other compounds to improve the efficiency of temozolomide, the current standard drug for glioblastoma treatment [42]. Anti-cancer activity of carbidopa hydrate have not yet been reported. It will be interesting to experimentally validate the prediction.
Our extensive benchmark studies show that REMAP outperforms existing algorithms in most of the cases for the off-target prediction. Compared with other state-of-the-art matrix factorization algorithms, the predictive power of REMAP comes from several improvements. First, we formulated the drug-target prediction as a one-class collaborative filtering problem; thus the negative data are not required for the training. Second, a priori knowledge including known negative data can be incorporated into the matrix factorization with imputation and weighting. Finally, using global imputation and weighting, the algorithm is computationally efficient without significantly sacrificing its performance.
The efficiency and effectiveness of REMAP allows us to predict proteome-wide target binding profiles of hundreds of thousands of chemicals. As the proteome-wide target binding profile is more correlated with phenotypic response than a single target binding, REMAP will facilitate linking molecular interactions in the test tube with in vivo drug activity. When using a multi-target binding profile predicted by REMAP as the signature of a chemical compound, seven drugs were found to be associated with anti-cancer therapeutics, although they do not have detectable chemical structural similarity. Among them, the anti-cancer activity of six drugs was supported by experimental evidences. Thus, REMAP could be a useful tool for drug repurposing.
Although REMAP showed its high potential on genome-wide off-target predictions as discussed above, two issues remain: the cold start problem and suboptimal protein-protein similarity metrics. Similar to matrix factorization algorithms such as NRLMF, REMAP suffers from cold start problem, also known as new user or new item problem. In other words, it is difficult to recommend a product for a new user if the new user has never purchased or reviewed a product in the database [28]. For novel chemicals that do not have any known target in the dataset, REMAP did not show better performance than PRW. Moreover, if the target of the novel chemical has 5 or fewer known ligands, the recovery of REMAP is lower than 0.5 (S8A Fig). When the novel chemical is similar to those chemicals in the database, the recovery of REMAP reached above 90% (S8B Fig). These results suggest that, in practice, existing matrix factorization-based methods, including REMAP, are not the optimal choice if the chemicals of interest do not have any known target. To resolve this issue, it is possible to design an algorithm that combines the benefits of PRW or other algorithms with REMAP. The use of confidence weights and a priori imputation makes it straightforward for REMAP to incorporate additional information. In addition, the time and memory efficiency of REMAP makes it possible to apply active learning to overcome the cold start problem [43–46].
The suboptimal performance of REMAP may arise from the lack of molecular-level biochemical details in deriving the protein-protein similarity metrics. When testing the ZINC dataset, we found that REMAP performs better as lower weight was assigned for protein-protein sequence similarity data (Fig 6B). In addition, the predictive power of REMAP improved when about half of the randomly selected protein-protein similarity scores were removed, further confirming that noise confounds relating global sequence similarity to ligand binding (Fig 5B). It is not surprising that proteins with similar sequences do not necessarily bind to similar chemicals, as protein-ligand interaction is governed by the spatial organization of amino acid residues in the protein structure [47]. Amino acid mutations/post-translational modifications and conformational dynamics may alter the binding of the ligand through direct modification of the ligand binding site or allosteric interaction. A protein may also consist of multiple binding sites that accommodate different types of ligands. Thus, two proteins with high sequence similarity do not necessarily bind the same ligands because the two proteins may possess different 3D conformations, especially in their binding pockets [47]. In contrast, two proteins with low sequence similarity can bind to the same ligands if their binding pockets are similar [48, 49]. The binding site similarity can be a more biologically sensitive measure of protein-protein similarity for the off-target prediction [50–55]. Such work is on-going.
In silico drug-target screening is an essential step to reduce costly experimental steps in drug development. In this study, we showed that dual-regularized one-class collaborative filtering algorithm, a class of computational methods frequently used in user-item preference recommendations, may be applied to drug-target association predictions. Our study presents REMAP, a collaborative filtering algorithm with capability of running whole human genome-level predictions within two hours. Other studies on some types of cancer treatment support our algorithm’s ability to capture drug-drug similarities based on both the drug-target interaction profile and the chemical structural similarity. Our study shows the limitation of REMAP in evaluating new chemicals or accommodating biochemical details. Further development of the computational tools for better prediction is needed.
|
10.1371/journal.pcbi.1000709 | Comparing Families of Dynamic Causal Models | Mathematical models of scientific data can be formally compared using Bayesian model evidence. Previous applications in the biological sciences have mainly focussed on model selection in which one first selects the model with the highest evidence and then makes inferences based on the parameters of that model. This “best model” approach is very useful but can become brittle if there are a large number of models to compare, and if different subjects use different models. To overcome this shortcoming we propose the combination of two further approaches: (i) family level inference and (ii) Bayesian model averaging within families. Family level inference removes uncertainty about aspects of model structure other than the characteristic of interest. For example: What are the inputs to the system? Is processing serial or parallel? Is it linear or nonlinear? Is it mediated by a single, crucial connection? We apply Bayesian model averaging within families to provide inferences about parameters that are independent of further assumptions about model structure. We illustrate the methods using Dynamic Causal Models of brain imaging data.
| Bayesian model comparison provides a formal method for evaluating different computational models in the biological sciences. Emerging application domains include dynamical models of neuronal and biochemical networks based on differential equations. Much previous work in this area has focussed on selecting the single best model. This approach is useful but can become brittle if there are a large number of models to compare and if different subjects use different models. This paper shows that these problems can be overcome with the use of Family Level Inference and Bayesian Model Averaging within model families.
| Mathematical models of scientific data can be formally compared using Bayesian model evidence [1]–[3], an approach that is now widely used in statistics [4], signal processing [5], machine learning [6], natural language processing [7], and neuroimaging [8]–[10]. An emerging area of application is the evaluation of dynamical system models represented using differential equations, both in neuroimaging [11] and systems biology [12]–[14].
Much previous practice in these areas has focussed on model selection in which one first selects the model with the highest evidence and then makes inferences based on the parameters of that model [15]–[18]. This ‘best model’ approach is very useful but, as we shall see, can become brittle if there are a large number of models to compare, or if in the analysis of data from a group of subjects, different subjects use different models (as is the case for a random effects analysis [19]). This brittleness, refers to the fact that which is the best model can depend critically on which set of models are being compared. In random effects analysis, augmenting the comparison set with a single extra model can, for example, reverse the ranking of the best and second best models. To address this issue we propose the combination of two further approaches (i) family level inference and (ii) Bayesian model averaging within families.
We envisage that these methods will be useful for the comparison of large numbers of models (eg. tens, hundreds or thousands). In the context of neuroimaging, for example, inferences about changes in brain connectivity can be made using Dynamic Causal Models [20],[21]. These are differential equation models which relate neuronal activity in different brain areas using a dynamical systems approach. One can then ask a number of generic questions. For example: Is processing serial or parallel? Is it linear or nonlinear? Is it mediated by changes in forward or backward connections? A schematic of a DCM used in this paper is shown in Figure 1. The particular questions we will address in this paper are (i) which regions receive driving input? and (ii) which connections are modulated by other experimental factors?
This paper proposes that the above questions are best answered by ‘Family level inference’. That is inference at the level of model families, rather than at the level of the individual models themselves. As a simple example, in previous work [19] we have considered comparison of a number of DCMs, half of which embodied linear hemodynamics and half nonlinear hemodynamics. The model space was thus partitioned into two families; linear and nonlinear. One can compute the relative evidence of the two model families to answer the question: does my imaging data provide evidence in favour of linear versus nonlinear hemodynamics? This effectively removes uncertainty about aspects of model structure other than the characteristic of interest.
We have provided a simple illustration of this approach in previous work [19]. We now provide a formal introduction to family level inference and describe the key issues. These include, importantly, the issue of how to deal with families that do not contain the same number of models. Additionally, this paper shows how Bayesian model averaging can be used to provide a summary measure of likely parameter values for each model family. We provide an example of family-level inference using data from neuroimaging, a DCM study of auditory word processing, but envisage that the methods can be applied throughout the biological sciences. Before proceeding we note that the use of Bayesian model averaging is a standard approach in the field of Bayesian statistics [4], but has yet to be applied extensively in computational biology. The use of model families is also accomodated naturally within the framework of hierarchical Bayesian models [1] and is proposed to address the well known issue of model dilution [4].
This section first briefly reviews DCM and methods for computing the model evidence. We then review the fixed and random effects methods for group level model inference, which differ as to whether or not subjects are thought to use the same or a different model. This includes the description of a novel Gibbs sampling method for random effects model inference that is useful when there are many models to compare. We then show that, for random effects inference, the selection of the single best model can be critically dependent on the set of models that are to be compared. This then motivates the subsequent subsection on family level inference, in which inferences about model characteristics are invariant to the comparison set. We describe family level inference in both a fixed and random effects context. The final subsection then describes a sample-based algorithm for implementing Bayesian model averaging using the notion of model families.
Dynamic Causal Modelling is a framework for fitting differential equation models of neuronal activity to brain imaging data using Bayesian inference. The DCM approach can be applied to functional Magnetic Resonance Imaging (fMRI), Electroencephalographic (EEG), Magnetoencephalographic (MEG), and Local Field Potential (LFP) data [22]. The empirical work in this paper uses DCM for fMRI. DCMs for fMRI comprise a bilinear model for the neurodynamics and an extended Balloon model [23] for the hemodynamics. The neurodynamics are described by the following multivariate differential equation(1)where indexes continuous time and the dot notation denotes a time derivative. The th entry in corresponds to neuronal activity in the th region, and is the th experimental input.
A DCM is characterised by a set of ‘exogenous connections’, , that specify which regions are connected and whether these connections are unidirectional or bidirectional. We also define a set of input connections, , that specify which inputs are connected to which regions, and a set of modulatory connections, , that specify which intrinsic connections can be changed by which inputs. The overall specification of input, intrinsic and modulatory connectivity comprise our assumptions about model structure. This in turn represents a scientific hypothesis about the structure of the large-scale neuronal network mediating the underlying cognitive function. A schematic of a DCM is shown in Figure 1.
In DCM, neuronal activity gives rise to fMRI activity by a dynamic process described by an extended Balloon model [24] for each region. This specifies how changes in neuronal activity give rise to changes in blood oxygenation that are measured with fMRI. It involves a set of hemodynamic state variables, state equations and hemodynamic parameters, . In brief, for the th region, neuronal activity causes an increase in vasodilatory signal that is subject to autoregulatory feedback. Inflow responds in proportion to this signal with concomitant changes in blood volume and deoxyhemoglobin content .(2)Outflow is related to volume through Grubb's exponent [20]. The oxygen extraction is a function of flow where is resting oxygen extraction fraction. The Blood Oxygenation Level Dependent (BOLD) signal is then taken to be a static nonlinear function of volume and deoxyhemoglobin that comprises a volume-weighted sum of extra- and intra-vascular signals [20](3)where is resting blood volume fraction. The hemodynamic parameters comprise and are specific to each brain region. Together these equations describe a nonlinear hemodynamic process that converts neuronal activity in the th region to the fMRI signal (which is additionally corrupted by additive Gaussian noise). Full details are given in [20],[23].
In DCM, model parameters are estimated using Bayesian methods. Usually, the parameters are of greatest interest as these describe how connections between brain regions are dependent on experimental manipulations. For a given DCM indexed by , a prior distribution, is specified using biophysical and dynamic constraints [20]. The likelihood, can be computed by numerically integrating the neurodynamic (equation 1) and hemodynamic processes (equation 2). The posterior density is then estimated using a nonlinear variational approach described in [23],[25]. Other Bayesian estimation algorithms can, of course, be used to approximate the posterior density. Reassuringly, posterior confidence regions found using the nonlinear variational approach have been found to be very similar to those obtained using a computationally more expensive sample-based algorithm [26].
This section reviews methods for computing the evidence for a model, , fitted to a single data set . Bayesian estimation provides estimates of two quantities. The first is the posterior distribution over model parameters which can be used to make inferences about model parameters . The second is the probability of the data given the model, otherwise known as the model evidence. In general, the model evidence is not straightforward to compute, since this computation involves integrating out the dependence on model parameters(4)
A common technique for approximating the above integral is the Variational Bayes (VB) approach [27]. This is an analytic method that can be formulated by analogy with statistical physics as a gradient ascent on the ‘negative variational Free Energy’ (or Free Energy for short), , of the system. This quantity is related to the model evidence by the relation [27],[28](5)where the last term in Eq.(5) is the Kullback-Leibler (KL) divergence between an ‘approximate’ posterior density, , and the true posterior, . This quantity is always positive, or zero when the densities are identical, and therefore is bounded below by . Because the evidence is fixed (but unknown), maximising implicitly minimises the KL divergence. The Free Energy then becomes an increasingly tighter lower bound on the desired log-model evidence. Under the assumption that this bound is tight, model comparison can then proceed using as a surrogate for the log-model evidence.
The Free Energy is but one approximation to the model evidence, albeit one that is widely used in neuroimaging [29],[30]. A simpler approximation, the Bayesian Information Criterion (BIC) [11], uses a fixed complexity penalty for each parameter. This is to be compared with the free energy approach in which the complexity penalty is given by the KL-divergence between the prior and approximate posterior [11]. This allows parameters to be differentially penalised. If, for example, a parameter is unchanged from its prior, there will be no penalty. This adaptability makes the Free Energy a better approximation to the model evidence, as has been shown empirically [6],[31].
There are also a number of sample-based approximations to the model evidence. For models with small numbers of parameters the Posterior Harmomic Mean provides a good approximation. This has been used in neuroscience applications, for example, to infer based on spike data whether neurons are responsive to particular features, and if so what form the dependence takes [32]. For models with a larger number of parameters the evidence can be well approximated using Annealed Importance Sampling (AIS) [33]. In a comparison of sample-based methods using synthetic data from biochemical networks, AIS provided the best balance between accuracy and computation time [13]. In other comparisons, based on simulation of graphical model structures [6] the Free Energy method approached the performance of AIS and clearly outperformed BIC. In this paper model evidence is approximated using the Free Energy.
Neuroimaging data sets usually comprise data from multiple subjects as the perhaps subtle cognitive effects one is interested in are often only manifest at the group level. In this and following sections we therefore consider group model inference where we fit models to data from subjects . Every model is fitted to every subjects data. In Fixed Effects (FFX) Analysis it is assumed that every subject uses the same model, whereas Random Effects (RFX) Analysis allows for the possibility that different subjects use different models. This section focusses on FFX.
Given that our overall data set, , which comprises data for each subject, , is independent over subjects, we can write the overall model evidence as(6)Bayesian inference at the model level can then be implemented using Bayes rule(7)Under uniform model priors, , the comparison of a pair of models, and , can be implemented using the Bayes Factor which is defined as the ratio of model evidences(8)Given only two models and uniform priors, the posterior model probability is greater than 0.95 if the BF is greater than twenty. Bayes Factors have also been stratified into different ranges deemed to correspond to different strengths of evidence. ‘Strong’ evidence, for example, corresponds to a BF of over twenty [34]. Under non-uniform priors, pairs of models can be compared using Odds Ratios. The prior and posterior Odds Ratios are defined as(9)resepectively, and are related by the Bayes Factor(10)When comparing two models across a group of subjects, one can multiply the individual Bayes factors (or exponentiate the sum of log evidence differences); this is referred to as the Group Bayes Factor (GBF) [16]. As is made clear in [19] the GBF approach implicitly assumes that every subject uses the same model. It is therefore a Fixed Effects analysis. If one believes that the optimal model structure is identical across subjects, then an FFX approach is entirely valid. This assumption is warranted when studying a basic physiological mechanism that is unlikely to vary across subjects, such as the role of forward and backward connections in visual processing [35].
An alternative procedure for group level model inference allows for the possibility that different subjects use different models. This may be the case in neuroimaging when investigating pathophysiological mechanisms in a spectrum disease or when dealing with cognitive tasks that can be performed with different strategies. RFX inference is based on the characteristics of the population from which the subjects are drawn. Given a candidate set of models, we denote as the frequency with which model is used in the population. We also refer to as the model probability.
We define a prior distribution over which in this paper, and in previous work [19], is taken to be a Dirichlet density (but see later)(11)where is a normalisation term and the parameters, , are strictly positively valued and can be interpreted as the number of times model has been observed or selected. For the density is convex in -space, whereas for it is concave.
Given that we have drawn subjects from the population of interest we then define the indicator variable as equal to unity if model has been assigned to subject . The probability of the ‘assignation vector’, , is then given by the multinomial density(12)The model evidence, , together with the above densities for model probabilities and model assignations constitutes a generative model for the data, (see figure 1 in [19]). This model, can then be inverted to make inferences about the model probabilites from experimental data. Such an inversion has been described in previous work, which developed an approximate inference procedure based on a variational approximation [19] (this was in addition to the variational approximation used to compute the Free Energy for each model). The robustness and accuracy of this method was verified via simulations using data from synthetic populations with known frequencies of competing models [19]. This algorithm produces an approximation to the posterior density on which subsequent RFX inferences are based.
As we shall see in the following section, unbiased family level inferences require uniform priors over families. This requires that the prior model counts, , take on very small values (see equation 24). These values become smaller as the number of models in a family increases. It turns out that although the variational algorithm is robust for , it is not accurate for . This is a generic problem with the VB approach and is explained further in the the supporting material (see file Text S1). For this reason, in this paper we choose to take a Gibbs sampling instead of a VB approach. Additionally, the use of Gibbs sampling allows us to relax the assumption made in VB that the posterior densities over and factorise [19]. Gibbs sampling is the Monte-Carlo method of choice when it is possible to iteratively sample from the conditional posteriors [1]. Fortunately, this is the case with the RFX models as we can iterate between sampling from and . Such iterated sampling eventually produces samples from the marginal posteriors and by allowing for a sufficient burn-in period after which the Markov-chain will have converged [1]. The procedure is described in the following section.
We have so far described procedures for Bayesian inference over models . These models comprise the comparison set, . This section points out a number of generic features of Bayesian model comparison.
First, for any data set there exists an infinite number of possible models that could explain it. The purpose of model comparison is not to discover a ‘true’ model, but to determine that model, given a set of plausible alternatives, which is most ‘useful’, ie. represents an optimal balance between accuracy and complexity. In other words Bayesian model inference has nothing to say about ‘true’ models. All that it provides is an inference about which is more likely, given the data, among a set of candidate models.
Second, we emphasise that posterior model probabilities depend on the comparison set. For FFX inference this can be clearly seen in equation 7 where the denominator is given by a sum over . Similarly, for RFX inference, the dependence of posterior model probabilities on the comparison set can be seen in equation 14. Other factors being constant, posterior model probabilities are therefore likely to be smaller for larger .
Our third point relates to the ranking of models. For FFX analysis the relative ranking of a pair of models is not dependent on . That is, if then for any two comparison sets and that contain models and . This follows trivially from equation 7 as the comparison set acts only as a normalisation term.
However, for group random effects inference the ranking of models can be critically dependent on the comparison set. That is, if then it could be that where is the posterior expected probability of model given comparison set . The same holds for other quantities derived from the posterior over , such as the exceedance probability (see [19] and later). This means that the decision as to which is the best model depends on . This property arises because different subjects can use different models and we illustrate it with the following example.
Consider that comprises just two models and . Further assume that we have subjects and model is preferred by 7 of these subjects and by the remaining 10. We assume, for simplicity, that the degrees of preference (ie differences in evidence) are the same for each subject. The quantity then simply reflects the proportion of subjects that prefer model [19]. So , and for comparison set model 2 is the highest ranked model. Although the differences in posterior expected values are small the corresponding differences in exceedance probabilities will be much greater. Now consider a new comparison set that contains an addditional model . This model is very similar to model such that, of the ten subjects who previously preferred it, six still do but four now prefer model . Again, assuming identical degrees of preference, we now have , and . So, for comparison set model is now the best model. So which is the best model: model one or two?
We suggest that this seeming paradox shows, not that group random effects inference is unreliable, but that it is not always appropriate to ask which is the best model. As is usual in Bayesian inference it is wise to consider the full posterior density rather than just the single maximum posterior value. We can ask what is common to models two and three. Perhaps they share some structural assumption such as the existence of certain connections or other characteristic such as nonlinearity. If one were to group the models based on this characteristic then the inference about the characteristic would be robust. This notion of grouping models together is formalised using family-level inference which is described in the following section. One can then ask: of the models that have this characteristic what are the typical parameter values? This can be addressed using Bayesian Model Averaging within families.
To implement family level inference one must specify which models belong to which families. This amounts to specifying a partition, , which splits S into disjoint subsets. The subset contains all models belonging family and there are models in the th subset.
Different questions can be asked by specifying different partitions. For example, to test model space for the ‘effect of linearity’ one would specify a partition into linear and nonlinear subsets. One could then test the same model space for the ‘effect of seriality’ using a different partition comprising serial and parallel subsets. The subsets must be non-overlapping and their union must be equal to S. For example, when testing for effects of “seriality”, some models may be neither serial or parallel; these models would then define a third subset.
The usefulness of the approach is that many models (perhaps all models) are used to answer (perhaps) all questions. This is similar to factorial experimental designs in psychology [36] where data from all cells are used to assess the strength of main effects and interactions. We now relate the two-levels of inference: family and model.
So far, we have dealt with inference on model-space, using partitions into families. We now consider inference on parameters. Usually, the key inference is on models, while the maximum a posteriori (MAP) estimates of parameters are reported to provide a quantitative interpretation of the best model (or family). Alternatively, people sometimes use subject-specific MAP estimates as summary statistics for classical inference at the group level. These applications require only a point (MAP) estimate. However for completeness, we now describe how to access the full posterior density on parameters, from which MAP estimates can be harvested.
The basic idea here is to use Bayesian model averaging within a family; in other words, summarise family-specific coupling parameters in a way that avoids brittle assumptions about any particular model. For example, the marginal posterior for subject and family is(27)where is our variational approximation to the subject specific posterior and is the posterior probability that subject uses model . We could take this to be under the FFX assumption that all subjects use the same model, or under the RFX assumption that each subject uses their own model (see equation 14).
Finally, to provide a single posterior density over subjects one can define the parameters for an average subject(28)and compute the posterior density from the above relation and the individual subject posteriors from equation 27.
Equation 27 arises from a straightforward application of probability theory in which a marginal probability is computed by marginalising over quantities one is uninterested in (see also equation 4 for marginalising over parameters). Use of equation 27 in this context is known as Bayesian Model Averaging (BMA) [4],[37]. In neuroimaging BMA has previously been used for source reconstruction of MEG and EEG data [9]. We stress that no additional assumptions are required to implement equation 27.
One can make small or large. If we make , the entire model-space, the posteriors on the parameters become conventional Bayesian model averages where . Conversely, if we make , a single model, we get conventional parameter inference of the sort used when selecting the best model; i.e., . This is formally identical to using under the assumption that the posterior model density is a point mass at . More generally, we want to average within families of similar models that have been identified by inference on families.
One can see from equation 27 that models with low probability contribute little to the estimate of the marginal density. This property can be made use of to speed up the implementation of BMA by excluding low probability models from the summation. This can be implemented by including only models for which(29)where is the minimal posterior odds ratio. Models satisfying this criterion are said to be in Occam's window [38]. The number of models in the window, , is a useful indicator as smaller values correspond to peakier posteriors. In this paper we use . We emphasise that the use of Occam's window is for computational expedience only.
Although it is fairly simple to compute the MAP estimates of the Bayesian parameter (MAP) averages analytically, the full posteriors per se have a complicated form. This is because they are mixtures of Gaussians (and delta functions for models where some parameters are precluded a priori). This means the posteriors can be multimodal and are most simply evaluated by sampling. The sampling approach can be implemented as follows. This generates samples from the posterior density . For each sample, , and subject we first select a model as follows. For RFX we draw from(30)where the th element of the vector is the posterior model probability for subject , (we will use the expected values from equation 14). For FFX the model probabilities are the same for all subjects and we draw from(31)where is the vector of posterior model probabilities with th element equal to . For each subject one then draws a single parameter vector, from the subject and model specific posterior(32)These samples can then be averaged to produce a single sample(33)One then generates another sample by repeating steps 30/31, 32 and 33. The samples then provide a sample-based representation of the posterior density from which the usual posterior means and exceedance probabilities can be derived. Model averaging can also be restricted to be within-subject (using equations 30/31 and 32 only). Summary statistics from the resulting within-subject densities can then be entered into standard random effects inference (eg using t-tests) [19].
For any given parameter, some models assume that the parameter is zero. Other models allow it to be non-zero and its value is estimated. The posterior densities from equation 27 will therefore include a delta function at zero, the height of which corresponds to the posterior probability mass of models which assume that the parameter is zero. For the applications in this paper, the posterior densities from equation 27 will therefore correspond to a mixture of delta functions and Gaussians because for DCMs have a Gaussian form. This is reminiscent of the model selection priors used in [39] but in our case we have posterior densities.
We illustrate the methods using neuroimaging data from a previously published study on the cortical dynamics of intelligible speech [17]. This study applied dynamic causal modelling of fMRI responses to investigate activity among three key multimodal regions: the left posterior and anterior superior temporal sulcus (subsequently referred to as regions P and A respectively) and pars orbitalis of the inferior frontal gyrus (region F). The aim of the study was to see how connections among regions depended on whether the auditory input was intelligible speech or time-reversed speech. Full details of the experimental paradigm and imaging parameters are available in [17].
An example DCM is shown in figure 1. Other models varied as to which regions received direct input and which connections could be modulated by ‘speech intelligibility’. Given that each intrinsic connection can be either modulated or not, there are possible patterns of modulatory connections. Given that the auditory stimulus is either a direct input to a region or is not there are possible patterns of input connectivity. But we discount models without any input so this leaves 7 input patterns. The 64 modulatory patterns were then crossed with the 7 input patterns producing a total of different models. These models were fitted to data from a total of subjects (see [17] for details). Overall DCMs were fitted. The next two sections focus on family level inference. As this is a methodological paper we present results using both an FFX and RFX approach (ordinarily one would use either FFX or RFX alone).
Our first family level inference concerns the pattern of input connectivity. To this end we assign each of the models to one of input pattern families. These are family A (models 1 to 64), F (65 to 128), P (129 to 192), AF (193 to 256), PA (257 to 320), PF (321 to 384) and PAF (285 to 448). Family PA, for example, has auditory inputs to both region P and A.
The first two numerical columns of Table 1 show the posterior family probabilities from an FFX analysis computed using equation 21. These are overwhelmingly in support of models in which region P alone receives auditory input (alternative probability ). The last two columns in Table 1 show the corresponding posterior expectations and exceedance probabilities from an RFX analysis computed using equation 25. The conclusions from RFX analysis are less clear cut. But we can say, with high confidence (total exceedance probability, ) that either region A alone or region P alone receives auditory input. Out of these two possibilities it is much more likely that region P alone receives auditory input (exceedance probability ) rather than region A (exceedance probability ). Figure 2 shows the posterior distributions , from an RFX analysis, for each of the model families.
Having established that auditory input most likely enters region we now turn to a family level inference regarding modulatory structure. For this inference we restrict our set of candidate models, , to the 64 models receiving input to region . We then assign each of these models to one of modulatory families. These were specified by first defining a hierarchy with region P at the bottom, A in the middle and F at the top; in accordance with recent studies that tend to place F above A in the language hierarchy [40]. For each structure we then counted the number of forward, , and backward, , connections and defined the following families: predominantly forward (F, ), predominantly backward (B, ), balanced (BAL, ), or None.
The first two numerical columns of Table 2 show the posterior family probabilities from an FFX analysis. We can say, with high confidence (total posterior probability, ) that . The last two columns in Table 2 show the posterior expectations and exceedance probabilities from an RFX analysis. These were computed from the posterior densities shown in Figure 3. The conclusions we draw, in this case, are identical to those from the FFX analysis. That is, we can say, with high confidence (total exceedance probability, ) that .
Family level posteriors are related to model level posteriors via summation over family members according to equation 21 for FFX and equation 22 for RFX. Figure 4 shows the how the posterior probabilities over input families break down into posterior probabilities for individual models. Figure 5 shows the same for the modulatory families.
The maximum posterior model for the input family inference is model number 185 having posterior probability . Given that all families have the same number of members, the model priors are uniform, so the maximum posterior model is also the one with highest aggregate model evidence. This model has input to region P and modulatory connections as shown in Figure 6(a).
The model evidence for the DCMs fitted in this paper was computed using the free energy approximation. This is to be contrasted with previous work in which (the most conservative of) AIC and BIC was used [17]. One notable difference arising from this distinction is that the top-ranked models in [17] contained significantly fewer connections than those in this paper (one sample t-test, ). The top 10 models in [17] contained an average 2.4 modulatory connections whereas those in this paper contained an average of 4.5. This difference reflects the fact that the AIC/BIC approximation to the log evidence penalizes models for each additional connection (parameter) without considering interdependencies or covariances amongst parameters, whereas the free energy approximation takes such dependencies into account.
We now follow up the family-level inferences about input connections with Bayesian model averaging. As previously discussed, this is especially useful when the posterior model density is not sharply peaked, as is the case here (see Figure 4. All of the averaging results in this paper are obtained with an Occam's window defined using a minimal posterior odds ratio of .
For FFX inference the input was inferred to enter region P only. We therefore restrict the averaging to those 64 models in family P. This produces 16 models in Occam's window (itself indicating that the posterior is not sharply peaked). The worst one is with . The posterior odds of the best relative to the worst is only (the largest it could be is ), meaning these models are not significantly better than one another. Four of the models in Occam's window are shown in Figure 6. Figure 7 shows the posterior densities of average modulatory connections (averaging over models and subjects). The height of the delta functions in these histograms correspond to the total posterior probability mass of models which assume that the connection is zero.
For RFX inference the input was inferred to most likely enter region P alone (posterior exceedance probability, ). In the RFX model averaging the Occam's windowing procedure was specific to each subject, thus each subject can have a different number of models in Occam's window. For the input model P family there were an average of models in Occam's window and Figure 8 shows the posterior densities of the average modulatory connections (averaging over models and subjects). Both the RFX and FFX model averages within family P show that only connections from P to A, and from P to F, are facilitated by speech intelligibility.
This paper has investigated the formal comparison of models using Bayesian model evidence. Previous application of the method in the biological sciences has focussed on model selection in which one first selects the model with the highest evidence and then makes inferences based on the parameters of that model. We have shown that this ‘best model’ approach, though useful when the number of models is small, can become brittle if there are a large number of models, and if different subjects use different models.
To overcome this shortcoming we have proposed the combination of two further approaches (i) family level inference and (ii) Bayesian model averaging within families. Family level inference removes uncertainty about aspects of model structure other than the characteristic one is interested in. Bayesian model averaging can then be used to provide a summary measure of likely parameter values for each family.
We have applied these approaches to neuroimaging data, specifically a DCM study of auditory word processing using fMRI. Our results indicate that spoken words most likely stimulate a region in posterior STS and that if the word is intelligible connections are strengthened both to anterior STS and an inferior frontal region. These conclusions were drawn based on family level inference and Bayesian model averaging.
The model evidence for the DCMs fitted in this paper was computed using the free energy approximation whereas previous work used (the most conservative of) AIC and BIC [17]. This resulted in the highly ranked models containing significantly more connections than in the previous study. This is due to a bias in the AIC/BIC criterion which leads to overly simple models being selected. Previous work in graphical models favours the free energy approach over BIC [6] and work on biochemical models finds AIS to be the best of the more computationally expensive sampling methods. The relative merits of the different model selection criteria, as applied to brain imaging models and data, will be addressed in a future publication. The family level inference procedures described in this paper can be applied whatever method is used for estimating the model evidence.
Interestingly, the use of BMA produced an average network structure with speech input to region P, and modulatory connections from P to A and from P to F. This is exactly the winning model from earlier work [17] (based on AIC/BIC approximation of model evidence). It is not, however, the best model as indicated by the free energy. The model with the highest free energy (see figure 6(a)) does not, however, have significantly higher evidence than the second best model, or indeed, any model in Occam's window. This indicates that in the particular example we have studied the use of Bayes factors or posterior odds ratios would be inconclusive, whereas clear conclusions can be drawn from family level inference.
This paper has also introduced a Gibbs sampling method for RFX model level inference when the number of models is large. This sampling method should be preferred to the previously suggested VB method [19] when the number of models exceeds the number of subjects (ie. ). We do emphasise, however, that for RFX model level inferences involving a small number of models (as in previous work [19]) the VB approach is perfectly valid, and is indeed the preferred approach because it is faster.
The issue of family versus model level inference is orthogonal to the issue of random versus fixed effects analysis. The same critera re. FFX versus RFX apply at the family level as at the model level. For the data in this paper one might use RFX analysis as auditory word processing is part of the high level language system and one expect might expect differences in the neuronal instantiation (eg. lateralisation). If the issue remains unclear one could adopt a more pragmatic approach by first implementing a FFX analysis, and if there appear to be outlying subjects, then one could follow this up with an RFX analysis.
Family level inferences under FFX assumptions are simple to implement. Families with (the same and) different numbers of models are accommodated by setting model priors using equation 20, model posteriors are computed using equation 7, and family level posteriors using equation 21. This is a simple non-iterative procedure. Family level inferences under RFX assumptions are more subtle and have been the main focus of this paper. Families with (equal and) unequal numbers of models are accommodated using the model priors in equation 24, model posteriors are computed using an iterative Gibbs sampling procedure, and family level posteriors are computed using equation 22. We envisage that family level inference under RFX assumptions will be particularly useful in neuroimaging studies of high level cognition or for clinical groups where there is a high degree of intersubject variability. Where subjects can be clearly divided into two or more groups on behavioural or other grounds (e.g. patients and controls), then it would be correct to group the models accordingly, and proceed with a between group analysis on selected parameters of the averaged models.
Finally, we comment on the broader issue of comparison of discrete models (the ‘Discrete’ approach adopted in this work) versus a hierarchical approach embodying Automatic Relevance Determination (ARD) in which irrelevant connections are ‘switched off’ during model fitting [41] (for the case of DCMs the ARD approach is currently hypothetical as no such algorithm has yet been implemented). The ARD approach provides an estimate of the marginal density directly without recourse to Bayesian model averaging. The Discrete approach allows for quantitative family-level inferences about issues such as whether processing is serial or parallel, linear or nonlinear. Additionally, Bayesian Model Averaging can be used with the Discrete approach to provide estimates of the marginal density . Overall, the ARD approach is probably the preffered method if one is solely interested in the marginal density over parameters, because it will likely be faster. If one is additionally interested in quantitative family-level inference then the Discrete approach would be the method of choice.
We expect that the comparison of model families will prove useful for a range of model comparison applications in biology, from connectivity models of brain imaging data, to behavioural models of learning and decision making, and dynamical models in molecular biology.
|
10.1371/journal.ppat.1004669 | Essential Domains of Anaplasma phagocytophilum Invasins Utilized to Infect Mammalian Host Cells | Anaplasma phagocytophilum causes granulocytic anaplasmosis, an emerging disease of humans and domestic animals. The obligate intracellular bacterium uses its invasins OmpA, Asp14, and AipA to infect myeloid and non-phagocytic cells. Identifying the domains of these proteins that mediate binding and entry, and determining the molecular basis of their interactions with host cell receptors would significantly advance understanding of A. phagocytophilum infection. Here, we identified the OmpA binding domain as residues 59 to 74. Polyclonal antibody generated against a peptide spanning OmpA residues 59 to 74 inhibited A. phagocytophilum infection of host cells and binding to its receptor, sialyl Lewis x (sLex-capped P-selectin glycoprotein ligand 1. Molecular docking analyses predicted that OmpA residues G61 and K64 interact with the two sLex sugars that are important for infection, α2,3-sialic acid and α1,3-fucose. Amino acid substitution analyses demonstrated that K64 was necessary, and G61 was contributory, for recombinant OmpA to bind to host cells and competitively inhibit A. phagocytophilum infection. Adherence of OmpA to RF/6A endothelial cells, which express little to no sLex but express the structurally similar glycan, 6-sulfo-sLex, required α2,3-sialic acid and α1,3-fucose and was antagonized by 6-sulfo-sLex antibody. Binding and uptake of OmpA-coated latex beads by myeloid cells was sensitive to sialidase, fucosidase, and sLex antibody. The Asp14 binding domain was also defined, as antibody specific for residues 113 to 124 inhibited infection. Because OmpA, Asp14, and AipA each contribute to the infection process, it was rationalized that the most effective blocking approach would target all three. An antibody cocktail targeting the OmpA, Asp14, and AipA binding domains neutralized A. phagocytophilum binding and infection of host cells. This study dissects OmpA-receptor interactions and demonstrates the effectiveness of binding domain-specific antibodies for blocking A. phagocytophilum infection.
| Anaplasma phagocytophilum causes the potentially deadly bacterial disease granulocytic anaplasmosis. The pathogen replicates inside white blood cells and, like all other obligate intracellular organisms, must enter host cells to survive. Multiple A. phagocytophilum surface proteins called invasins cooperatively orchestrate the entry process. Identifying these proteins’ domains that are required for function, and determining the molecular basis of their interaction with host cell receptors would significantly advance understanding of A. phagocytophilum pathogenesis. In this study, the binding domains of two A. phagocytophilum surface proteins, OmpA and Asp14, were identified. The specific OmpA residues that interact with its host cell receptor were also defined. An antibody cocktail generated against the binding domains of OmpA, Asp14, and a third invasin, AipA, blocked the ability of A. phagocytophilum to infect host cells. The data presented within suggest that binding domains of OmpA, Asp14, and AipA could be exploited to develop a vaccine for granulocytic anaplasmosis.
| Human granulocytic anaplasmosis (HGA) is an emerging tick-borne zoonosis in the United States, Europe, and Asia [1]. The number of HGA cases reported to the U. S. Centers for Disease Control and Prevention rose nearly seven-fold between 2003 and 2012 [2,3]. Seroprevalence data indicate that the disease is underreported in some endemic regions [4–8]. HGA can also be spread via perinatal, nosocomial, and blood transfusion routes [6,9–13]. It is an acute illness characterized by fever, chills, headache, malaise, leukopenia, thrombocytopenia, and elevated liver enzymes. Complications can include shock, seizures, pneumonitis, rhabdomyolysis, hemorrhage, increased susceptibility to secondary infections, and death. Risk for complications and fatality is greater for the elderly, the immunocompromised, and when proper diagnosis and/or antibiotic therapy are delayed [1]. The causative agent of HGA is Anaplasma phagocytophilum, an obligate intracellular bacterium that exhibits a tropism for neutrophils [1]. A. phagocytophilum is carried by a variety of wild animal reservoirs and, in addition to humans, causes disease in domestic animals including dogs, cats, horses, and sheep [14].
A. phagocytophilum exhibits a biphasic developmental cycle similar to that of Chlamydia spp., Ehrlichia spp., and Coxiella burnetii [15–18]. The A. phagocytophilum infectious dense-cored (DC) form promotes its receptor-mediated uptake into a host cell-derived vacuole. Within its vacuole, the DC develops into the non-infectious reticulate cell (RC) form that replicates to form a bacterial cluster called a morula [18,19]. RCs then convert back to DCs and are released to initiate the next infection cycle [18].
Sialyl Lewis x ([NeuAcμ(2–3)Galβ1–4(Fucα1–3)GlcNac]; sLex), an α2,3-sialylated and α1,3-fucosylated core-2 O-linked glycan that caps the N-termini of selectin ligands [20], is a critical A. phagocytophilum receptor [21]. sLex is richly expressed on mammalian cells that are permissive for A. phagocytophilum infection—neutrophils, bone marrow progenitors, and promyelocytic HL-60 cells [22–24]. A. phagocytophilum recognizes sLex that caps the N-terminus of P-selectin glycoprotein ligand-1 (PSGL-1) on these myeloid cells [21,25]. Neutrophils and HL-60 cells that have been treated with an sLex blocking antibody, from which surface sialic acids have been enzymatically removed, or that are devoid of sialyltransferase and/or α1,3-fucosyltransferase activity are resistant to A. phagocytophilum binding and infection [19,21,26,27]. A. phagocytophilum also infects rhesus monkey choroidal (RF/6A) endothelial cells, megakaryoblastic MEG-01 cells, and bone marrow-derived mast cells in tissue culture. Infection of these non-myeloid host cell types depends on sLex itself, α2,3-sialic acid, and/or α1,3-fucose [28–35]. Thus, sLex and possibly other closely related α2,3-sialylated and α1,3-fucosylated molecules are essential for efficient A. phagocytophilum infection of mammalian cells.
We identified A. phagocytophilum OmpA and α2,3-sialic acid (N-acetylneuraminic acid [Neu5Ac], further referred to as sialic acid throughout) as the bacterium’s first adhesin/invasin-receptor pair [19]. OmpA binding to the α2,3-sialic acid determinant of sLex on myeloid cells and to α2,3-sialylated glycans on RF/6A cells are vital steps in A. phagocytophilum invasion of these host cell types [19]. Exposure of OmpA on the A. phagocytophilum DC surface makes it accessible to antibodies [19], which could be used to exploit the bacterium’s obligatory intracellular nature to block the host cell invasion step that is essential for survival. The OmpA binding domain that recognizes α2,3-sialic acid lies within amino acids 19 to 74 [19], but has yet to be specifically identified. The A. phagocytophilum OMP that recognizes α1,3-fucose is unknown. OmpA functions in concert with two additional invasins that are also upregulated during tick transmission feeding, Asp14 (14-kDa A. phagocytophilum surface protein) and AipA (A. phagocytophilum invasion protein A), to promote optimal A. phagocytophilum entry into mammalian host cells [29,36]. Thus, the most effective anti-granulocytic anaplasmosis approach may require targeting of all three invasins. We defined the AipA binding domain as residues 9 to 21 [36]. Pinpointing the OmpA and Asp14 binding domains; dissecting the interactions of key OmpA amino acids with α2,3-sialic acid and potentially α1,3-fucose; and evaluating the efficacy of targeting the OmpA, Asp14, and AipA binding domains together would potentially benefit development of approaches to block A. phagocytophilum infection.
In this study, we used antibody blocking, in silico docking models, and site-directed mutagenesis to identify the A. phagocytophilum OmpA binding domain, specifically the key residues that are essential for its adhesin/invasin activity, and determined that it recognizes both α2,3-sialic acid α1,3-fucose. This work represents the most detailed study of any rickettsial adhesin/invasin-receptor pair to date. Furthermore, we identified the Asp14 binding domain and confirmed that an antibody cocktail targeting the binding domains of OmpA, Asp14, and AipA nearly abolishes A. phagocytophilum infection of host cells.
The OmpA region that is important for A. phagocytophilum infection of mammalian host cells lies within residues 19 to 74 (OmpA19–74) [19]. As a first step in further delineating the binding domain, we raised polyclonal antisera against peptides corresponding to OmpA amino acids 23 to 40, 41 to 58, and 59 to 74. We verified that the antisera were specific for OmpA by confirming that each recognized recombinant forms of mature OmpA (minus the signal sequence; corresponding to residues 19 to 205 and hereafter referred to as OmpA) and OmpA19–74, but neither OmpA75–205 nor Asp14 (S1 Fig. and Fig. 1A). Anti-OmpA41–58 and anti-OmpA59–74 were specific for their target peptides at all serum dilutions. Anti-OmpA23–40 was specific for its target peptide at most dilutions tested, but exhibited low level recognition of OmpA41–58 at dilutions below 1:12,800 (S1 Fig. and Fig. 1A). Next, we evaluated if any of the OmpA peptide antisera could inhibit A. phagocytophilum infection of host cells. Bacteria that had been treated with anti-OmpA or preimmune serum served as positive and negative controls, respectively. As previously observed [19], OmpA antibody reduced the percentage of A. phagocytophilum infected HL-60 cells by approximately 40% (Fig. 1B). OmpA59–74 antibody exhibited a dose-dependent inhibitory effect and, at a concentration of 200 ug/ml, reduced the percentage of infected HL-60 cells by approximately three-fold. Antisera targeting OmpA residues 23 to 40 and 41 to 58 exhibited very little to no inhibition of infection, regardless of concentration. Unless otherwise specified, all antisera were used at a concentration of 200 ug/ml in subsequent blocking experiments.
sLex-capped PSGL-1 is an A. phagocytophilum receptor on human myeloid cells [21,25], and OmpA has been shown to bind the sLex portion [19]. Because OmpA59–74 antibody significantly inhibited A. phagocytophilum infection of HL-60 cells, we rationalized that OmpA amino acids that are critical for engaging the receptor are within residues 59 to 74. To test our hypothesis, we assessed the abilities of antisera targeting various portions of OmpA to interfere with A. phagocytophilum binding to Chinese hamster ovary cells transfected to express sLex-capped PSGL-1 (PSGL-1 CHO cells) [37]. These cells are useful models for studying A. phagocytophilum interactions with sLex and/or PSGL-1 because they robustly support bacterial binding but not infection, while untransfected CHO [CHO (-)] cells that lack expression of these receptors poorly support bacterial binding [18,19,26,27,29]. Anti-OmpA59–74 reduced the mean number of bound A. phagocytophilum DC organisms per PSGL-1 CHO cell by approximately four-fold to nearly that of CHO (-) cells (Fig. 1C). Anti-OmpA reduced bacterial binding to PSGL-1 CHO cells by approximately two-fold. Anti-OmpA23–40, anti-OmpA41–58, and preimmune serum had no effect. These results indicate that the OmpA binding domain lies within amino acids 59 to 74 and this region is important for A. phagocytophilum recognition of sLex-capped PSGL-1.
To complement our antibody blocking experiments, molecular modeling and docking was used to identify the OmpA amino acids that possibly contact sLex. First, a three-dimensional model of the invasin was generated. A crystal structure for A. phagocytophilum OmpA has yet to be determined, but an abundance of crystal structures for similar bacterial proteins have. The Phyre2 (Protein Homology/ Analogy Recognition Server version 2.0) server (www.sbg.bio.ic.ac.uk/phyre2), which predicts three-dimensional structures for protein sequences and threads the predicted models on known crystal structures [38], was used to generate a tertiary structure model for OmpA (Fig. 2A). The resulting homology model predicted that OmpA residues 59 to 74 form part of a surface-exposed alpha helix (Fig. 2A), which could potentially interact with ligands. Surface electrostatic values calculated using the adaptive Poisson-Boltzmann solver (APBS) [39] plugin for PyMOL (pymol.org/educational) indicated that OmpA amino acids 19 to 74 have an overall cationic surface charge. The rest of the modeled protein exhibits an overall anionic surface charge (Fig. 2C). These findings are consistent with prior observations that bacterial and viral proteins that interact with sLex and/or sialic acid do so at cationic surface patches [40–44].
For docking predictions, the sLex glycan (Fig. 2B) was extracted from the crystal structure of sLex-capped PSGL-1 (DOI:10.2210/pdb1g1s/PDB). Autodock Vina was used to predict how OmpA might interact with sLex [45,46]. The search grid encapsulated OmpA19–74 (Fig. 2A). The top two docking models, each with the same predicted affinity value of -4.2 kcal/mol, displayed similar interactions between sLex and the OmpA region encompassed by amino acids 59 to 74. In both models, K64 of OmpA was predicted to bind the α2,3-sialic acid residue of sLex (Fig. 2, D and E). G61 was also predicted to interact with sLex in both models, though it was predicted to bind α2,3-sialic acid in one model and α1,3-fucose in the other. Lastly, K60 was predicted to bind the ß1,3-galactose residue of sLex in the docking model presented in Fig. 2E. Together, the in silico predictions and peptide antibody blocking results suggest that OmpA59–74 contains critical residues that interact with sLex to promote A. phagocytophilum infection of host cells.
Aligning the OmpA sequence from the A. phagocytophilum NCH-1 strain that we study, which was originally isolated from a HGA patient in Nantucket, MA [47], with those encoded by geographically diverse A. phagocytophilum isolates that had been recovered from infected humans, animals, and ticks [48–54] revealed that OmpA is highly conserved among these strains (S2A Fig.). Eight of the nine sequences were identical. The OmpA of NorV2 Norwegian sheep isolate [53] had only three amino acid differences, none of which were within the binding domain encompassed by residues 59 to 74. The high degree of OmpA sequence conservation further supports the invasin’s importance to A. phagocytophilum pathobiology. We next aligned NCH-1 OmpA residues 19 to 74 with corresponding regions of OmpA homologs from A. marginale and Ehrlichia spp., which are in the family Anaplasmataceae with A. phagocytophilum and infect bovine erythrocytes and human and animal monocytes, respectively [55–57]. A. phagocytophilum OmpA K64 that was predicted to potentially interact with sLex (Fig. 2, D and E), was the only binding domain residue that was conserved among all Anaplasmataceae OmpA regions examined (S2B Fig.). Additional residues within the A. phagocytophilum OmpA binding domain, including the other two predicted to interact with sLex, K60 and G61 (Fig. 2, D and E), were conserved among Anaplasma spp. but not Ehrlichia spp. OmpA proteins (S2B Fig.).
Because A. phagocytophilum is an obligate intracellular bacterium, developing a knock out-complementation system for this organism has proved challenging and has not been described. Therefore, we utilized a series of alternative approaches to further functionally evaluate OmpA. Recombinant OmpA can be used as a competitive agonist to block A. phagocytophilum access to its receptor and thereby inhibit infection [19]. We exploited this phenomenon to further define the OmpA amino acids that are critical for receptor recognition and bacterial uptake by assessing the competitive agonist abilities of OmpA proteins having site-directed amino acid changes. Our approach was built on the rationale that OmpA proteins in which the binding domain was disrupted would be unable to inhibit infection. First, we generated OmpA proteins N-terminally fused to glutathione-S-transferase (GST), each of which had an insertion of the peptide CLNHL at one of six different sites within residues 19 to 78. This approach has been used in previous studies to disrupt proteins’ binding domains without perturbing overall protein structure, and the insertion sequence that we devised for this purpose was a consensus of the insertion peptides used in those studies [58–60]. Incubating HL-60 cells with the positive control, GST-OmpA, prior to the addition of DC bacteria resulted in a significant reduction in the percentage of infected cells relative to GST alone (Fig. 3A), as shown previously [19]. GST-OmpA proteins carrying insertions between residues 67 and 68 and between 62 and 63 were completely and partially abrogated, respectively, in their abilities to inhibit A. phagocytophilum infection. GST-OmpA proteins bearing insertions at other sites were unaffected in their ability to inhibit infection.
We next set out to identify the specific amino acids of GST-OmpA that were critical for it to inhibit A. phagocytophilum infection. We repeated the competitive agonist assay using GST-OmpA proteins in which select amino acids had been mutated to alanine (S2 Fig. and Fig. 3, B and C). Many of the targeted residues were within OmpA amino acids 59 to 74. R32 and D53 were selected because they lie outside of residues 59 to 74, and, accordingly, we anticipated that substituting them would not alter OmpA function. GST-OmpAK64A was considerably reduced in its ability to inhibit A. phagocytophilum infection (Fig. 3, B and C), thereby indicating that this highly conserved residue was critical for GST-OmpA to serve as a competitive agonist. K65, however, was dispensable for this function, as the blocking ability of GST-OmpAK65A was uncompromised and the blocking ability of GST-OmpAKK6465AA was no greater than that of GST-OmpAK64. GST-OmpAG61A displayed a modest but significant decline in its competitive agonist ability. Replacement of both G61 and K64 with alanines yielded an additive effect that was greater than substituting either residue alone, as GST-OmpAGK6164AA was unable to inhibit infection. GST-OmpA proteins in which R32, D53, K60, E69 and E72 had been mutated to alanine were each unaffected in the ability to hinder infection.
Given that K64 and G61 are vital and contributory, respectively, to the ability of recombinant OmpA to competitively inhibit A. phagocytophilum infection, we evaluated if these residues mediate binding to mammalian host cell surfaces. RF/6A and HL-60 cells were incubated with His-tagged OmpA proteins. After unbound proteins were washed away, bound proteins were detected by flow cytometry using a His-tag antibody. His-tagged OmpA and OmpAD53A bound equally well to RF/6A cells (Fig. 3D). His-OmpAG61A bound poorly, His-OmpAK64A even more so, and His-OmpAGK6164AA could not bind to host cells. Collectively, these data are consistent with the invasin-receptor contacts predicted by the OmpA-sLex docking models and underscore the importance of OmpA K64 and G61 to OmpA-receptor interactions.
α1,3-fucose is critical for A. phagocytophilum to bind PSGL-1-modeled glycopeptides, to bind and invade human and murine myeloid cells, and to establish infection in laboratory mice [25–27]. Consistent with these observations, PSGL-1 CHO cells that had been pretreated with α1,3/4-fucosidase were approximately three-fold less permissive for A. phagocytophilum binding (S3A Fig.). Multiple lines of evidence led us to hypothesize that OmpA binds α1,3-fucose. First, OmpA binds α2,3-sialic acid [19], which is in close proximity to α1,3-fucose on sLex [46]. Second, the docking model in Fig. 2E predicted that OmpA residues within the binding domain contact both α2,3-sialic acid and α1,3-fucose of sLex. Third, OmpA is important for A. phagocytophilum infection of not only myeloid, but also endothelial cells [19]. Fourth, fucose residues are critical for the pathogen to invade RF/6A endothelial cells, as pretreatment of the host cells with α1,3/4-fucosidase made them significantly less permissive to A. phagocytophilum binding (S3B Fig.) and infection (S3C Fig.).
To determine if OmpA recognizes fucose, His-tagged OmpA was incubated with RF/6A cells that had been treated with α1,3/4-fucosidase and binding was assessed by immunofluorescence microscopy and flow cytometry. α2,3/6-sialidase-treated RF/6A cells were included as a positive control for a treatment that would make the host cells less permissive to recombinant OmpA binding [19]. To verify the efficacy and specificity of both glycosidases, treated and untreated host cells were screened with AAL (Aleuria aurantia lectin) and MAL II (Maackia amurensis lectin II). AAL recognizes fucose residues that are in α1,3- and α1,6-linkages with N-acetylglucosamine [61,62]. MAL II detects sialic acids that are in α2,3-linkages with galactose [63]. Fucosidase treatment abolished AAL but not MAL II binding, while sialidase treatment eliminated MAL II but not AAL binding (Fig. 4, A to D). Thus, the glycosidases were effective and specific. His-OmpA binding to both sialidase- and fucosidase-treated RF/6A cells was comparably reduced relative to vehicle control treated cells, while binding of His-Asp14 was unaffected. Incubating the host cells with MAL II or AAL prior to the addition of His-OmpA competitively reduced the efficiency of His-OmpA binding by similar degrees as sialidase or fucosidase, respectively (Fig. 4E). Overall, these observations demonstrate that optimal adhesion of OmpA to host cells involves both α2,3-sialic acid and α1,3-fucose and that Asp14 utilizes neither sialic acid nor fucose to bind to host cells.
Because His-OmpA binding to RF/6A cells involved recognition of α2,3-sialic acid and α1,3-fucose (Fig. 4), we hypothesized that OmpA interacts with sLex or a sLex-like receptor on these host cells. sLex and the sLex-like molecule, 6-sulfo sLex (Neu5Ac(α2–3)Gal(β1–4)[Fuc(α1–3)][HSO3(3–6)]GlcNAc1) (Fig. 5A) have both been detected on the surfaces of high endothelial venal, vascular, cancerous, and/or inflamed endothelial cells [64–76]. To assess if either glycan is present on RF/6A cells, we screened them with sLex antibodies, CSLEX1 [77] and KM93 [78] and the 6-sulfo-sLex antibody, G72 [64]. Robust G72 signal but little to no CSLEX1 or KM93 signal was detected on RF/6A cells (Fig. 5, B and C). Binding of His-OmpA to RF/6A cells that had been pretreated with G72 was pronouncedly reduced relative to cells that had been incubated with CSLEX1, KM93, or isotype control antibody (Fig. 5D). Thus, A. phagocytophilum OmpA recognizes 6-sulfo-sLex on RF/6A endothelial cells.
The ability of recombinant OmpA to bind to non-phagocytic RF/6A endothelial cells [19] (Figs. 3–5), suggests that, in addition to functioning as an invasin, it may also exhibit adhesin activity. Furthermore, while OmpA on the A. phagocytophilum surface acts cooperatively with Asp14 and AipA to mediate bacterial binding to and invasion of mammalian host cells [19,29,36], its ability to mediate these processes by itself is unknown. Therefore, we assessed the ability of recombinant OmpA to confer adhesiveness and invasiveness to inert particles. His-OmpA was coupled to red fluorescent microspheres that were 1.0 μm in diameter, a size similar to that of the diameter of an A. phagocytophilum DC organism (0.8 ± 0.2 μm) [18]. Successful conjugation of His-OmpA to the beads was confirmed by immunofluorescence using OmpA antiserum (Fig. 6A). RF/6A cells were incubated with recombinant OmpA-coated or non-coated control beads and screened with OmpA antibody to determine the numbers of beads bound per cell. To assess bead uptake, the cells were incubated for an additional 1 to 8 h and trypsin was used to remove non-internalized beads prior to screening. Immunofluorescence microscopy revealed that significantly more OmpA coated beads bound to and were internalized by RF/6A cells versus control beads (Fig. 6, B and D). Scanning electron microscopy corroborated these results, as OmpA coated bead were observed bound to and inducing the formation of filopodia-like structures on the surfaces of RF/6A cells or covered by plasma membrane (Fig. 6C). Thus, OmpA alone was sufficient to mediate bead binding to and uptake by non-phagocytic RF/6A endothelial cells.
We next assessed the ability of His-OmpA coated beads to bind and enter HL-60 cells and, if so, whether these processes involve the OmpA myeloid cell receptor, sLex. Scanning electron microscopy revealed that OmpA beads bound to and induced their own uptake into HL-60 cells (Fig. 7A). Relative to the results obtained using RF/6A cells (Fig. 6D), OmpA coated bead binding to HL-60 cells was reduced (Fig. 7B). However, of the OmpA beads that did bind, approximately half of them were internalized (Fig. 7C). Approximately three-fold fewer control beads than OmpA coated beads bound to and were taken in (Fig. 7, B and C). OmpA bead uptake, but not adherence was pronouncedly inhibited when the assay was performed at 4°C versus 37°C (S4 Fig.). Beads coated with OmpAG61A, OmpAK64A, OmpAGK6164AA, and OmpAKK6465AA were significantly compromised in their abilities to bind to and be internalized by HL-60 cells (Fig. 7, B and C). OmpA bead cellular adherence and entry were significantly inhibited and neutralized, respectively, for host cells that had been pretreated with α2,3/6-sialidase or α1,3/4-fucosidase (Fig. 7, D and E). Moreover, the sLex-specific antibody, CSLEX1 significantly reduced binding and blocked internalization of OmpA beads into HL-60 cells (Fig. 7, F and G). KPL-1, an antibody that is specific for and blocks A. phagocytophilum binding to the PSGL-1 N-terminus [25,79,80], did not affect OmpA bead adherence or uptake (Fig. 7, H and I). These data indicate that OmpA coated beads bind and enter myeloid cells in a sLex-dependent manner and require OmpA residues G61 and K64 to optimally do so.
Of the three invasins that cooperatively function to facilitate A. phagocytophilum infection of mammalian host cells [19,29,36], only the binding domain of Asp14 had yet to be defined. Asp14 is a 124-amino acid (13.8 kDa) protein, and its binding domain lies within residues 101 to 124 [29]. To further narrow down this region, antisera were raised against residues 98 to 112 and 113 to 124. Both antisera recognized GST-Asp14, but not GST-Asp141–88 or GST alone (S5 Fig.). Also, antiserum targeting Asp1498–112 but not Asp14113–124 detected GST-Asp141–112 and each antiserum was specific for the peptide against which it had been raised. Next, the abilities of anti-Asp1498–112 and anti-Asp14113–124 to inhibit A. phagocytophilum infection of HL-60 cells were assessed. Incubating DC bacteria with Asp14113–124 antibody reduced the percentages of infected cells in a dose-dependent manner, whereas Asp1498–112 antibody had no effect (Fig. 8A). When used together, antisera against Asp14113–124 and OmpA59–74 reduced A. phagocytophilum by approximately four-fold (Fig. 8B). The observed blocking effect was significantly greater than that achieved with either antiserum alone or when either was paired with antisera that targeted irrelevant regions of OmpA or Asp14. To ensure that the blocking effects achieved by the OmpA59–74 and Asp14113–124 antisera were specific, fragment antigen binding (Fab fragment) portions of OmpA23–40, OmpA41–48, OmpA59–74, Asp1498–112, Asp14113–124, or OmpA59–74 and Asp14113–124 antibodies were prepared and assessed for the ability to inhibit A. phagocytophilum infection of HL-60 cells. Consistent with results obtained using intact antibodies, OmpA59–74 Fab, Asp14113–124 Fab, and the combination thereof achieved the greatest reductions in the percentage of infected cells and morulae per cell (Fig. 8, C and D).
We previously showed that a combination of antisera that had been raised against the entireties of OmpA, AipA, and Asp14 strongly inhibited A. phagocytophilum infection of mammalian host cells [36]. To refine this blocking approach, DC organisms were treated with a cocktail of antibodies specific for OmpA59–74, Asp14113–124, and AipA9–21 prior to incubating the bacteria with HL-60 cells. This antibody combination significantly attenuated infection, reducing the percentage of infected cells and number of morulae per cell by approximately five-fold (Fig. 9, A and B). The reduction in infection achieved using the combination antisera was due to effective blocking of bacterial adhesion to HL-60 cell surfaces, as combination antisera specific for OmpA59–74, Asp14113–124, and AipA9–21 reduced the numbers of bound A. phagocytophilum organisms per cell by more than four-fold relative to the same amount of preimmune serum (Fig. 9C). The observed reductions in bacterial adhesion and infection achieved by targeting all three binding domains were greater than those achieved using (1) antibodies that targeted only one or two of the binding domains and (2) combinations of antibodies against one or two of the binding domains together with antibodies against irrelevant portions of OmpA, Asp14, or AipA. Thus, targeting the OmpA, Asp14, and AipA binding domains together produced a synergistic blocking effect that protects host cells from A. phagocytophilum infection.
This study identified the OmpA and Asp14 binding domains and defined the OmpA residues that are critical for adhesion and invasion. The OmpA binding domain lies within amino acids 59 to 74 and it, like the rest of the protein, is highly conserved among A. phagocytophilum strains known to cause disease in humans and animals. Antibody against OmpA59–74 inhibited bacterial binding to PSGL-1 CHO cells and infection of HL-60 cells. OmpA59–74 is predicted to be a solvent exposed alpha helix and part of a cationic surface patch that binds sLex, an interaction that is similar to those between staphylococcal superantigen-like (SSL) protein family members and sLex. SSL4, SSL5, and SSL11 each use basic residues within cationic surface pockets to interact with α2,3-sialic acid of sLex [40,41,81]. Likewise, other pathogens’ sialic acid binding proteins, including uropathogenic Escherichia coli sialic acid-specific S fimbrial adhesin [82], pertussis toxin of Bordetella pertussis [43], influenza viral neuraminidase [44], canine adenovirus 2 capsid protein [83], and rhesus rotovirus VP4 [42] all use basic residues localized within cationic surface pockets to target sialic acid. The Asp14 binding domain is within amino acids 113 to 124. Antibody specific for Asp14113–124 abrogated bacterial binding and infection of host cells. As Asp14 bears no semblance to any known crystal structure, it could not be modeled. However, from the data presented herein it can be inferred that Asp14 amino acids 113 to 124 are exposed on the surfaces of A. phagocytophilum and the invasin itself.
OmpA K64 is essential for and G61 contributes to the ability of OmpA to bind to mammalian host cells. These experimental findings support the top two OmpA-sLex docking models, both of which predicted the involvement of K64 and G61 in interacting with α2,3-sialic acid and α1,3-fucose of sLex. The actual interactions between OmpA and sLex are likely a hybrid of those predicted by the two docking models because, while both predicted the involvement of K64 and G61, one also predicted the involvement of K60, which was found to be negligible for OmpA to act as a competitive agonist. OmpA K64 and G61 may play functionally conserved roles among members of the family Anaplasmataceae and the genus Anaplasma. K64 is present in all Anaplasma and Ehrlichia spp. OmpA proteins, while G61 is conserved among Anaplasma but not Ehrlichia spp. OmpA proteins. A. marginale agglutinates bovine red blood cells in a sialidase-sensitive manner [84], indicating that it interacts with sialylated glycans on erythrocyte surfaces. Given the similarities between A. phagocytophilum and A. marginale OmpA proteins [19] and the conservation of residues implicated in receptor recognition, it will be worth investigating whether A. marginale OmpA is important for infection of bovine erythrocytes, and, if so, if it involves interactions between conserved OmpA lysine and glycine residues with sialylated glycans. E. chaffeensis OmpA contributes to infection of monocytic cells [85]. Compared to the conservation exhibited among Anaplasma spp. OmpA proteins, A. phagocytophilum and E. chaffeensis OmpA proteins are more divergent in sequence [19], especially in the binding domain, which may contribute to these pathogens’ tropisms for different leukocytes. Still, because of its conservation, the E. chaffeensis OmpA residue that corresponds to A. phagocytophilum OmpA K64 may be involved in binding to a sLex-related glycan on monocytic cells.
Together with α2,3-sialic acid, α1,3-fucose is critical for A. phagocytophilum binding and infection [21,25–27]. OmpA binds α1,3-fucose, as can be inferred from our observations that recombinant OmpA bound poorly to RF/6A endothelial cells from which α1,3/4-fucose residues had been removed or that had been incubated with the α1,3/6-fucose-specific lectin, AAL. The ability of OmpA to bind α2,3-sialic acid and α1,3-fucose is consistent with the close proximity of the two sugar residues to each other in sLex and related glycans and also with OmpA-sLex molecular docking predictions. Yet, RF/6A cells, which support A. phagocytophilum binding and infection [19,28,29,31,33,34,36], express very little to no sLex. Rather, they express 6-sulfo-sLex, which presents α2,3-sialic acid and α1,3-fucose in the same orientation and proximity to each other as sLex. Recombinant OmpA binding to RF/6A cells was significantly reduced in the presence of 6-sulfo-sLex antibody, but not sLex antibodies, thereby supporting that 6-sulfo-sLex is an A. phagocytophilum receptor on these cells. Thus, A. phagocytophilum OmpA interacts with glycans that present α2,3-sialic acid and α1,3-fucose in a similar manner as sLex.
OmpA by itself functions as both an adhesin and an invasin, as demonstrated by the ability of His-OmpA to confer adhesive and internalization capabilities to inert beads. Approximately half of the His-OmpA beads that bound to host cells were internalized, a degree of uptake that was similar to that reported for C. burnetii OmpA coated beads [86]. Twenty-fold more OmpA coated beads bound to RF/6A cells than to HL-60 cells. Similarly, recombinant OmpA binding to RF/6A cells but not to HL-60 cells could be detected by immunofluorescence microscopy and flow cytometry. Nonetheless, the ability of recombinant OmpA to competitively antagonize A. phagocytophilum binding and infection of HL-60 cells demonstrates its ability to bind to the host cells, but it apparently does so at too low an avidity to remain bound during the wash steps associated with sample preparation for the detection methods used. The observed differences in OmpA binding to HL-60 versus RF/6A cells could be due to differences in the levels of sLex and 6-sulfo-sLex on HL-60 and RF/6A cell surfaces or perhaps due to the presence of an additional, undefined OmpA receptor on RF/6A cells. Yet another possibility is that the bacterium binds with a greater avidity to 6-sulfo-sLex than to sLex.
Because of the essential and cooperative roles that OmpA, Asp14, and AipA play in the A. phagocytophilum lifecycle [19,26,27,29,36], blocking their ability to function can prevent both infection and bacterial survival. Moreover, directing the immune response to their binding domains could enhance protective efficacy. In this study, an antibody cocktail specific for the OmpA, Asp14, and AipA binding domains blocked A. phagocytophilum infection of host cells. This finding could potentially pave the way for development of a multi-invasin targeting vaccine that can protect against or treat human and veterinary granulocytic anaplasmosis. The relevance of this work extends to other obligate intracellular pathogens that use multiple invasins, including A. marginale [87], E. chaffeensis [85,88,89], spotted fever rickettsiae [90–94], Chlamydia spp. [95–99], Mycobacterium spp. [100–102], and Orientia tsutsugamushi [103,104], as their survival hinges on their abilities to enter host cells.
Uninfected and A. phagocytophilum infected (NCH-1 strain) HL-60 cells (ATCC CCL-240) and RF/6A cells (ATCC CRL-1790, Manassas, VA) were maintained as previously described [18,28]. CHO (-) and PSGL-1 CHO cells were cultivated as described [105].
pGST-OmpA, which encodes OmpA19–205 N-terminally fused to GST, was previously constructed [19]. Using pGST-OmpA as template and primers from suppl. S1 Table, the QuikChange Lightning (Agilent Technologies, Santa Clara, CA) protocol was used per the manufacturer’s guidelines to perform site-directed insertions and point mutagenesis of the ompA insert sequence. For site directed insertions, a five-amino acid insert sequence (CLNHL) was selected based on previous studies that had successfully employed the linker-scanning method [58,60], which is used to insert peptide “linkers” to disrupt protein binding domains without perturbing overall protein structure. The sequence chosen for the insertion peptide, CLNHL, was a consensus sequence based on the most common amino acids at their respective positions in the insertion peptides used in prior studies [58,60]. The nucleotide sequence, 5'-TGCCTGAACCACCTG-3', which encoded CLNHL, was inserted in the ompA sequence of pGST-OmpA between ompA nucleotides 102 and 103, 162 and 163, 186 and 187, 201 and 202, 216 and 217, and 231 and 232 to yield plasmids that encoded GST-OmpA proteins bearing CLNHL inserts between OmpA amino acids 34 and 35, 54 and 55, 62 and 63, 67 and 68, 72 and 73, and 77 and 78, respectively. Likewise, the QuikChange protocol was used to perform site directed mutagenesis to yield plasmids that encoded GST-OmpA proteins having R32, D53, K60, G61, K64, K65, E69, and/or E72 converted to alanine. GST-OmpA mutants were expressed and purified as previously described [19]. Plasmids encoding His-tagged wild type and site-directed mutant OmpA proteins were generated by amplifying wild type and mutant ompA sequences using primers containing ligase-independent cloning (LIC) tails and annealing the amplicons into the pET46 Ek/LIC vector (Novagen, EMD Millipore, Darmstadt, DE) per the manufacturer’s instructions. His-OmpA proteins were expressed and purified by immobilized metal-affinity chromatography as previously described [106].
To obtain a putative three-dimensional OmpA protein structure, the mature OmpA sequence was threaded onto the solved crystal structures of proteins with similar sequences using the PHYRE2 server (www.sbg.bio.ic.ac.uk/phyre2/html/page.cgi) as previously described [19,38]. Amino acids 19 to 150 (73% of the mature OmpA sequence) were modeled with greater than 90% confidence to known structures for similar proteins (Protein Data Bank [PDB] files 2aiz [Haemophilus influenzae OmpP6 peptidoglycan associated lipoprotein (PAL)], 4g4v [Acinetobacter baumanni PAL], 4b5c [Burkholderia pseudomallei PAL], 3ldt [Legionella pneumophila OmpA], 2kgw [Mycobaterium tuberculosis OmpATb]). The remainder of the protein lacked sufficient homology to any experimentally derived structure, but could be modeled using the Poing method [38], which was performed as part of the Phyre2 analyses. The sLex-PSGL-1 peptide (residues 61 to 77) and the sLex glycan itself was extracted from the solved crystal structure of PSGL-1 (PDB 1G1S) in PyMol (www.pymol.org) and saved as an individual PDB file. Open Babel software was used to convert PDB files to PDBQT (Protein Data Bank, Partial Charge and Atom Type) format in order to perform OmpA-sLex docking analysis [107]. AutoDock Tools software (autodock.scripps.edu/resources/adt) was used to generate the docking output files for both the OmpA protein structure and the sLex ligand. The search location for OmpA was generated in AutoDock Tools by setting a search grid that encompassed OmpA residues 19 to 74 [19]. Molecular docking was performed using AutoDock Vina (http://vina.scripps.edu/) to identify potential points of interaction between OmpA and sLex [45]. The top two OmpA-sLex models generated by AutoDock Vina had the same predicted affinity value of -4.2 kcal/mol and were selected for analysis in PyMol to determine potential points of contact.
To generate antisera specific to the OmpA and Asp14 binding domains, peptides corresponding to OmpA residues 23 to 40, 41 to 58, and 59 to 74 and Asp14 residues 98 to 112 and 113 to 124 were synthesized, conjugated to keyhole limpet hemocyanin, administered to rabbits, and the resulting OmpA and Asp14 peptide-specific sera were affinity-purified by New England Peptide (Gardner, MA). Each peptide antiserum’s specificity for the peptide against which it had been raised and for its protein target was determined by ELISA using the TMB substrate kit (Thermo Scientific, Waltham, MA) following the manufacturer’s instructions or by Western blot analysis as previously described [108]. Mouse anti-AipA peptide antisera have been previously described [36]. sLex antibodies CSLEX1 (BD Biosciences, San Jose, CA) and KM93 (Millipore, Darmstadt, DE) and PSGL-1 N-terminus-specific antibody KPL-1 (BD Biosciences) were obtained commercially. Fab fragments of OmpA and Asp14 peptide-specific antisera were generated using the Fab Preparation Kit (Pierce, Rockford, IL) according to the manufacturer’s instructions. Reiji Kannagi (Aichi Medical University, Nagukute, Aichi, Japan) kindly provided 6-sulfo-sLex antibody, G72. His tag and Alexa Fluor 488-conjugated secondary antibodies and Alexa Fluor 488-conjugated streptavidin were obtained from Invitrogen (Carlsbad, CA). Biotinylated AAL and MAL II were obtained from Vector Labs (Burlingame, CA). Glycosidases used in this study were α2,3/6-sialidase (Sigma-Aldrich, St. Louis, MO) and α1,3/4-fucosidase (Clontech, Mountain View, CA).
The NCH-1 gene sequence for ompA (APH0338) was previously determined [19,29,36]. A Protein BLAST (basic local alignment search tool) [109] search using the NCH-1 OmpA predicted protein sequence as the query was used to identify homologs in other Anaplasmataceae species and in A. phagocytophilum strains HZ [50], HGE1 [54], Dog [53], JM [52], MRK [48,49], CRT35, CRT38 [51], and NorV2 [53], for which the genomes are available [53,110]. All of these strains except for NorV2 had been originally isolated from clinically affected humans and animals. HZ and HGE1 were recovered from human patients in Westchester, NY, USA and Minnesota, USA, respectively [50,54]. The Dog and JM strains were isolated from a dog in Minnesota, USA and a meadow jumping mouse (Zapus hudsonius) in Camp Ripley, MN, USA [52,53]. MRK had been recovered from a horse in California, USA [48,49]. CRT35 and CRT38 are isolates of the A. phagocytophilum Ap-variant 1 strain that were recovered from ticks collected at Camp Ripley, MN, USA [51]. NorV2 is a naturally occurring A. phagocytophilum isolate that was maintained in an experimentally infected lamb, exhibits reduced virulence in sheep, and differs in its 16S rRNA gene sequence when compared to other sheep isolates [53,111]. OmpA sequence alignments were generated using Clustal W [112].
For binding of His- or GST-tagged OmpA proteins to host cells, RF/6A or HL-60 cells were incubated with 4 μM recombinant protein in culture media for 1 h in a 37°C incubator supplemented with 5% CO2 and a humidified atmosphere. To assess for the presence of sLex or 6-sulfo-sLex on RF/6A cell surfaces, the cells were fixed in 4% PFA in PBS for 1 h at room temperature followed by incubation with CSLEX1, KM93, or G72 for 1 h at room temperature. Antibody incubations and washes were performed as described previously [79]. Spinning-disk confocal microscopy using an Olympus BX51 microscope affixed with a disk-spinning unit (Olympus, Center Valley, PA) and/or flow cytometry using a BD FACS Canto II (BD Biosciences) were performed to assess binding of antibodies or His-OmpA proteins to host cell surfaces as previously described [19,29]. In some cases, RF/6A cells were pretreated with α2,3/6-sialidase, α1,3/4-fucosidase, AAL, MAL II, or sLex- or 6-sulfo-sLex-specific antibodies prior to incubation with His-OmpA.
Competitive inhibition assays utilizing recombinant protein or antibody were performed and analyzed by spinning-disk confocal microscopy as previously described [19,29]. To determine if A. phagocytophilum binding to PSGL-1 CHO cells or infection of RF/6A cells involved bacterial binding to host cell surface fucose residues, the host cells were treated with α1,3/4-fucosidase (10 μU/mL) prior to the addition of DC organisms and assessment for bacterial binding or infection as previously described [18,19]. For competitive inhibition assays using antisera raised against OmpA or Asp14 peptides, A. phagocytophilum DC bacteria were incubated with serially diluted concentrations of antiserum. Preimmune rabbit serum (200 μg/mL) was a negative control. Assays using combinations of two or three different OmpA, Asp14, or AipA peptide antibodies were performed using 100 μg/mL per antibody. Preimmune serum (200 μg/mL or 300 μg/mL, based on the combined total of peptide antisera) served as a negative control. Competitive inhibition assays using OmpA and/or Asp14 Fab fragments were performed exactly as described for antisera. Preimmune Fab fragments served as a negative control.
1.8 x 107 red fluorescent sulfate-modified 1.0 μm diameter microfluorospheres (Life Technologies, Carlsbad, CA) were mixed by rotation with 8 μg of His-OmpA, or His-OmpA proteins bearing alanine substitutions, in 400 μL of 50 mM phosphate-buffered saline (PBS) supplemented with 0.9% NaCl at room temperature overnight in the absence of light. The His-OmpA coated beads were centrifuged at 5,000 g for 25 min, followed by three washes in 50 mM PBS. Coated beads were resuspended in 400 μL of 50 mM PBS, 0.9% NaCl, 1% BSA and stored at 4°C until use. To validate that the beads were conjugated with His-OmpA, 1.8x104 of the beads were screened by immunofluorescent microscopy using mouse polyclonal OmpA antisera followed by Alexa Fluor 488-conjugated goat anti-mouse IgG as described [19]. To assess binding to and uptake by HL-60 or RF/6A cells, His-OmpA coated beads or uncoated control beads were resuspended in the appropriate culture medium and added to host cells at a concentration of 500 beads/cell. For adherent RF/6A cells, beads were centrifuged onto the host cells at 1,000 g for 5 min. The cells plus beads were incubated for 1 h at 37°C in a 5% CO2 supplemented humidified incubator followed by washing the cells three times with PBS to remove unbound beads. Non-adherent HL-60 cells were mixed with the beads in suspension, incubated as described above, and three PBS washes were performed intermittently between five-min spins performed at 300 g. To assess binding, the host cells were fixed in 4% paraformaldehyde (PFA) in PBS, mounted with ProLong Antifade Gold gel mounting medium containing 4',6-diamidino-2-phenylindole (DAPI) (Invitrogen), and analyzed by spinning-disk confocal microscopy as previously described [19]. For uptake assays, after the final wash, the host cells were resuspended in culture medium and cultivated for an additional 7 h. The cells were washed three times in PBS, incubated with a 0.25% trypsin solution (Hyclone, Thermo Scientific, Waltham, MA) for 10 min at 37°C to cleave host cell surface proteins and consequently remove non-internalized beads, and washed three times with PBS. HL-60 cells were cytospun onto glass microscope slides and fixed, mounted, and screened as described above. RF/6A cells were added to wells containing coverslips, incubated overnight in a 37°C incubator supplemented with 5% CO2 and a humidified atmosphere to allow the host cells to adhere prior to further processing. To determine if His-OmpA coated bead binding or uptake was temperature sensitive, some experiments were performed at 4°C. To assess the contribution of sLex or PSGL-1 determinants to His-OmpA coated bead binding and uptake, host cells were pretreated with α2,3-sialidase (5 μg/mL), α1,3/4-fucosidase (10 μU/mL), sLex-specific antibody CSLEX1 (10 μg/mL), PSGL-1 N-terminus-specific antibody KPL-1 (10 μg/mL), or vehicle or isotype controls as previously described [19] prior to the bead binding and uptake assays.
Coverslips of RF/6A cells were incubated with OmpA coated or control beads as described above. The coverslips were fixed in 2.0% glutaraldehyde in 0.1 M sodium cacodylate for 1 h at room temperature. The coverslips were subjected to two 10-min washes in 0.1 M sodium cadodylate and fixed in 1.0% osmium tetroxide in 0.1 M sodium cacodylate for 1 h. The coverslips were rinsed two more times with 0.1 M sodium cadodylate buffer for 10 min each. The samples were dehydrated by successive 5-min incubations in 50% ethanol, 70% ethanol, 80% ethanol, 95% ethanol, and three 10-min washes in 10% ethanol. Next, the samples were incubated three times for 30 min each in hexamethyldisilazane, air-dried, mounted with silver paint, and sputter coated with gold before imaging on a Zeiss EVO 50XVP scanning electron microscope (Thornwood, NY). For HL-60 cells incubated in suspension with beads, the samples were retained on a 0.1 μm filter and processed exactly as described for RF/6A cells.
The Prism 5.0 software package (Graphpad, San Diego, CA) was used to determine the statistical significance of data using one-way analysis of variance (ANOVA) or the Student’s T-test, as previously described [19]. Statistical significance was set to P < 0.05.
|
10.1371/journal.pntd.0003646 | Trypanosoma cruzi, Etiological Agent of Chagas Disease, Is Virulent to Its Triatomine Vector Rhodnius prolixus in a Temperature-Dependent Manner | It is often assumed that parasites are not virulent to their vectors. Nevertheless, parasites commonly exploit their vectors (nutritionally for example) so these can be considered a form of host. Trypanosoma cruzi, a protozoan found in mammals and triatomine bugs in the Americas, is the etiological agent of Chagas disease that affects man and domestic animals. While it has long been considered avirulent to its vectors, a few reports have indicated that it can affect triatomine fecundity. We tested whether infection imposed a temperature-dependent cost on triatomine fitness. We held infected insects at four temperatures between 21 and 30°C and measured T. cruzi growth in vitro at the same temperatures in parallel. Trypanosoma cruzi infection caused a considerable delay in the time the insects took to moult (against a background effect of temperature accelerating moult irrespective of infection status). Trypanosoma cruzi also reduced the insects’ survival, but only at the intermediate temperatures of 24 and 27°C (against a background of increased mortality with increasing temperatures). Meanwhile, in vitro growth of T. cruzi increased with temperature. Our results demonstrate virulence of a protozoan agent of human disease to its insect vector under these conditions. It is of particular note that parasite-induced mortality was greatest over the range of temperatures normally preferred by these insects, probably implying adaptation of the parasite to perform well at these temperatures. Therefore we propose that triggering this delay in moulting is adaptive for the parasites, as it will delay the next bloodmeal taken by the bug, thus allowing the parasites time to develop and reach the insect rectum in order to make transmission to a new vertebrate host possible.
| Parasites are often assumed to cause little harm to their arthropod vectors, even though they commonly reproduce inside the arthropods and exploit their nutrients, even causing lesions when crossing internal barriers. Thus, the interests of parasite and vector may well not be aligned and we can expect the parasite to exploit its vector just as it does with its main host, with consequent negative effects on the vector’s fitness. Here, we show that this occurs with Trypanosoma cruzi in its bug vector (T. cruzi causes Chagas disease, affecting ca. 8 million people and disease management is principally attained via vector control). Our results indicate that the parasites delay insect moulting, which is likely beneficial to them as they need time to develop in the insect before the next bloodmeal (that only occurs post-moult). We also show parasite-induced mortality over the narrow range of temperatures which the insect prefers and over which it performs best. In vitro growth of the parasite increases with temperature and we discuss how this may help explain the effects in vivo. Overall, these results will be important to understand the epidemiology of Chagas disease and provide an evolutionary context to explain the parasite′s interaction with its vector.
| A long-standing implicit assumption in the literature on vector-borne diseases is that the parasite does little or no harm to its vector (see [1] for a review). This makes considerable intuitive sense as the parasite relies on the vector for its transmission, so negative effects on the vectors’ fitness could be expected to reflect negatively on the parasites’ fitness. This was perhaps best formulated (verbally rather than mathematically) in Ewald’s classic treatise on the evolution of virulence [2]. With the rapid development of theory on the evolution of virulence in recent years [3], it has become clear that the vector should to a large degree be considered an alternative host for the parasite, one in which a certain degree of host exploitation (and consequent virulence to this ‘host’) is to be expected [1]. Meanwhile, empirical studies that are aimed at detecting fitness effects of parasite infection have become more refined, looking beyond fecundity and mortality to hunt for more subtle life history or behavioral effects, for example. This can be seen particularly in studies of mosquito (Culicidae) infection with pathogens, such as negative effects of dengue virus on fecundity and oviposition success in Aedes [4]. Perhaps the most elegant demonstration that the interests of parasite and vector are not entirely aligned is parasite-induced increases in biting rates in mosquitoes [5–9], sand flies [10] and tsetse flies [11,12]—this is likely to increase transmission (and thereby fitness) of the parasite while the vector is liable to suffer a reduction in fitness due to excessive energy expenditure and increased risk of mortality when attempting to bite. Meanwhile, evidence of an interplay between parasite and vector strategies towards one another can be seen in the case of several parasites of plants that are transmitted by insect vectors. In several systems where parasite and vector are believed to have shared a coevolutionary history, the parasite increases its vector’s fitness indirectly via effects on the host plant (e.g. [13,14]). This positive interaction is illustrative as the vector will likely spend several generations on the main host (the plant), a situation very different from most vectors of parasite diseases of humans that interact only briefly with the main hosts and in which negative effects can be expected.
For vector-borne diseases of humans, such considerations are of great importance for vector management, especially when novel technologies are under consideration. In strategies such as the release of transgenic vectors, paratransgenesis or use of biocontrol agents that interfere with transmission, the life history and behavior of the vector are key factors [15,16], as are possible evolutionary responses of vector and parasite [17]. It is vital, then, to understand how vector and parasite interact in terms of their respective fitnesses and possible patterns of selection. Chagas disease is one such example. Trypanosoma cruzi is a digenetic protozoan that infects mammals and triatomines in the Americas. As a result of anthropic activities this enzootic infection affects man and domestic animals, causing to the first a disease with different levels of pathology. As a comparatively recently described disease (Chagas disease was first described by Carlos Chagas in 1909) research has focused on interactions between the parasite and man, with little consideration of parasite effects on the invertebrate hosts. Further, as earlier studies showed no parasite-induced alterations in triatomine physiology [18], the parasite has long been considered avirulent to its vectors [19–21]. Few studies showing alterations on fecundity rates of infected females have been conducted [22, 23]. Furthermore, our group has recently shown that T. cruzi affects fecundity and fertility rates of R. prolixus depending on the temperature at which insects are raised [24].
We sought, then, to investigate how T. cruzi might affect its triatomine hosts. As the parasite does not invade the insects’ body but develops rather in its intestine, we might expect effects on the insects’ fitness to be marginal. Previous studies showed no effect of T. cruzi on the development of Nocardia sp. and Rhodococcus rhodnii, gut symbionts of Triatoma infestans and Rhodnius prolixus, respectively [25]. However, as a consequence of living only in the insect intestinal tract, T. cruzi probably competes with its host for nutritional resources. In addition, most laboratory studies of T. cruzi-triatomine interactions have evaluated fitness parameters under conditions that aim to maximize vector development and survival. Changes in mortality rates in mosquitoes under glucose deprivation have been demonstrated for Plasmodium [26,27] and dengue virus infections [4]. Therefore, our prediction is that the parasite might have negative effects on its host’s fitness under less than optimal (and therefore more realistic) environmental conditions [28].
In addition, temperature is a factor of particular importance in host-parasite interactions, especially when the host is ectothermic. It can be a key factor in determining whether a host-parasite interaction eventually favors host or parasite, while in some instances the nature of the interactions can only really be understood by observing the host-parasite interaction under different thermal conditions [29, 30]. We therefore chose to conduct our study under four thermal regimes, and to use a comparatively narrow range of temperatures to keep the test conservative.
All experiments using live animals were performed in accordance with FIOCRUZ guidelines on animal experimentation and were approved by the Ethics Committee in Animal Experimentation (CEUA/FIOCRUZ) under the approved protocol number L-058/08. The protocol is from CONCEA/MCT (http://www.cobea.org.br/), which is associated with the American Association for Animal Science (AAAS), the Federation of European Laboratory Animal Science Associations (FELASA), the International Council for Animal Science (ICLAS) and the Association for Assessment and Accreditation of Laboratory Animal Care International (AAALAC).
Rhodnius prolixus used in assays were obtained from a laboratory colony which is derived from insects collected in Honduras around 1990. The colony was maintained by the Vector Behaviour and Pathogen Interaction Group in Centro de Pesquisas René Rachou, FIOCRUZ, Brazil. Insects were reared at 26 ± 1°C and relative humidity of 65 ± 10%, with natural illumination. They were fed on chicken and mice anesthetized with an intraperitoneal injection of a ketamine (150 mg/kg; Cristália, Brazil)/xylazine (10 mg/kg; Bayer, Brazil) mixture. When insects were infected, they were fed on an artificial feeder containing a suspension of freshly collected and inactivated human blood (56°C/30min) [24], using standard aseptic procedures. Therefore, the parasites did not enter into contact with the anesthetic mixture. Note also that as a routine procedure, T. cruzi cultures were checked for bacterial contamination in every passage under the microscope. Therefore, these procedures assured no bacterial contamination in the blood or T. cruzi cultures.
The T. cruzi used was ‘CL’ strain, originally isolated from naturally-infected T. infestans from southern Brazil [31] and subsequently kept in laboratory cultures. Epimastigote forms were cultured at 27°C in liver infusion tryptose (LIT) medium supplemented with 15% fetal bovine serum, 100mg/ml streptomycin and 100units/ml penicillin. Parasite passages were performed twice a week, i.e. ca. once every three days. As it has been shown that trypanosomes tend to lose infectivity if they are not frequently exposed to hosts [32], parasites were passed through mice and triatomines every 6 months. Briefly, 5th instar nymphs were infected with culture epimastigotes through artificial feeding. One month after infection, these insects were fed and their urine containing metacyclic trypomastigotes was collected and inoculated into a Swiss mouse. Two weeks after inoculation the parasites were recovered by cardiac puncture and used to perform a hemoculture. For infection assays, 50–100μl of culture were washed in sterile PBS (0.15M NaCl, 0.01M sodium phosphate, pH 7.4; 2,000 RPM) and resuspended in a final volume of 50μl.
Seven day old second instar nymphs (n = 126) were fed on a suspension of freshly collected and inactivated human blood (56°C/30min) with culture epimastigotes. Aiming to prepare 5ml of inoculum at 1x107 parasites/ml of blood, we took a volume of culture that would give us 5x107 parasites total. This volume of culture was washed in PBS, centrifuged and resuspended in 50μl of PBS. This was then added to 5ml of blood. Since a second instar nymph ingests between 20–24μl of blood, we estimated that each one ingested approximately 200,000 parasites. Insects used for the control group were fed on the same inactivated blood at the same conditions, except for the parasite presence (n = 132). One day after feeding, the insects were transferred to Petri dishes whose bases were lined with filter paper discs (up to seven insects per plate; 5 plates/treatment). These were placed in temperature control chambers at 21±0.2, 24±0.2, 27±0.2 or 30±0.2°C and no further food was offered during the experiment. The times taken to reach third instar and mortality rates were recorded up to 90 days after the first moult. Insect mortalities were recorded for both treatments at 30, 60 and 90 days after ecdysis to the third instar. The entire intestinal tracts of infected insects—dead or alive at the end of the experiment—were macerated and examined to confirm parasite infection.
Culture epimastigotes were transferred at an initial concentration of 1x106/ml to cell culture flasks (25cm2) containing fresh LIT medium to a final volume of 8ml. The flasks were immediately transferred to four independent controlled temperature chambers (21±0.2, 24±0.2, 27±0.2 and 30±0.2°C) and kept there for seven days. Two replicate culture samples were simultaneously tested for each temperature. Daily, a 50μl sample was collected from each flask and stained for flow cytometry absolute counts and viability analysis.
Dual label fluorescent staining procedures were performed per sample to determine the absolute counts of live and dead parasites in each sample. For this purpose, 50μl of culture were incubated in the presence of 120μl of PBS and 25μl of fluorescein diacetate (FDA) at 7μg/ml plus 5μl of propidium iodide (PI) at 25μg/ml, both from Sigma (St Louis, MO, USA) for 10 min at room temperature. FDA (Sigma 7378) stock solution was prepared at 1mg/ml in acetone and stored at −20°C until use. PI stock solution was prepared in ddH2O at 1 mg/ml and stored at −20°C until use.
Following incubation, 20μl of fluorosphere suspension were added to each tube immediately before flow cytometric acquisition. As many flow cytometers cannot directly provide the cell concentration or absolute count of cells in a sample, the Flow-Count Fluorosphere (lot #7548025 bead counts of 986 beads/μl, Beckman Coulter, Inc., Miami Lakes, FL, USA) were used as a calibration device to directly obtain absolute counts of parasites using flow cytometry.
Quantitative flow cytometric double labeling assay, calibrated with fluorospheres, was used to simultaneously determine the number of parasites along the growth curve, as well as to calculate the mortality rate. In order to obtain the number of total epimastigotes/μL of LIT cultures, following flow cytometer acquisition of approximately 5,000 fluorospheres per sample, data analysis was carried out as follows: A bidimensional pseudocolor graph of granularity (SSC) versus non-related fluorescence 3 chart was created to exclude autofluorescent (FL3 positive) events outside the region R1. Following this, the events inside the R1 were displayed on size versus granularity plots to select and quantify the bead cluster (BEADS) and epimastigote population (EPI) as illustrated in S1 Fig. EPI gated events were then analyzed further on FL1 (FDA) versus FL2 (PI) charts to quantify the frequency of PI+FDA positive events (DEAD EPI = MORTALITY RATE) as well as FDA single positive cells (LIVE EPI) (S1 Fig). The calculation of the final concentration of TOTAL EPI and the VIABLE EPI counts were achieved with the following equations:
Totalepi=epi50×beads÷19,720
where EPI = number of epimastigote event counts for a given tube, 50 = volume of culture suspension added to each tube; BEADS = number of fluorosphere beads aspirated in a given tube and 19,720 = number of fluorosphere beads added to each tube, considering the volume of 20ml of bead suspension.
Viableepi=totalepi×liveepi100
where liveepi = percentage of FDA single positive events and 100 = the percentage conversion factor.
A dye-free sample was used as a control. A FACScan Becton Dickinson flow cytometer (La Jolla, CA, USA) was used for acquisition and the FlowJo software 9.6.3 (San Diego, CA, USA) used for data analysis using pseudocolor charts. Representative flow cytometry charts are provided in the figures.
The times that infected and uninfected insects took to die were estimated using Kaplan—Meier survival analyses. Comparisons were made with log-rank tests. Intermoult periods were compared through a nested ANOVA with Petri dish groups nested within both feeding and infection status. As no significant differences were found among the Petri dishes (F = 1.178, p = 0.286) post hoc comparisons (Tukey HSD test) were performed adding all data of the five groups of the respective temperatures.
Analyses of temperature effects on in vitro parasite growth were conducted in R version 2.13.0 [33]. The first step was to determine growth rates (i.e. regression slopes) for each replicate (bottle) for each temperature treatment. For this, live parasite population sizes were log-transformed (i.e. log10 of parasite number +1) and linear mixed effects models were used to account for the repeated measures (i.e. days 1, 2, 3 and so on). Eight growth rate values were therefore obtained (two replicates x four temperatures). These were subjected to regression analyses aimed at detecting temperature effects on growth rates, in particular, to test whether growth rates could be seen to peak at different temperatures.
The time taken to moult from second to third instar was affected by both infection (nested ANOVA, F = 67.445, p = 0.00001) and temperature (nested ANOVA, F = 68.967, p = 0.00001). The period was reduced by increasing temperatures up to one third for uninfected insects and half for infected insects (Fig. 1A-D). Meanwhile, infection with T. cruzi delayed moult by 6–11 days (Fig. 1A-D, Tukey HDS; 32.1±8.3 (control) vs. 43.5±9.2 (infected) days for 21°C (p = 0.00003), 23.2±9.0 (control) vs. 30.0±7.7 (infected) days for 24°C (p = 0.017), 17.8±8.9 (control) vs. 23.6±6.3 (infected) days for 27°C (p = 0.08), and 13.3±3.2 (control) vs. 23.3±10.4 (infected) days for 30°C (p = 0.00008)).
At the lowest temperature (21°C), mortality in uninfected control insects was more than 20% after 30 days (Fig. 1H). This initial mortality of uninfected insects was much reduced at the higher temperatures but by the end of the observations (90 days), these uninfected insects had almost all died at the higher temperatures. Against this background, infection with T. cruzi was found to accelerate mortality in the two intermediate temperatures, 24 (P = 0.02) and 27°C (P = 0.0001) (Fig. 1F & G), but not at 21 or 30°C (Fig. 1E & H).
The population growth of T. cruzi in vitro increased consistently with increasing temperature (Fig. 2; p<0.0001 for temperature effect). The best-fit regression of growth rates against temperature (Fig. 2B) was not curved, so peak growth would have occurred at or above 30°C.
Fluorescein diacetate-propidium iodide staining made it possible to distinguish live cells from those that had already started to die, these last being stained by both dyes (S1 Fig). Mortality rates of T. cruzi were below 5% at 27 and 30°C, reaching ca. 15% at 24°C and over 20% at 21°C (S1 Fig).
To the best of our knowledge, the present study is the first report of temperature-modulated mortality caused by a protozoan parasite of medical importance to its arthropod vector. In the case of malaria-mosquito systems, decreased mosquito survival during infection is only seen in unnatural combinations, although natural combinations exhibit a tendency towards such increases in mortality [34]. More recently, a natural combination of Plasmodium-mosquito (in this case an avian malaria system), showed an increase in longevity associated with a decrease in fecundity in infected mosquitoes [35]. It is now well-established that arboviruses can be virulent to their culicid vectors, depending on the taxonomic groups and the mode of virus transmission [36]. In dengue virus-Aedes associations it has been observed that the virus presence affected several mosquito fitness parameters such as survival, fecundity and oviposition success [4].
This accelerated mortality of R. prolixus infected with T. cruzi, under conditions of starvation (commonly experienced by these insects—[28]), was seen over a narrow range of temperatures (i.e., at 24 and 27°C but not at 21 or 30°C). High temperatures associated with prolonged starvation were lethal to insects, independently of parasite infection. It is well known that high temperatures promote an increase in the metabolism of insects (reviewed by [37]). Therefore, an increased mortality would be expected in starved insects submitted to higher temperatures, as already seen in previous studies [38]. According to the effect of temperature on T. cruzi growth in culture media, the mortality of infected insects would be expected to occur as a consequence of large parasite populations developed at higher temperatures. Nevertheless, there were no differences in mortality rates between infected and healthy insects kept at 30°C. This was probably a result of a lack of nutritional resources for parasite development in starved insects. In fact, it has already been demonstrated that triatomines can eliminate T. cruzi infections after long periods of starvation [39]. Curiously, R. prolixus prefers temperatures of 25.0–25.4°C when offered a choice and performs best around these temperatures [40]. Furthermore, temperatures in the sylvatic ecotopes in which this insect is to be found oscillate closely around 25°C [41]. While we might have expected the vector to be less affected by the parasite at temperatures near to its optimum (as is the case with locusts infected with the fungus Metarhizium anisopliae, [29]) we might also expect the parasite to be adapted to perform optimally at exactly these temperatures. If this is the case, then we must conclude that T. cruzi’s strategy, in its vector, results in direct physiological harm to its vector, that can be observed as vector mortality. At this range of temperatures, the parasite has a high in vitro growth rate (Fig. 2) so we hypothesize that its strategy in the vector is close to unrestrained growth, trading off an increased chance of transmission (due to a high population density in the intestine) with the cost of killing its vector and so effecting zero transmission. Given these insects are able to display temperature preferences [40, 42–45], we might expect them, when infected with T. cruzi, to alter their thermal preferences. All of these factors are liable to affect T. cruzi transmission dynamics and ultimately, epidemiology.
The number of T. cruzi parasites increased in direct relation to temperature. In fact, after seven days, parasites kept at 30°C increased their numbers close to 28 times, almost doubling their growth rate at 27°C. Previous studies have shown that both T. cruzi epimastigote and trypomastigote forms grow when exposed to 37°C [46]. The lowest temperature tested here seemed to have a harmful effect on T. cruzi, since their mortality at this temperature was close to 20%. It has been suggested that low temperatures affect the endocytic processes in T. cruzi epimastigotes [47]. Low-temperature blockage of endocytosis has also been reported in many eukaryotic cells [48–51]. Whether these effects of low temperature on parasite endocytosis are related to the poor performance of T. cruzi at 21°C deserves to be analyzed in future experiments. Temperature is important in the development and within-host dynamics of several other protozoan parasites. Leishmania species differ in their susceptibility to temperature stress, as reflected in their ability to establish infections at different sites in the mammalian body [52]. The temperature resistance of Leishmania spp. has been related with the parasite tropism, as visceral species are more temperature resistant than cutaneous species [53]. Temperature has also been shown to be important to regulate the membrane potential across the plasma membrane and the internal pH in Trypanosoma brucei [54]. In addition, the reduction in temperature from 37 to 27°C and the addition of cis-aconitate are enough to trigger the transformation of the monomorphic T. brucei from bloodstream to procyclic trypomastigotes in culture medium [55].
Beyond mortality, infection with T. cruzi considerably delayed moult in R. prolixus, across the range of temperatures tested. In contrast to results observed with other triatomine species [18–22], moulting in R. prolixus second instar nymphs was delayed by more than 10 days in a single developmental stage. In an entire life cycle the accumulation of this effect could possibly prolong by more than a month the time needed to reach the adult stage. It is highly likely that such a delay would affect insect fitness. While it will be interesting to look for physiological explanations for this (there is some evidence indicating a possible competition for lipids in this host-parasite system [56]), there may be a very good adaptive explanation, in terms of the parasite’s fitness. Trypanosoma cruzi has been reported to take approximately up to month at 28°C to colonize the intestine, reach the rectum and differentiate into infective stages [57, 58]. As triatomines will only feed again after they have moulted, it would benefit the parasite if their moult, and thus the next bloodmeal, were delayed until such a time as the parasite is in the right place and life stage to be transmitted to a vertebrate host. Such a delay would then favor parasite transmission.
Effects of T. cruzi infections on triatomine fitness have previously been described in the literature. Schaub and Lösch [59] observed that the resistance of infected T. infestans was reduced when insects were starved. However, in subsequent studies from the same group the parasite was considered subpathogenic to its invertebrate hosts since, apparently, it does not damage the vector under optimal conditions [60,61]. Meanwhile, Botto-Mahan [62] evaluated the time to moult during the ontogeny of Mepraia spinolai infected by T. cruzi (kept at 26°C) and showed that infected insects presented a delayed moult and an increased mortality when compared with control ones. However, insects from the infection treatment were always fed on infected mice, and as mentioned by the author, it is not possible to be sure that the observed effects were not a result of differences in blood quality between infected and non infected mice. Nevertheless, the deleterious effects of T. cruzi described in these studies altogether with the results presented in this study and the alteration of the reproductive fitness of R. prolixus induced by T. cruzi recently demonstrated by our group [24] represent a bulk of evidence confirming fitness costs induced by this parasite.
To conclude, we have shown that the medically-important parasite T. cruzi can exert virulence effects on the vector R. prolixus. This effect is strongest over exactly the temperature range preferred by the insect and in which it is to be found in the wild (often infected with T. cruzi). The ability of T. cruzi to develop over a broad temperature range might have contributed to its adaptation to a larger number of triatomines. It will be important to investigate virulence effects in other vector species, behavioural responses of the insects to infection (see [63] for example) and impacts on transmission dynamics.
|
10.1371/journal.pcbi.1003650 | Model Selection in Systems Biology Depends on Experimental Design | Experimental design attempts to maximise the information available for modelling tasks. An optimal experiment allows the inferred models or parameters to be chosen with the highest expected degree of confidence. If the true system is faithfully reproduced by one of the models, the merit of this approach is clear - we simply wish to identify it and the true parameters with the most certainty. However, in the more realistic situation where all models are incorrect or incomplete, the interpretation of model selection outcomes and the role of experimental design needs to be examined more carefully. Using a novel experimental design and model selection framework for stochastic state-space models, we perform high-throughput in-silico analyses on families of gene regulatory cascade models, to show that the selected model can depend on the experiment performed. We observe that experimental design thus makes confidence a criterion for model choice, but that this does not necessarily correlate with a model's predictive power or correctness. Finally, in the special case of linear ordinary differential equation (ODE) models, we explore how wrong a model has to be before it influences the conclusions of a model selection analysis.
| Different models of the same process represent distinct hypotheses about reality. These can be decided between within the framework of model selection, where the evidence for each is given by their ability to reproduce a set of experimental data. Even if one of the models is correct, the chances of identifying it can be hindered by the quality of the data, both in terms of its signal to measurement error ratio and the intrinsic discriminatory potential of the experiment undertaken. This potential can be predicted in various ways, and maximising it is one aim of experimental design. In this work we present a computationally efficient method of experimental design for model selection. We exploit the efficiency to consider the implications of the realistic case where all models are more or less incorrect, showing that experiments can be chosen that, considered individually, lead to unequivocal support for opposed hypotheses.
| Mathematical models provide a rich framework for biological investigation. Depending upon the questions posed, the relevant existing knowledge and alternative hypotheses may be combined and conveniently encoded, ready for analysis via a wealth of computational techniques. The consequences of each hypothesis can be understood through the model behaviour, and predictions made for experimental validation. Values may be inferred for unknown physical parameters and the actions of unobserved components can be predicted via model simulations. Furthermore, a well-designed modelling study allows conclusions to be probed for their sensitivity to uncertainties in any assumptions made, which themselves are necessarily made explicit.
While the added value of a working model is clear, how to create one is decidedly not. Choosing an appropriate formulation (e.g. mechanistic, phenomenological or empirical), identifying the important components to include (and those that may be safely ignored), and defining the laws of interaction between them remains highly challenging, and requires a combination of experimentation, domain knowledge and, at times, a measure of luck. Even the most sophisticated models will still be subject to an unknown level of inaccuracy – how this affects the modelling process, and in particular experimental design for Bayesian inference, will be the focus of this study.
Both the time and financial cost of generating data, and a growing understanding of the data dependancy of model and parameter identifiability [1], [2], has driven research into experimental design. In essence, experimental design seeks experiments that maximise the expected information content of the data with respect to some modelling task. Recent developments include the work of Liepe et. al [2] that builds upon existing methods [3]–[8], by utilising a sequential approximate Bayesian computation framework to choose the experiment that maximises the expected mutual information between prior and posterior parameter distributions. In so doing, they are able to optimally narrow the resulting posterior parameter or predictive distributions, incorporate preliminary experimental data and provide sensitivity and robustness analyses. In a markedly different approach, Apgar et. al [8] use control theoretic principles to distinguish between competing models; here the favoured model is that which is best able to inform a controller to drive the experimental system through a target trajectory.
In order to explore the effects of model inaccuracies we work with a computationally efficient experimental design framework. We build on the methods of Flassig and Sundmacher [9] where expected likelihoods are predicted using efficient Sigma-point approximations and leveraged for optimal experimental design, and Busetto et al. [10] where choosing the optimal measurement readouts and time points is undertaken in an iterative fashion, using Sigma-point approximations to update the posterior distributions. Here we show how mixtures distributions may be exploited to cope with non-Gaussian parameter and predictive distributions and further, derive an extension to the case of stochastic state space models. The intuition behind the approach (described fully in Materials and Methods) is shown in Figure 1, where for identical inputs, two ODE models (illustrated in blue and red respectively) are simulated for a range of parameter values, with times and representing two possible choices of times at which the true system can be measured and data gathered. Time represents an uninformative experimental choice since the behaviour of the two models is very similar, while data obtained at time is more likely to favour one model over another, since the distributions of simulated trajectories completely separate. More formally, the key steps in the method are as follow: Firstly we define the limited range of experimental options to be explored and encode them as parameterised extensions of the competing models. Secondly, the so called unscented transform (UT) [11] is used to approximate the prior predictive distribution as a mixture of Gaussians, for each model and a given experiment. Finally, optimisation is performed over the experiment parameters in order to best 'separate' the prior predictive distributions of the competing models. Parameters obtained by this optimisation represent an experiment whose generated data is predicted to maximise the differences in the subsequent marginal likelihood values of the models.
The contributions of this article are threefold; firstly, we extend a promising and computationally efficient experimental design framework for model selection to the stochastic setting, with non-Gaussian prior distributions; secondly, we utilise this efficiency to explore the robustness of model selection outcomes to experimental choices; and finally, we observe that experimental design can give rise to levels of confidence in selected models that may be misleading as a guide to their predictive power or correctness. The latter two points are undertaken via high-throughput in-silico analyses (at a scale completely beyond the Monte Carlo based approaches mentioned above) on families of gene regulatory cascade models and various existing models of the JAK STAT pathway.
We first illustrate the experimental design and model selection framework in the context of crosstalk identification. After observing how the choice of experiment can be crucial for a positive model selection outcomes, the example will be used to illustrate and explore the inconsistency of selection between misspecified models.
We consider pairs of regulatory cascades, each consisting of four transcription factors, modelled by ordinary differential equations of the form, for , where is the rate at which protein degrades, represents the maximal rate of production of , is the amount of the transcription factor, , needed for half the maximal response, and is called the Hill-coefficient, and determines the steepness of the response. A range of crosstalk models are formed (Figure 2) by inserting additional regulatory links between and with the same kinetics as above. A single model is chosen as the 'true' biological system to which we perform experiments, and six others with equal prior probabilities are proposed as models of the true system – our task will be to identify the most suitable one.
An experiment is defined by the parameter , where denotes the strength of an external stimulus to the production of , which is modelled as a term, (1)(2)(3)added to the relevant ODE equations. The time delay between the two stimulus applications is given by , and is the time at which a single measurement of the system (of species only) is taken. Prior distributions for the model parameters are set as Gaussian with means of and covariances of for both the and respectively, with the Hill coefficient fixed at .
The results of this round of experimental design are shown in the top left of Figure 3, where a good choice of is found to be , with a corresponding score of . From the figure, it can be seen that this experiment is predicted to distinguish some pairs of models better than others. In particular, the distribution of scores suggests that while the marginal likelihoods of most pairs of models are separated as desired, there is no power to discriminate between models and , or models and . Indeed, data obtained by performing the experiment upon our 'true' system, leads to posterior probabilities for each model with the same pattern.
As a sanity check, we first choose the true model from amongst the set of competing models (), and as expected find that it is recovered by model selection with probability 1. However if the true model is not represented by (a far more realistic case) but instead the crosstalk model with a single connection from to , then models and are found to have similar posterior probabilities of approximately . Likewise, and share a posterior probability of , while a clear difference exists between any other pair of models. To distinguish further between the pair of highest scoring models, a further round of experimental design was performed, with the resulting experiment and data providing strong evidence in favour of model .
In an attempt to evaluate the added value of choosing rationally for this example, we calculate scores for a uniform sample of values of from the same range as explored above. The resulting score distribution shown in Figure 4a, peaks in the interval which corresponds to an average Hellinger distance of between the maximally separated marginal likelihoods of each pair of models. This is in contrast to the experiment found by our approach which lives in the tail of the distribution, with an average Hellinger distance of , and highlights how unlikely it is to find suitable experiments by chance alone. Experiments with even higher information content are found, which suggests that more care could be taken with the optimisation of , by for example, increasing the population size, or number of generations of the genetic algorithm used.
Perhaps unnervingly, the evidence in the first experiment is found to contradict (though not significantly in this case) the decision in favour of model over , which is based on additional data from the second experiment. This suggests the possibility that the choice of experiment influences not only the amount of information available to select a particular model, but also the outcome of the model selection itself. Indeed the distribution of independently selected models from data generated by random experiments is surprisingly flat (Figure 4b). Even at very low levels of assumed noise, the most frequently selected model is chosen for less than half the experiments undertaken. This has been, to our knowledge, completely overlooked by the experimental design literature, but has important implications that we will explore further below.
To examine this last observation in more detail, we work with three of the crosstalk models described above, with connections between, , and respectively. The last of these is designated as the true model, and the others are considered as competing hypotheses about the location of the crosstalk connection. We perform 36100 experiments to collect data sets of size 1, 2, 4 and 8 equally spaced time points, each consisting of simulating the true model with different values of that correspond to changes in the delay between stimulus applications, and variation of the time at which the state of is first measured. An independent round of model selection is performed for each data set, and the posterior probabilities for each model are calculated.
The results for data sets of size 1 and 8 are illustrated in Figure 4c and 4d as heatmaps of posterior probabilities of the first model, and show that the vast majority of the space of experiments is split into distinct regions of high, low and equal probability for each model. In the case of a single time point, most of the explored experiment subspace is found to be uninformative, with the data providing equal support for each model. Three other distinct regions are identified, of which two show decisive support (on the Jeffreys scale) for the first model, and one for which the second model is chosen decisively. In other words, by varying the experimental conditions an unequivocal choice (in isolation) for either model can be obtained. As more data points are considered, the uninformative region grows smaller, but regions of decisive support for each model remain. Interestingly, these regions are located in distinctly different places for single or multiple time points, although they remain similar for 2 or more time points. This reflects the added value of time series experiments – the marginal likelihoods now balance the ability of the models to reproduce each time point, with their ability to capture the autocorrelation of the time series.
In order to establish whether the observed inconsistencies are an artefact of the UT approximations, we perform a similar but necessarily course grained study using MultiNest [12], [13], an implementation of nested sampling (a Monte Carlo based technique with convergence rate [14]). Results obtained using MultiNest (shown in the upper right of figure 5) are almost identical to those of figure 4c, displaying the same regions of decisive support for each model. Given how difficult it is to estimate marginal likelihoods in general, the excellent performance of the UT (with only one Gaussian component) may seem rather surprising, until one notes that for the models and experiments considered, the prior predictive distributions are approximately Gaussian themselves (Figure 5). We discuss how the framework can deal with non-Gaussian effects, such as those found in the next examples, in the appendix.
In this section we undertake an analysis of three mass action models of varying degrees of resolution of the JAK-STAT signalling pathway [15]. Each model describes the initial pathway activity after receptor activation (Figure 6), but before any feedback occurs. In brief, the signalling process consists of a receptor binding to JAK to form a complex that can dimerise in the presence of interferon- (IFN). This dimer is activated by phosphorylation by JAK, and in turn deactivated after being bound by tyrosine phosphatase (SHP_2). In its active state, the receptor complex phosphorylates cytoplasmic STAT1, which is then able to dimerise and act as a transcription factor [16].
We take the most detailed model, , with 17 state variables and 25 parameters (published by Yamada et al. [16]), as our true system to which in-silico experiments can be performed, and select between two of the other models proposed by Quaiser et al. The first of these competing models, , simplifies the true system, by neglecting a reaction – the re-association of phosphorylated STAT1 to the activated receptor – and thereby reducing the system to 16 states and 23 parameters. A series of five other 'biologically inspired' simplifications leads to our second model, , which has 9 states and 10 parameters (these steps are summarised in Figure 6).
We set the parameter priors as a component mixture of Gaussians fit to a uniform sample from the hypercube , where is the parameter dimension, such that all the parameter values inferred for each model by Quaiser et al. are supported. We define and undertake two classes of experiment upon the true model (with parameters fixed to the published values); in the first, the IFN stimulus strength and the initial time point of a time series of 8 equally spaced measurements of the amount of JAK bound to the receptor are varied, and in the second, the species to be measured and the time at which this first measurement takes place are adjusted.
Model selection outcomes for each experiment (shown in Figure 7) show similar features to those for the crosstalk models, with distinct region of high posterior probability for each model. For the first class of experiments, selection between models and reveals strong support for the simpler model when data is gathered at earlier time points. The more complex model, , is generally favoured for later time series, and also for a very limited range of IFN stimuli strengths at early time series. For the second class of experiments, the model selection outcome is found to depend strongly upon which species is measured. The simpler model is chosen decisively and almost independently of the measurement times considered when cytoplasmic phosphorylated STAT1, in monomeric or dimeric form, or two forms of the receptor complex (IFN_R_JAKPhos_2 and IFN_R_JAK) are measured. The same is true of the complex model for measurements of two other forms of the receptor complex (IFN_R_JAK2 and IFN_R_JAKPhos_2_SHP_2). Otherwise the model selection outcome is time dependant or the choice of species is found to be uninformative.
Both these case studies make it clear that under the realistic assumption that all models are more or less incorrect, model selection outcomes can be sensitive to the choice of experiment. This observation has particular importance for studies that treat models as competing hypotheses that are decided between using experimental data; it is quite possible that if different experiments are undertaken, the conclusions drawn will also be different. In particular, the confidence calculated for such a conclusion (using the Jeffreys scale or another measure) can be misleading as a guide to how correct or predictive a model is (Figure 8a); in both the examples studied here, conditions exist such that any of the competing models can score a 'decisive' selection. The model selection outcome and associated confidence must therefore be strictly interpreted, as only increasing the odds of one model (with respect to others) for the data gathered under the specific experimental conditions.
In light of this observation, the role of experimental design may need to be examined further. Since different models can be selected depending on the experiment undertaken, the use of experimental design will necessarily lead to choosing the model which, for some 'optimal' experiment, has the highest possible predicted level of confidence i.e. experimental design implicitly makes confidence a selection criterion. Is it misleading to claim high confidence in a model selection result when the models have been set up (by extensions to mimic the optimal experiment) for this purpose? Is a bias introduced into the inference via experiment design? In the context of experiment design for parameter estimation, MacKay suggests this is not a problem [17], stating that Bayesian inference depends only on the data collected, and not on other data that could have been gathered but was not. Our situation here is different since we consider changes not only to the data collection procedure, but also the data generation process and in turn the competing models themselves. It seems plausible that some models will gain or lose more flexibility than others with regards to fitting data for a particular choice of experiment. Even if the actual model selection is not biased, the confidence we associate with it will scale with the optimality of the experiment. After performing the optimal experiment, should there be any surprise that the selected model seems to have high support from the data? We feel these questions need further investigation.
In practical terms, the important question seems to be: how wrong does the model structure (or parameter values) have to be before the less predictive model (or that which captures less about the true system) is chosen? Clearly the answer is sensitive to the system and models under study, and moreover, the issue of how to compare the size of different structural inaccuracies is non trivial. Here, as a first attempt, we limit ourselves to considering the simple case of parameter inaccuracies in linear ODE models.
We define a 'base' model as the linear ode system defined by its Jacobian matrix with entries, and 'extensions' to this model as an extra row and column,
Biologically such an extension may represent the inclusion of an extra molecular species into the model, along with rules for how it interacts with components of the original system. Defining true base and extension models by and , we consider two models, and where and , are competing (true and false) hypotheses about the structure of the model extension, with a zero or indicating a belief that species does not directly affect the rate of increase of species . Parameters , are the unknown strengths of these interactions, over which we place a component mixture of Gaussians prior, fit to a uniform distribution over the interval for each parameter. We represent inaccuracies in modelling the base as additive perturbations and . Data was generated by simulating the state of the first variable of the true model at times , for initial condition .
Model selection outcomes for different pairs of values for the perturbations , are shown in Figure 9. Distinct regions for each possible outcome are found and colour coded in the figure, with red indicating that the true extension has been identified successfully, yellow representing a decision in favour of the false extension, orange that evidence for either model is not substantial on the Jeffreys scale, and finally blue indicating that the marginal likelihood for both models is found to be less than , for which any conclusion would be subject to numerical error. Increasing this threshold has the effect of replacing red areas with blue.
In the majority of cases tested, the true extension is correctly identified despite inaccuracies in the base model. However, a set of perturbations are seen to confound the selection, and allow the false extension to obtain substantial support. Furthermore, the selection outcome is found to be more sensitive in some directions than others, with relatively small perturbations to base model entry causing a change in outcome and creating decision boundaries near the lines and . Prior to our analysis, it would be hard to predict these observations even when the true model is known and as simple as that explored here.
In real applications, where the true model is unknown and more complex, it may not be possible to tell whether a conclusion is an artefact of model inaccuracies, even when the truth of the conclusion itself can be tested by direct experimental measurement. However, the type of analysis undertaken here at least gives a measure of robustness for the conclusion to a range of model inaccuracies. Unfortunately, this remains difficult to implement in a more general setting – for example, in climatology, where the accepted method of coping with structural uncertainty is through the use of large ensembles of similar models produced by various research groups [18], a luxury that cannot be afforded on the scale of the most ambitious systems biology projects. While the practical challenges of dealing with large numbers of models is somewhat overcome by the model selection algorithm described above, a harder conceptual problem exists of how to define perturbations to more complicated classes of model, and to compare their strengths.
Finally, the example also highlights the difficulty of testing a hypothesis that represents only part of a model. The study shows that the implicit assumption that the base model is accurate, is not necessarily benign, and can affect any conclusions drawn – a result that is borne out by the logical principle that from a false statement, anything is provable.
The scale of the analyses detailed above, comprising thousands of marginal likelihood computations, requires extreme computational efficiency. Indeed it is completely beyond Monte Carlo based methods such as that recently developed by Liepe et al. [2], which are limited to exploring small sets of models and experiments. Here, the efficiency was obtained by using the unscented transform for propagating Gaussian mixture distributions through non-linear functions. Further computational savings can be made by exploiting the highly parallelizable nature of Flassig and Sundmacher's method [9], which we have extended for use with mixture distributed priors and stochastic state space models.
This efficiency has allowed us to explore model selection problems involving relatively large numbers of models and experiments, and investigate the robustness of model selection results to both changes in experimental conditions and inaccuracies in the models. Results from the latter two studies illustrate some common, but often ignored, pitfalls associated with modelling and inference. Firstly, we show that the conclusions of a model selection analysis can change depending on the experiment undertaken. Related to this, we observe that confidence in such a conclusion is not a good estimator of the predictive power of a model, or the correctness of the model structure. Further we note that the use of experimental design in this context maximises the expected discriminatory information available, and implicitly makes confidence in the outcome a criterion for model selection. In the future we intend to investigate the desirability of this property and how it affects the interpretation of the confidence associated with model selection outcomes.
At the heart of these issues is a lack of understanding of the implications of model (or parameter) inaccuracies. Often improved fits to data or better model predictions are interpreted as evidence that more about the true system is being captured. This assumption underlines a guiding paradigm of systems biology [19], where a modelling project is ideally meant to be a cycle of model prediction, experimental testing and subsequent data inspired model/parameter improvement. However, it is possible that improved data fitting and predictive power (although desirable in their own right) can be achieved by including more inaccuracies in the model. In the context of parameter estimation, this concept of local optima is widely known, and their avoidance is a challenge when performing any non-trivial inference. One simple method to do so is to include random perturbations in the inference, in order to 'kick' the search out of a local optimum. Perhaps a similar strategy might be included in the modelling paradigm; by performing random experiments, or adding or removing interactions in a model structure, data might be gathered or hypotheses generated that allows a leap to be made to a more optimal solution.
While we have been concerned solely with the statistical setting, it is reasonable to expect similar results can be found for alternative model discrimination approaches e.g the use of Semidefinite programming to establish lower bounds on the discrepancy between candidate models and data [20]. Here the particular subset of models that are invalidated will be dependent upon the experiment undertaken. However, emphasis on invalidating wrong models instead of evaluating the relative support for each at least reduces the temptation for extrapolated and, perhaps, false conclusions.
George E. P. Box famously stated that 'Essentially, all models are wrong, but some are useful'. Here we would add that if nothing else, models provide a natural setting for mathematicians, engineers and physicists to explore biological problems, exercise their own intuitions, apply theoretical techniques, and ultimately generate novel hypotheses. Whether the hypotheses are correct or not, the necessary experimental checking will reveal more about the biology.
The UT is a method that describes how the moments of a random variable, , are transformed by a non-linear function, . The algorithm begins by calculating a set of weighted particles (called sigma-points) with the same sample moments up to a desired order as the distribution . For the results shown here, we use a scaled sigma-point set that captures both means and covariances [21], where is the dimension of , and are the mean and covariance of , represents the th column of a matrix , and
The sigma-point weights are given by, and finally, the parameters , and may be chosen to control the positive definiteness of covariance matrices, spread of the sigma-points, and error in the kurtosis respectively. For the results in this article we take as is standard in the literature [22], and which is optimal for Gaussian input distributions, while , controlling the spread of sigma-points is taken small as to avoid straddling non-local non-linear effects with a single Gaussian component [21].
The mean and covariance of the variable , can be estimated as the weighted mean and covariance of the propagated sigma-points, (4)(5)
We denote the resulting approximate probability density function for , by .
By matching terms in the Taylor expansions of the estimated and true values of these moments, it can be shown that the UT is accurate to second order in the expansion. More generally, if the sigma-point set approximates the moments of up to the order then the estimates of the mean and covariance of will be accurate up to the term [11]. Crucially, the number of points required ( for this scheme) is much smaller than the number required to reach convergence with Monte-Carlo methods.
We will consider discrete time state space models, , with state–transition () and observation () functions both parametrized by ,(6)(7)where , is the time series of dimensional measurements that we are trying to model, is the dimensional true state of the system at time , and , and are independent, but not necessarily additive, Gaussian white-noise process and measurement terms. Bayesian model selection compares competing models, , by combining the a priori belief in each model, encoded by the model prior distribution , with the evidence for each model in the data , as quantified by the marginal likelihood,where is the parameter prior for model . In the Bayesian setting, the relative suitabilities of a pair of models are often compared using the ratio of posterior probabilities, known as the Bayes factor,with a Bayes factor of seen as substantial [23]. However, for complex or stochastic models, the marginal likelihood can be intractable, and so approximate likelihood free methods, such as Approximate Bayesian Computation are becoming increasingly important and popular within the biosciences [24]. A big drawback of such Monte-Carlo based algorithms is the large number of simulations – and associated computational cost – required to estimate the posterior distributions or Bayes factors. Even with GPU implementation [25], applications are currently still limited to comparing pairs or handfuls of models.
In order to address the issues raised above, a higher-throughput model selection algorithm is needed. Our approach will be to fit mixture of Gaussian models to the prior parameter distribution for each model, so that we can exploit the UT within the state-space framework to drastically reduce the number of simulations necessary to estimate the distribution of the output of the model. Gaussian mixture measurement and process noise can also be considered, as in the work on Gaussian sum filters [26], [27], although the number of mixture components required to model the output at each time point then increases exponentially, and in the case of long time series, component reduction schemes need to be implemented.
With this approximation, the marginal likelihood may be expressed as the sum, (8)(9)(10)where the components, , can be determined using the UT as described below. Note that the accuracy of the approximation can be controlled by the number of components used. However, in the presence of nonlinearities, choosing the number and position of components solely to fit the prior distribution may not be adequate. This is because we need to have enough flexibility to also fit a complex and possibly multi-modal output. Indeed, except at the asymptotic limit of dense coverage by the mixture components, it is possible to construct badly behaved mappings that will lead to loss of performance. For the applications visited in this article, the models proved well behaved enough such that a single component and 10 components respectively for the crosstalk and JAK-STAT systems sufficed for sufficient agreement with the nested sampling and Monte Carlo results. An improvement to the method described here would be to update the number of components automatically with respect to the model behaviour in a manner similar to how Gaussian mixtures can be adaptively chosen in particle based simulation of Liouville-type equations [28], [29].
For the deterministic case including the examples considered in this article, we have , and the state–space model simplifies to, where might represent the simulation of certain variables of a system of ODEs, parameterised by , with additive measurement error . In this case the marginal likelihood can then be expressed as, where each component is obtained simply through application of the UT with input distribution , and liklihood that is Gaussian with mean, , and variance, .
To estimate the marginal likelihood in the stochastic case (), we assume the observation function takes the form of a linear transformation of the true state and measurement noise at time with additive noise, (11)where is an matrix. In practice this might correspond to the common situation where observations are scaled measurements of the abundance of various homo- or heterogeneous groups of molecules.
We may then write the mean of the observation, , in terms of the statistics of , (12)for any , and from the bilinearity of the covariance function, the covariance between any pair of observations, , as, (13)
(14)since is independent of for all and . We now need to find expressions for the process state covariance terms in equation 14. To do so we apply the UT iteratively for to transform the state-variable, through the state-transition function , with input distribution given by,
The result is a Gaussian approximation to the joint distribution for each , and hence also to the conditional distributions . Given that is a Markov process and that the product of Gaussian functions is Gaussian, we also have a Gaussian expression for the joint distribution, ,
The covariance between any pair of observations and , may then be found by substituting relevant entries from the covariance matrix of the density of Equation into Equation 14. The subsequent Gaussian approximation to the joint distribution of , given , constitutes one component in the mixture approximation of the marginal likelihood given in Equation 10.
We first introduce a vector of experiment parameters, , that describes how the dataset is created, specifying, for example, the times at which the system is stimulated, the strengths and targets of the stimuli, knockouts or knockdowns, along with the choice of observable to be measured at each time point. We can then model the system and experiments jointly, extending the to include terms describing the possible experimental perturbations, and the to capture the measurement options, (15)(16)
We assume that there is overlap between the system observables appearing in each model so that experiments that allow model comparison can be designed.
To illustrate how this might be done in practice, we consider a typical set of ordinary differential equations used to describe a gene regulatory mechanism, (17)(18)where are the parameters controlling the rates of production and degradation of an mRNA, , and a protein, , subject to the concentration of a repressor protein, . We define the state transition function as their solution evaluated at the next measurement time-point which is now dependant on the choice of , given the state at time , and subject to some additive noise . These equations have be extended as, (19)(20)to model a range of possible experimental perturbations, e.g. setting mimics a knockout of the gene producing mRNA , and an input stimulus to species . The observation function , as before can be some linear function of the states, however, the selection of variables and coefficients is now an experimental choice specified by ,
Given a particular set of experimental options, , the marginal likelihood of model for any possible data set (the prior predictive distribution) can be estimated efficiently from equation 10, with the components calculated with respect to the extended system and experiment model. Comparisons between such prior predictive distributions for competing models provides a means to predict the discriminatory value of a proposed experiment. Intuitively, values of , for which the prior predictive distributions of two models are separated, correspond to experimental conditions under which the models make distinct predictions of the system behaviour. Data gathered under these conditions are thus more likely to yield a significant model selection outcome. More formally, we can quantify the value of an experiment , using the Hellinger distance between the prior predictive distributions, which takes the following closed form for multivariate Gaussian distributions, and , where, or for Gaussian mixtures, it can be evaluated using the method suggested in [30].
The experimental design problem may then be posed as an optimisation problem (the results in this article used a genetic algorithm [31] of population size and generations) over - we search for the set of experimental parameters, , for which the Hellinger distance between the competing models , , is maximal. will then specify the experiment that gives the greatest chance of distinguishing between and . In the case where more than two models are considered, the cost function is taken as where the sum of exponentials is introduced to encourage selection of experiments with a high chance of distinguishing between a subset of the model pairs, over experiments with less decisive information for any pair of models, but perhaps a larger average Hellinger distance over all model pairs.
|
10.1371/journal.pcbi.1002906 | Pairwise Analysis Can Account for Network Structures Arising from Spike-Timing Dependent Plasticity | Spike timing-dependent plasticity (STDP) modifies synaptic strengths based on timing information available locally at each synapse. Despite this, it induces global structures within a recurrently connected network. We study such structures both through simulations and by analyzing the effects of STDP on pair-wise interactions of neurons. We show how conventional STDP acts as a loop-eliminating mechanism and organizes neurons into in- and out-hubs. Loop-elimination increases when depression dominates and turns into loop-generation when potentiation dominates. STDP with a shifted temporal window such that coincident spikes cause depression enhances recurrent connections and functions as a strict buffering mechanism that maintains a roughly constant average firing rate. STDP with the opposite temporal shift functions as a loop eliminator at low rates and as a potent loop generator at higher rates. In general, studying pairwise interactions of neurons provides important insights about the structures that STDP can produce in large networks.
| The connectivity structure in neural networks reflects, at least in part, the long-term effects of synaptic plasticity mechanisms that underlie learning and memory. In one of the most widespread such mechanisms, spike-timing dependent plasticity (STDP), the temporal order of pre- and postsynaptic spiking across a synapse determines whether it is strengthened or weakened. Therefore, the synapses are modified solely based on local information through STDP. However, STDP can give rise to a variety of global connectivity structures in an interconnected neural network. Here, we provide an analytical framework that can predict the global structures that arise from STDP in such a network. The analytical technique we develop is actually quite simple, and involves the study of two interconnected neurons receiving inputs from their surrounding network. Following analytical calculations for a variety of different STDP models, we test and verify all our predictions through full network simulations. More importantly, the developed analytical tool will allow other researchers to figure out what arises from any other type of STDP in a network.
| Spike timing-dependent plasticity (STDP) is a widespread mechanism that modifies synapses on the basis of the intervals between ensembles of pre- and postsynaptic spikes [1], [2]. The most prevalent form of STDP involves potentiation of the synapse when presynaptic spikes precede postsynaptic spikes, and depression for the reverse ordering [3]. STDP is inherently a local synaptic modification rule because the determinant of synaptic modification is the timing of pre- and postsynaptic spikes. Neurons, on the other hand, are typically embedded in interconnected networks in which each neuron receives thousands of synapses from other neurons [4], [5]. A number of studies have explored how STDP shapes the distribution of synaptic weights for a population of synapses converging onto a single neuron [6]–[10]. Here, we consider the more difficult problem of bridging the gap between the locality of STDP and the global structures that it generates in a recurrent network of spiking neurons.
The problem of STDP in a recurrent network has been addressed before in a number of studies [11]–[22]. The generally antisymmetric shape of the STDP window, in which reversing the ordering of pre- and postsynaptic spikes reverses the direction of synaptic change, led to the proposal that this synaptic modification rule should eliminate strong recurrent connections between neurons [11], [23]. This idea has recently been expanded by Kozloski and Cecchi [21] to larger polysynaptic loops in the case of “balanced” STDP in which the magnitudes of potentiation and depression are equal. These authors also showed that balanced STDP organizes network neurons into in- and out-hubs. The possibility of enhancing recurrent connections through pair-based STDP was also proposed by Song and Abbott [11] and is further explored by Clopath and colleagues [22] in a more complex model. An excessively active group of neurons has been shown to decouple from the rest of the network through STDP [13], and in presence of axonal delays, STDP enhances recurrent connections when the neurons fire in a tonic irregular mode [14]. Here, we show that, surprisingly, all of these network properties can be explained through an understanding of the effect of STDP on pairwise interactions of neurons. This provides an analytically tractable way of relating the structures arising in a network to properties of the STDP model being used to modify synapses.
STDP is characterized by a change of synaptic strength, , induced by a pair of pre- and postsynaptic action potentials with time difference (pairing interval) . The functional relation between the synaptic modification and the pairing interval is given by(1)
The positive parameters and specify the maximum potentiation and depression, respectively. We express the synaptic strengths in units of the membrane potential (mV), so and have mV units. The time constants and determine the temporal spread of the STDP window for potentiation and depression. The parameter , when it is nonzero, introduces a shift in the STDP window such that for positive values of even in cases where a presynaptic action potential precedes a postsynaptic spike by a short interval (), the corresponding synapse gets depressed [10]. Conversely, for negative values of the synapse gets potentiated even when a postsynaptic action potential precedes a presynaptic spike by a short interval () [14]. We recover conventional STDP by setting . For conventional STDP, spike interactions are all-to-all, meaning that all possible pre-post pairs contribute to plasticity. However, using an STDP model with only nearest-neighbor interactions does not qualitatively alter the obtained results (see Text S1). The main motivation for using all-to-all interactions in the first cases analyzed below is the clarity of explaining the resulting synaptic dynamics with these interactions. In the case of shifted STDP (), we use only nearest-neighbor spike pairs to drive plasticity, for reasons of stability explained in [10]. Moreover, as the results show, the background firing rates of neurons play an important role in synaptic dynamics with nearest-neighbor interactions if the STDP window is shifted.
The effect of STDP on synaptic weights depends not only on properties of the STDP window function of equation 1, but also on how boundaries are imposed on the range of allowed synaptic weights [7], [9], [24]. In the case of “soft” boundaries, the synaptic dynamics will be confined only to a narrow region in the middle of the range of synaptic strengths [24] (see Text S1, figure S2), whereas in the case of “hard boundaries”, the synapses can explore the whole range of their allowed strengths. Therefore, we restrict our analysis to the case of hard boundaries as it results in more interesting network structures.
To gain analytical insights into the structures that arise from STDP in a network and then to verify those insights, we take a dual approach for each form of STDP we study. First, we consider a pair of connected neurons that we imagine to have been extracted from a full network, representing the remaining neurons of the network as independent Poisson input to each of these neurons (figure 1, middle, with network inputs shown as gray connections). This simplified model allows us to perform a detailed analytic study of how STDP affects the synapses between the two explicitly modeled neurons (drawn in yellow and green in figure 1, middle). Modifications of synapses from the rest of the “network” onto these two neuron is not modeled directly, but is duplicated by changing the mean effective network input into each neuron, which alters its baseline firing rate. Despite these simplifications, many of the structures induced by STDP in a large network can be explained by analogy with properties of this two-neuron system. As the second component of our approach, we verify that analytic predictions extracted from the simplified model apply to full networks with STDP acting at all synapses by simulating these full networks.
The two representative excitatory neurons drawn from the network are labeled neuron 1 and neuron 2, and are reciprocally connected (figure 1, middle). We denote the strength of the synapse from neuron 1 to neuron 2 as and the strength of the synapse from 2 to 1 as . Due to the additional “network” inputs, each neuron fires at a baseline rate, given by and , respectively. We assume that there are no significant correlations between the baseline spike trains of neurons 1 and 2, because the recurrent inhibition is strong. In the presence of a strong recurrent inhibition, spontaneous fluctuations in the activity of excitatory and inhibitory populations accurately track each other and cancel the effect of shared input [25], and this is true for the networks we study except where otherwise indicated (figure S3). It is worth noting that we do not assume that the baseline firing rates are constant; they are functions of the synaptic strengths in the network and therefore change with STDP. We are interested in the local dynamics of pairs of reciprocal synapses given the current values of baseline firing rates. As we will see below, this pairwise analysis can predict how the baseline firing rates will change over time as STDP affects the network. Under the conditions we assume, synapses are modified primarily by random pre-post pairings of their baseline spike trains. The average amount of modification due to this baseline activity is the same for both synapses (i.e. the same for both and ). On top of the baseline firing, the reciprocal synaptic connections induce correlations between the spike trains of the two neurons. Each spike arriving from neuron 1 to neuron 2 transiently increases the firing rate of neuron 2 proportional to (figure 1, yellow areas). This transient increase (or causal bump) potentiates (figure 1, top) and depresses (figure 1, bottom). Likewise, the causal bump in the cross-correlation induced by neuron 2 into neuron 1 (figure 1, green areas) potentiates and depresses . Taken together, the average drift of the synaptic pair can be expressed as(2)where the coefficients , and can be calculated from the parameters of the neuronal and plasticity models (see Text S1). In each equation, the coefficient represents the potentiation induced on a synaptic weight by the causal effect it has on the firing of its postsynaptic neuron, represents the depression induced in the same synapse by the causal effect of its reciprocal synapse on presynaptic firing, and characterizes the synaptic modification due to random pairings of the baseline spike trains of the two neurons. Because the synaptic strengths are bounded between 0 and , the drift of the synaptic pair, as described by these equations, is restricted to a limited region in the state space . As we will see in the following sections, this restriction results in a number of interesting effects that would not arise in a strictly linear system.
In what follows, we first examine the effect of different parameterizations of the STDP window on the synaptic pair. This leads to a number of predictions about the structures that arise from STDP in networks. We then test each prediction through numerical simulations of a large network. In the case in which the STDP window is not shifted (), the time constants of the window are assumed to be equal (), but we vary the balance between potentiation and depression by changing the maximum values and . The same qualitative results hold when the maximum values are set equal and the potentiation/depression balance is modified by changing the time constants (see Text S1).
The simplest form of STDP that we consider is balanced with equal potentiation and depression domains, i.e. with . In this case, the coefficient vanishes because the baseline potentiation and depression cancel each other. In addition, the coefficients and are equal. These conditions greatly simplify the system of equations (2). We visualize the dynamics of the synaptic pair by phase planes (figure 2). Note that all phase planes throughout this paper are snapshots of the dynamics for a given network firing rate, and as we will see below, they should be recomputed if the baseline firing rates change appreciably. When the baseline firing rates are equal (), the values of the synaptic weights do not change when and are equal, i.e. the synaptic drift is zero on the line (figure 2A, solid line). However, this equilibrium is unstable. If the two synapses have unequal strengths, the stronger synapse grows even stronger and weakens the other synapse until they reach the boundary of their allowed range (figure 2A, arrows). Then, the synapses continue their dynamics along the boundary edge until they reach the upper-left () or lower-right () corner of the state space (figure 2A, filled circles), depending on which synapse was stronger to begin with. These “attractors” of the synaptic dynamics indicate that, at steady-state, loops between pairs of neurons are eliminated by this form of STDP. A linear system of differential equations cannot have more than one attractor. The existence of two attractors here is a consequence of restricting the dynamics to a limited range, and it suggests that STDP favors unidirectional connections and eliminates loops in a network (figure 3A), in agreement with the results of [21]. The attractors also imply that, for each neuron, the strengthening of an incoming synapse is accompanied by the weakening of an outgoing synapse and vice versa. If we consider the effect of this interplay at the level of a network, it is expected that for each neuron the number of above-threshold incoming and outgoing synapses will be linearly related with a coefficient of -1, such that their sum remains constant. Our simulation results and previous work [21] also confirm this prediction (figure 3B).
When the baseline rates of the two neurons are not equal, the line of equilibrium is tilted to (figure 2B, see Text S1). As a result, the size of the basins of the two attractors differ, and the outgoing synapses of the neuron with the higher firing rate are more likely to strengthen, while its incoming synapse are likely to weaken (figure 2B, top-left corner). Conversely, outgoing synapses from the neuron with the lower firing rate are more likely to weaken and its incoming synapses strengthen. If we generalize this behavior to the context of a network, an important prediction can be made: neurons with low initial firing rates should attract strong excitatory synapses onto themselves but project weaker synapse to other neurons. Neurons with high firing rates should experience the opposite trend; they lose incoming synaptic input through synaptic weakening, while their outgoing synapses strengthen. Therefore, if the external input is biased to give a sub-population of excitatory neurons an initially higher (lower) firing rate than the rest of the network, these neurons will become out-hubs (in-hubs) through STDP. We tested this by setting the mean of the external input to the neurons such that the initial firing rate of the first hundred excitatory neurons (1–100) was , the initial firing rate of the next hundred excitatory neurons (101–200) was , and the initial firing rate for the rest of the excitatory neurons (201–500) together with that of the inhibitory neurons was . The results show that the sub-population with high initial rate indeed turns into out-hubs, while the sub-population with low initial rate turns into in-hubs once the synaptic weights reach steady state (figure 3C).
Another related prediction is that the firing rates of neurons in the network tend to equalize through STDP. This is because neurons with high initial firing rate become out-hubs, thereby receiving less input from the rest of the network, which lowers their final, equilibrium firing rates. At the same time, they share their initial high firing rate with the other neurons of the network through the strengthening of their outgoing synapses. The opposite happens to neurons with low initial firing rates; they turn into in-hubs. As a result the final firing rates of the neurons become homogenized across the network at steady-state. To test this prediction, we tracked the evolution of the average firing rates of the above three sub-populations throughout the simulation. The results confirm that the final firing rates of all three sub-populations equalize once the synaptic weights reach steady-state (figure 3D).
As another prediction, the steady-state mean of the synaptic weights is expected to converge to the midpoint of its allowed range, regardless of the initial distribution of weights and the initial firing rate of the network. This is due to the precise balance between the potentiation and depression domains of STDP in this case. If the initial mean is already at the midpoint of the allowed range, the potentiation and depression events have equal probabilities across the network due to this balance. If the initial mean is smaller than the midpoint, a number of depression events would not be fully realized because they are likely to push the synaptic weights to , so equality of potentiation and depression is disrupted in favor the former, and the mean tends to increases. Similarly, if the initial mean is larger than the midpoint, some of the potentiation events push the corresponding weight above the maximum value and are truncated. This decreases the mean. Therefore, the mean tends to the midpoint in both cases, and the baseline firing rate (which already has the tendency to equalize) does not change this scenario. The simulation results confirm this prediction (figure 3F). If the initial value of the weights are drawn from a uniform distribution, the mean will be at the midpoint from the very beginning, and it remains there throughout the simulation, making the final network firing rate equal to the initial rate (figure 3E). Note that in figure 3E, and in similar figures to follow, we plot quantities as a function of the initial firing rate of the network, as opposed, for example, the external input used to modify this rate. We do this to make apparent changes in the network firing rate caused by STDP.
In studies of a single neuron receiving Poisson input through synapses that are modified by STDP, it has been shown that stability requires depression to dominate over potentiation [6]–[9]. If potentiation dominates in this case, all the synaptic weights get potentiated to the maximum allowed value. Interestingly, for STDP within a network of neurons, potentiation dominated STDP can be stable because this instability is counteracted by the depression induced on reciprocal pairs of synapses. In other words, if one synapse between a reciprocally connected pair of neurons grows, the other synapse is likely to be weakened, preventing the outcome in which all the synapses are maximally potentiated.
When the potentiation/depression balance is tipped in favor of potentiation ( and in our examples), the coefficient in equation (2) becomes larger than (see Text S1). In addition, the baseline parameter is positive. By setting the right-hand-sides of equations (2) to zero, the fixed point values of the two synaptic weights are found to be and . Both of these values are negative, so the fixed point lies outside the allowed range of synaptic strengths. Furthermore, this fixed point is unstable (in both directions), which means that the weights tend to drift away from it (figure 4; see Text S1).
We now examine the influence of the outlying, unstable fixed point on the dynamics within the allowed region of synaptic values when the baseline firing rates of the two neurons are equal. If the initial weights are fairly close to each other (figure 4A, red area), they eventually end up at the attractor in the upper-right corner of the phase space due to repulsion from the outlying, unstable fixed point. This attractor corresponds to strong recurrent connections. Trajectories of weights that hit the upper boundary () perpendicularly, form another fixed point that is unstable (figure 4, open circle on top). Trajectories to the left of this critical line are eventually absorbed by the attractor at the top-left corner (corresponding to a unidirectional connection), while trajectories to its right are absorbed by the top-right attractor (corresponding to recurrent connections). A similar unstable fixed point exists on the rightmost boundary (; figure 4, open circle on the right). As a result, the state-space of the weights is partitioned into three basins of attraction: one leading to the attractor corresponding to recurrent connections (figure 4, red shading) and the others to attractors that produce unidirectional connections (yellow and green shadings).
The appearance of the attractor corresponding to recurrent connections leads to a prediction about networks: STDP with dominant potentiation can generate loops in a network in contrast to balanced STDP. This prediction is confirmed by our numerical simulations showing that there are more loops induced in the steady-state weight matrix of a network in this case (figures 5A).
As the baseline firing rates of the neurons increase, the basin for the attractor with recurrent connections expands (figure 4B, red area). This leads to the prediction that when a network is driven by stronger external input and consequently has a higher initial average firing rate, it will have more loops. Numerical simulation confirms this observation (figure 5B). To quantify the degree of recurrence in a network, we define a “recurrence index” as the sum of the number of loops with less than synapses divided by the sum of similar loops in a shuffled version of the network (see Methods). Simulation results show that the recurrence index increases as a function of the initial firing rate of the network and rises rather abruptly when the initial rate exceeds , and slightly decreases when the initial rate exceeds (figure 5B). The eventual decrease of the recurrence index is due to the effect of correlations in the baseline firing, which appear in high rates and are not included in our analysis (see figure S3). The baseline correlations originate from the shared input that the neurons receive from the embedding network and is not related to their pairwise connectivity. Therefore it induces modifications to the reciprocal synapses regardless of the attractor structure explained here.
The existence of the attractor corresponding to the recurrent connections also leads to the prediction that a network modified by potentiation dominant STDP settles into higher steady-state firing rates than a network with balanced STDP, starting from the same initial conditions. The simulations confirm this prediction as well (figure 5C). The steady-state mean synaptic weight is expected to increase as a function of the initial firing rate because the basin of attraction corresponding to recurrent connections expands at high firing rates. The simulation results agree with this expectation up to the initial firing firing rate of , after which the steady-state mean decreases slightly (figure 5D). As in the case of the recurrence index (figure 5B) this decrease is due to the baseline correlations that appear at high firing rates.
If depression dominates over potentiation in STDP ( and in our examples), the coefficient in equations (2) is larger than (see Text S1), and the baseline parameter is negative. For these conditions, both elements of the fixed point of the weights, and , are negative, which is once again outside of the allowed range of synaptic values. In this case, however, the fixed point is a saddle node, which attracts trajectories from one direction and repels them from the other (see Text S1).
As before, we consider two neurons with equal baseline firing rates. The weight trajectories tend to move toward the outlying fixed point in the direction that passes through the origin (; see figure 6A, arrows). This tendency makes the origin an attractor of the dynamics within the allowed range of synaptic weights. This attractor correspond to completely disconnected neurons. Because the outlying fixed point is a saddle node, the trajectories also tend to drift away from it in the direction perpendicular to the positive-slope diagonal. This tendency produces attractors corresponding to unidirectional connections (figure 6, top-left and bottom-right). Once again, trajectories that hit the borders perpendicularly partition the weight space into three basins of attractions corresponding to each attractor (figure 6).
The dynamics of the synaptic pair we have considered suggests that some pairs of neurons in a network should become disconnected when depression dominates over potentiation. This is a more potent mechanism for eliminating loops than the previous cases, so we expect that STDP with dominant depression eliminates more loops in a large network than the other forms we have considered. Numerical simulations confirm that, indeed, there are fewer loops in the steady-state of a network with depression-dominated STDP compared to the previous cases (compare figure 7A to figures 3A and 5A). In addition, the number of disconnected pairs is large, as predicted (figure 7B).
When the baseline rates of the two neurons increase, the basin of the attractor corresponding to disconnected pair becomes larger (figure 6B). In a newtork, when neurons become excessively active, more connections should thus be eliminated, and the average rate should return to a lower value. Thus, the steady-state firing rate of a network with depression-dominated STDP should be lower than that of a network with balanced STDP starting from the same initial conditions. Simulation results corroborate this observation by showing that the steady-state firing rate of the network increases moderately as a function of the initial firing rate (figure 7C, compare with figure 3E), so depression dominant STDP implements a partial buffering of steady-state firing rates. Finally, the mean synaptic weight is a decreasing function of the initial firing rate in this case (figure 7D).
The rightward shifted STDP model, in which nearly synchronous pre- and postsynaptic action potentials induce depression, has been shown to stabilize the distribution of the synaptic weights converging onto a single neuron. The rightward shift can arise from the finite rise time of activation of NMDA receptors [10]. Here, we study this model in the context of a network. The restriction of spike pairings that induce plasticity to those between nearest neighbor pre- and postsynaptic spikes, which is necessary in this case [10], makes the dynamics of the pair of weights more complicated than in the previous cases, because the coefficients , and in equations (2) depend on the baseline firing rates (see Text S1). Furthermore, the coefficient can become negative at high firing rates, which makes the behavior of the system even more complicated. However, if we divide the analysis into three different rate regimes, we can elucidate the full range of behaviors. If depression dominates over potentiation in this model, the synaptic dynamics will be tantamount to the depression-dominant unshifted STDP described above, and the shift only makes depression even more dominant. Novel properties of this model only arise when potentiation dominates over depression, thus we assume that the potentiation domain is larger than the depression domain ( and as in [10]), and we set the amount of the shift to be .
When the initial baseline firing rates of the two neurons are low, the coefficients , and are all positive. This is because the pairing intervals are not typically short enough to fall into the depression domain caused by the shift. In addition, the coefficient is slightly smaller than . This makes the fixed point for the weights positive and large, meaning that once again it falls out of the putative range of allowed synaptic weights, but this time on the positive not the negative side (figure 8A). We use the term “putative” here because, as we will see, the upper limits on the synaptic weights are not actually required in this case. The fixed point is a saddle node (see Text S1) and attracts the trajectories of weights along the direction toward the top-right corner of the state space (figure 8A, arrows), which corresponds to recurrent connections. This case is qualitatively similar to what we found for STDP with dominant potentiation (compare figures 8A and 4A,B), so the baseline firing rate tends to become higher than its initial value and eventually the dynamics of the system falls into the regime described by figure 8B.
At higher baseline firing rates, the coefficient becomes negative. This occurs because the pairing intervals between presynaptic spikes and their causally induced postsynaptic spikes become short enough to fall into the depression domain caused by the shift. This creates a single stable fixed point for the two weights that lies within the putative allowed range of synaptic weights. Both weights are attracted to this fixed point, forming a recurrent connection (figure 8B, arrows; see Text S1), though not of maximal strength.
If the two neurons start with even higher baseline rates, the coefficients and are both negative. This follows because at very high firing rates, even the intervals between randomly paired spikes of the baseline activity are short enough to fall into the depression domain caused by the shift. This pushes the fixed point of the weights out of the allowed range (figure 8C) but, in this case, on the negative side. Because this fixed point is stable, the weights tend to approach it, creating an attractor at the origin that eliminates both weights and disconnects the neurons. This mechanism prunes the weights in the embedding network until the baseline firing rate decreases enough to make the parameter positive. Then, the regime with a stable fixed point within the allowed range (figure 8B) is restored. This is why no upper bounds on the synaptic weights are required in this case.
Combining these effects, we find that, if the shift is larger than a critical value ( in this case, see figure S1), a network will settle into a regime with a single stable fixed point within a narrow range of steady-state firing rates, regardless of the initial firing rate or the strength of the external input. The condition for this scenario to occur is that the fixed point of the weights becomes stable before it grows negative, as the initial firing rate increases. This happens when the potentiation domain is larger than depression domain and the shift is sufficiently large. The calculations show that a shift of fulfills this condition for our chosen values of potentiation and depression magnitudes (see figure S1). By generalizing from the dynamics of a pair of synapses, two predictions can be made. First, the steady-state matrix of synaptic weights should have many recurrent connections because there is no mechanism to eliminate loops, and reciprocal connections should tend to be strengthened. This prediction is confirmed by numerical simulations that show a highly recurrent steady-state connectivity (figure 9A). Second, because the pairwise connections settle into a regime with a single stable fixed point regardless of the initial baseline rate, the steady-state firing rate of the network should be resilient to changes in the external input or in the initial firing rate. Numerical simulations show that the steady-state firing rate of the network varies very slightly as a function of the initial firing rate (figure 9B). Interestingly, the narrow range of the steady-state firing rates agrees precisely with the prediction of the pairwise analysis (dashed lines in figures 9B and figure S1). Thus, rightward shifted STDP implements a homeostatic mechanism that strongly buffers the steady-state firing rates from external influences. Finally, the mean synaptic weight decreases with increased initial firing rate in this case (figure 9C).
A leftward shifted STDP model, in which synchronous pre- and postsynaptic spikes cause potentiation as a result of axonal conduction delays, has been shown to have a desynchronizing effect on population bursts and a synchronizing effect on random spiking in a recurrent network [14]. Here, we study this model within the framework of pairwise analysis. As in the previous section, we consider the interactions of nearest-neighboring spikes. If potentiation dominates over depression in this model, the synaptic dynamics will be tantamount to the potentiation-dominant unshifted STDP described above and the shift only makes potentiation further dominant. Therefore, in order to observe novel behaviors of this model, we assume that the depression domain is larger than potentiation domain ( and ), and we set the amount of the shift to be , i.e. the parameters are chose to be the flipped versions of those in the rightward shifted model above.
When the initial baseline firing rate is low, the coefficients and are positive () and is negative. This is because the pairing intervals are not typically short enough to fall into the potentiation domain caused by the shift. As a result the fixed point is positive, unstable in both directions, and out of the allowed range of weights (figure 10A). The weight trajectories tend to drift away from the fixed point in the direction that passes through the origin, so this behavior is qualitatively similar to what we found for STDP with dominant depression. The attractor at the origin corresponds to completely disconnected neurons, therefore the baseline firing rate tends to become less than its initial value.
For higher initial baseline firing rates, coefficient becomes negative, because the pairing intervals between pre- and postsynaptic spikes become short enough to fall into the potentiation domain caused by the shift. This turns the fixed point into a saddle node and pushes it into the allowed range of weights (figure 10B). The weights drift away from the fixed point in the directions that passes through the origin and the top-right corner, and are attracted to it in the perpendicular direction. As a result, both the origin and top-right corner turn into attractors, corresponding to disconnected and recurrently connected neurons respectively (figure 10B, closed circles). Because these two points are the only attractors of the system, the network is expected to become highly recurrent in this case and the neurons to become either recurrently connected or disconnected. This regime happens for a narrow range of initial firing rates. As the initial firing rate increases, the basin of the top-right attractor becomes larger (figure 10C). As a result, more recurrent connections form and hence the baseline firing rate increases, which eventually pushes the system into the regime described in the following paragraph.
For even higher initial baseline firing rates, not only coefficient becomes negative, but also turns positive and the fixed point is pushed out of the allowed range on the negative side (figure 10D). It remains a saddle node, so the weights are repelled from it in the direction that passes through the top-right corner, which becomes the only attractor of the system corresponding to recurrent connection. Therefore, it is expected that all the synapse in the network potentiate up to the upper limit of the weights in this case.
In summary, the above description shows that as the initial baseline firing rate increases, the networks undergoes three different phase: 1) for low initial rates it behaves similarly to depression-dominant STDP, i.e. recurrent connections are eliminated and the steady-state firing rate is partially buffered; 2) for higher initial rates the network becomes highly recurrent and the steady-state rate increases; 3) for even higher initial rates, all the weights become potentiated up to the maximum, and the firing rate is pathologically high. The simulation results confirm these predictions. When the initial rate is less than , the steady-state rate increases modestly (figure 11B, left) and the the mean of synaptic weights decreases (figure 11C, left) as a function of initial rate. The number of loops also decrease in this regime (figure 11A, blue). For higher initial rates, the mean synaptic weight and the steady-state rate increase rapidly (figures 11B-C, middle) and the network is highly recurrent (figure 11A, red). For initial rates higher than , the mean synaptic weight equals the maximum allowed value, implying that all the weights are maximally potentiated (figure 11C, right), and the steady-state rate is pathologically high. Although the simulation results qualitatively show the full range of behaviors predicted by pairwise analysis, the initial firing rate at which the transitions occur in simulations is lower than that predicted from calculations (see Text S1). This discrepancy is due to baseline correlations that appear at high rates (see figure S3). In presence of baseline correlations, the neurons tend to fire synchronously regardless of their pairwise connections, and hence the synapses get potentiated indiscriminately due to leftward shift of the STDP.
By analyzing pairwise interactions of neurons affected by STDP, we clarified how conventional pair-based STDP functions as a loop-eliminating mechanism in a network of spiking neurons and organizes neurons into in- and out-hubs, as reported in [21]. Loop-elimination increases when depression dominates, and turns to loop generation when potentiation dominates. STDP with dominant depression implements a partial buffering mechanism for network firing rates. Rightward shifted STDP can generate recurrent connections in a network and functions as a strict buffering mechanism to maintain a roughly constant network firing rate. STDP with leftward shift functions as a partial buffer of firing rates and a loop eliminator at low rates, and as a potent loop generator at higher rates.
All of our analytical results were obtained by considering the effect of imposing weight constraints on a linear system describing pairwise interactions of neurons in the presence of STDP. The effect of constraints on Hebbian plasticity has been explored before to explain the formation of visual receptive fields [26]. Our work can be viewed as an extension of this approach to a specific form of Hebbian plasticity that involves the timing of spikes, namely STDP. In the context of a recurrent network, this method can predict the outcome of STDP in shaping the connectivity of the network and qualitatively captures the direction of change of firing rates in the network. However, the steady-state firing rate of the network cannot be quantitatively calculated by this approach, since the analysis is focused on the snapshots of the weight dynamics given the current firing rates.
The network used in our numerical simulations was densely connected so that every neuron could potentially form a synaptic connection to every other one. However, our analytical results does not rely on any particular assumption about the density or sparsity of network connectivity. Instead, the results indicate that STDP can organize patterns of connectivity in particular ways within the framework provided by anatomical constraints, developmental hard-wiring and other physiological mechanisms, such as other forms of plasticity.
In a series of articles, Gilson and colleagues studied the structures that arise from STDP in a recurrent network in response to the patterns of correlations in the external input [15]–[20]. Here, we took a different approach. We focused on the network structures that arise in the absence of correlations either imposed by external input or originated from common inputs within the network, inspired by the observation that these are dramatically reduced by fast and strong recurrent inhibition [25]. Instead, we systematically studied the effect of the shape of the STDP window on the structures that arise in this decorrelated state. Our results can be viewed as a basis over which any structures induced by external correlations will be mounted.
A prominent feature of STDP is its ability to organize neurons into in- and out-hubs. The dependence of hub-formation on baseline firing rate shows how heterogeneity at the level of external inputs can influence the internal structure of a neural network. Moreover, this property of STDP can play an important protective role in pathological cases in which a sub-population of excitatory neurons fires at atypically high rates. Through STDP, most of the incoming synapses to this sub-population will be weakened to mitigate the excessive high firing rate. Decoupling of a highly active sub-population from an embedding network through STDP has been observed previously in networks with an excitation-inhibition balance [13].
A related study based on simulations of a small network showed that depending on the external input, an STDP rule that is phenomenologically similar to the triplet model [27] can either induce feedforward structures or recurrent connections, which was argued to be incompatible with simple pair-based STDP [22]. Although our study only addressed structures arising from pair-based STDP, our results show that recurrent connections can arise if potentiation dominates depression or the plasticity window is shifted. Interestingly, the dependence of the structures on external input in the case of leftward shifted STDP is similar to that of the more elaborate model studied by Clopath and colleagues [22].
A number of studies indicate that, apart from the timing of spikes, several other factors including firing rates, inhibitory inputs, dendritic spikes and neuromodulation influence plasticity induction [27]–[30]. Various STDP rules (including the multi-spike STDP models reviewed in [31]) have been proposed to incorporate some of these factors. The method we have developed can be used with these other STDP models, but we did not include an analysis of multi-spike STDP or more complex models because we did not want an excessive number of examples nor complexity in the STDP rule to obscure the basic approach and the insights that it provides.
The ability of pair-based STDP to generate recurrent connections has been shown previously [11]. Although in that case the depression domain was elongated, but the magnitude of potentiation domain was larger such that overall potentiation was dominant over depression, which agrees with our results on loop generation through STDP. Lubenov and colleagues have shown that STDP with leftward shifted window, arising from axonal conduction delays, can generate recurrent connections and thereby synchronize neurons when the network is initialized with a tonic irregular firing mode. In the bursting mode, leftward shifted STDP has the opposite effect, i.e. it eliminates loops and desynchronizes the neurons [14]. Because the networks we studied were in excitatory/inhibitory balanced state in which the firing patterns are irregular and asynchronous [28]–[30], our findings about loop generation through leftward shifted STDP agree with the results of [14], even though the same model can function as a loop eliminating mechanism at low initial firing rates.
A combination of axonal, synaptic and dendritic propagation delays can induce a leftward shift in STDP window [14]. On the other hand, the finite rise time of the NMDA receptor activation can give rise to a rightward shift in the window [10]. Thus the exact magnitude and direction of the shift depends on the relative contribution of these opposing factors. For instance, because the back-propagating postsynaptic spikes arrive at distal synapses with a longer delay than at proximal ones, leftward shifted STDP is expected to be observed more in the distal dendrites and rightward shift is expected at proximal sites. Moreover, the relative magnitude of potentiation and depression varies considerably along the dendritic tree [31]–[34]. Therefore, each of the different versions of STDP window analyzed in our study may be relevant in a particular region of the dendritic tree. A general prediction of our study is then that different regions of the dendritic tree may participate in different network structures as a result of differences in their STDP windows.
In a number of studies, clusters of three or four synaptically connected neurons have been observed in cortical slices at a higher frequency than expected from a random or distance-based connectivity pattern [35], [36]. We doubt that STDP can account for these clusters unless network synapses were unrealistically strong, so strong that the causal effect of single spikes from one neuron can pass through two or more synapses and transiently increase the firing rate of another neuron. Otherwise the effects of STDP would be restricted to mono-synaptically connected neurons, even in larger ensembles. In fact, our results show that loops of length 3 are usually the loops least affected by STDP. This can be explained by the direct effect of STDP being confined to loops of length 2. In loops of length 3, unlike longer loops, there is no contribution from reciprocally connected pairs of neurons (loops of length 2).
In conclusion, studying pairwise interactions of neurons through STDP provides a number of important insights about the structures that arise from this plasticity in large networks. This approach can be extended to networks with more complex STDP models and more structured external input.
We used leaky integrate-and-fire (LIF) model neuron in our simulations. The membrane potential of the LIF neuron obeys(3)where is the membrane time constant, is the resting potential, and is the synaptic input (see below). Although the input appears as a current, it is actually measured in units of the membrane potential (mV) because a factor of the membrane resistance has been absorbed into its definition. When the membrane potential reaches the firing threshold , the neuron fires an action potential and the membrane potential resets to the resting value .
A network of excitatory and inhibitory LIF neurons was simulated. Each neuron receives excitatory and inhibitory inputs from all the other neurons in the network. The strengths of the excitatory-to-inhibitory, inhibitory-to-excitatory and inhibitory-to-inhibitory synapses are fixed. At the beginning of each simulation, their strengths were drawn from uniform distributions ranging from 0 to , , and respectively. The excitatory-to-excitatory connections are modified by pair-based STDP as described below. They are also initialized at the beginning of each run to random values from a uniform distribution ranging between and . Although the inhibitory connections are stronger than excitatory connections (but inhibitory-to-excitatory and inhibitory-to-inhibitory connections are equally strong), the network settles into an excitation/inhibition balanced state with these initial conditions (see figure S3). In this state, individual neurons fire irregularly and asynchronously [28]–[30] and the strong recurrent inhibition causes the firing correlations due to shared input to be very week [25]. The connections are all to all and self connections are prohibited for all neurons.
Each presynaptic action potential arriving at an excitatory or inhibitory synapse induces an instantaneous jump or fall in synaptic input respectively, by an amount proportional to the appropriate synaptic weight. The input decays exponentially between presynaptic action potentials. In addition to synaptic inputs originating from the neurons within the network, the input to each neuron includes an external constant bias term and independent white noise. Taken together, the input to the excitatory or inhibitory neuron is described by(4)Here, the synaptic time constant is , denotes the full matrix of connections (, , and ) and the first sum runs over all neurons ( and for excitatory and inhibitory populations, respectively). The second sum runs over all the times of spikes produced by neuron prior to time , indexed by . The parameters and determine the mean and the variability of the input ( has not subscript because it is the same for all neurons), and satisfies and , with the brackets denoting averages. The parameter was set to to provide an average initial baseline firing rate of for the neurons in the network when is zero. In the simulations, the value of was changed systematically to modify initial firing rates. Each simulation is run until the excitatory-to-excitatory connections reached a steady-state in which the average firing rate, and the mean and variance of the synaptic weights remained constant.
To count the number of closed loops implied by the matrix of excitatory-to-excitatory synaptic weights (), we first turn the network into a directed graph [21]. This is done by comparing each synaptic weight to a threshold value , and assigning the value 1 to the synapse if its weight is greater than or equal to , and assigning a zero otherwise. This defines the adjacency matrix of the resultant directed graph, which can be written formally as(5)where is the Heaviside step function. The number of closed loops of length in the adjacency matrix is then(6)where denotes the matrix trace (the sum of the diagonal elements). To evaluate the degree of recurrence in a network, we compare the number of closed loops obtained from the above equation with the number in a randomly permuted (shuffled) version of the same matrix. This distinguishes recurrent connections formed by chance from those that arise from plasticity. In the following sections, whenever we mention the number of loops in a network, we are in fact referring to the number of loops in the adjacency matrix formed by turning the network into a directed graph as described above. In addition, when we refer to a “number” of synapses, we really refer to the number of synapses with strengths greater than the threshold . To obtain a loop count that is not biased by the overall strengths of the weights, we chose to be equal to the mean of the excitatory synaptic weights. For figures 3, 5, 7, 9, 11 ( initial rate) and 11( initial rate) , respectively, was set to .
|
10.1371/journal.pgen.1008202 | Exploring various polygenic risk scores for skin cancer in the phenomes of the Michigan genomics initiative and the UK Biobank with a visual catalog: PRSWeb | Polygenic risk scores (PRS) are designed to serve as single summary measures that are easy to construct, condensing information from a large number of genetic variants associated with a disease. They have been used for stratification and prediction of disease risk. The primary focus of this paper is to demonstrate how we can combine PRS and electronic health records data to better understand the shared and unique genetic architecture and etiology of disease subtypes that may be both related and heterogeneous. PRS construction strategies often depend on the purpose of the study, the available data/summary estimates, and the underlying genetic architecture of a disease. We consider several choices for constructing a PRS using data obtained from various publicly-available sources including the UK Biobank and evaluate their abilities to predict not just the primary phenotype but also secondary phenotypes derived from electronic health records (EHR). This study was conducted using data from 30,702 unrelated, genotyped patients of recent European descent from the Michigan Genomics Initiative (MGI), a longitudinal biorepository effort within Michigan Medicine. We examine the three most common skin cancer subtypes in the USA: basal cell carcinoma, cutaneous squamous cell carcinoma, and melanoma. Using these PRS for various skin cancer subtypes, we conduct a phenome-wide association study (PheWAS) within the MGI data to evaluate PRS associations with secondary traits. PheWAS results are then replicated using population-based UK Biobank data and compared across various PRS construction methods. We develop an accompanying visual catalog called PRSweb that provides detailed PheWAS results and allows users to directly compare different PRS construction methods.
| In the study of genetically complex diseases, polygenic risk scores (PRS) synthesize information from multiple genetic risk factors to provide insight into a patient’s inherited risk of developing a disease based on his/her genetic profile. These risk scores can be explored in conjunction with health and disease information available in electronic medical records. PRS may be associated with diseases that may be related to or precursors of the underlying disease of interest. In this paper, we demonstrate how PRS can be used in concert with the medical phenome to better understand the etiology of disease subtypes nested within a broad disease classification. This is done by examining the shared and distinct genetic risk factors across the related but heterogeneous disease subtypes and also through our comparison of the secondary associations across the phenome corresponding to the subtype specific PRS. We consider several PRS construction methods in our study. This framework of analysis is enabled by access to electronic health records and genetics data. Leveraging and harnessing the rich data resources of the Michigan Genomics Initiative, a biorepository effort at Michigan Medicine, and the large population-based UK Biobank study, we investigated the primary and secondary disease associations with PRS constructed for the three most common types of skin cancer: melanoma, basal cell carcinoma and cutaneous squamous cell carcinoma.
| The underlying risk factors of genetically complex diseases are numerous. Genome-wide association studies (GWAS) on thousands of diseases and traits have made great strides in uncovering a vast array of genetic variants that contribute to genetic predispositions to disease [1]. In order to harness the information from a large number of genetic variants, a popular approach is to summarize their contribution through polygenic risk scores (PRS). While the performance of PRS to predict disease outcomes at a population level has been modest for many diseases, including most cancers, PRS have successfully been applied for risk stratification of cohorts [2, 3] and recently have been used to screen a multitude of clinical phenotypes (collectively called the medical phenome) for secondary trait associations [4, 5]. The goal of these phenome-wide screenings is to uncover phenotypes that share genetic components with the primary trait that, if pre-symptomatic, could shed biological insights into the disease pathway and inform early interventions or screening efforts for individuals at risk. Phenome-wide studies using PRS rely on an easily-calculated single biomarker that combines information across a spectrum of genetic variants. Additionally, PRS may be routinely available in patients’ electronic health records (EHR) in the near future, making analyses based on PRS a useful route for agnostic interrogation of the medical phenome. Existing literature has explored how to construct PRS with respect to a single disease phenotype [6, 7]. In this paper, we demonstrate how polygenic risk scores can be used in concert with the medical phenome to better understand the etiology of disease subtypes nested within a broad disease classification. This is done by examining the shared and distinct genetic risk factors across the related but heterogeneous disease subtypes and also through our comparison of the secondary associations across the phenome corresponding to subtype specific PRS.
In the post-GWAS era and with the availability of large biobank data from multiple sources, there is great interest in using gene-based biomarkers such as PRS for risk stratification and exploration of disease etiology. However, it is not always clear how best to construct a PRS for a particular phenotype. A PRS of the general form ∑i=1Kβi^Gi requires specification of three things: a list of markers G1, G2, ⋯GK, the depth of the list or the number of markers (K), and the choice of the weights βi^. These choices can be based on information extracted from the latest GWAS or GWAS meta-analysis (when available), the NHGRI-EBI GWAS catalog of published results [1] (when available), or summary data for GWAS corresponding to each phenotype, e.g., from efforts that comprehensively screened the UK Biobank (UKB) phenome [8, 9]. While various methods for constructing PRS have been widely studied for predicting the primary phenotype collected through population-based sampling [6, 10], it is unknown how the different PRS will be associated with subtypes and related phenotypes and the associations PRS can help unravel across the medical phenome. The comparative performance of different PRS construction methods may depend on the phenotype of interest. For example, diseases such as depression, which are believed to involve genetic contribution of a large number of genetic variants, might perform differently than diseases such as cancer, which may involve a smaller number of causal variants. We provide important empirical results comparing different PRS construction methods in terms of their associations with secondary and related phenotypes and in terms of the associations they identify across the phenome.
In this paper, we first explore strategies for constructing a PRS using markers and weights obtained from either the latest GWAS or the NHGRI-EBI GWAS catalog that have reached genome-wide significance. We compare the PRS in terms of their performance [11] for the three most common skin cancer subtypes in the USA: basal cell carcinoma (MIM: 614740) [12], cutaneous squamous cell carcinoma [13] and melanoma (MIM: 155601) [14]. We compare the two strategies using an independent biobank of genetic, demographic, and phenotype data collected by the Michigan Genomics Initiative (MGI), a longitudinal biorepository effort within Michigan Medicine (University of Michigan) [4, 15]. Based on these results, we choose a PRS construction strategy for each skin cancer subtype for further analysis.
For the chosen PRS corresponding to each skin cancer subtype, we perform a phenome-wide association study (PheWAS) relating the PRS to the EHR-based phenome of MGI. We call such a study a PRS-PheWAS [4]. PRS-PheWAS results are then replicated using the population-based UK Biobank data. In order to identify secondary associations that are not driven by the primary phenotype, we perform an additional “exclusion” PRS-PheWAS for each skin cancer subtype in which we exclude subjects with any type of observed skin cancer [4]. These studies demonstrate similarities and differences in PheWAS results for PRS constructed for different disease subtypes, suggesting that PRS constructed for various disease subtypes can provide insight into shared and unique secondary associations. Our results further demonstrate the ability of such studies to reproduce known associations between secondary phenotypes and particular disease subtypes through use of PRS.
We then describe an approach for using PRS to more directly evaluate the shared and unique genetic architecture of disease subtypes and identify shared and unique secondary phenotype associations related to this genetic architecture. We define a new PRS for each skin cancer subtype using loci unique to that subtype’s chosen PRS. We further construct a composite PRS for general skin cancer consisting of loci common among all subtypes’ PRS. While merging distinct clinical entities into a compound PRS may seem counterintuitive in terms of specificity, such an approach may increase power to identify dermatological features through PheWAS that are shared by all three subtypes. Such features may help provide insight into the shared biological etiology and disease development across disease subtypes.
The NHGRI-EBI GWAS catalog and latest GWAS PRS construction methods are based on published GWAS studies, which only report risk variants that reached genome-wide significance (usually defined by a P-value threshold of P < 5x10-8). However, it is likely that there are additional risk variants below this threshold that could be associated with the trait but have not reached statistical significance [16]. Incorporating non-significant variants may conceivably improve the predictive power of a PRS but may also add additional random false positive signals, which in turn could dilute the discriminatory power of the true risk variants and diminish any predictive gain [6, 17]. To explore whether a PRS constructed using additional non-significant loci may perform differently than a PRS using only loci reaching genome-wide significance, we evaluated PRS constructed using publicly available genome-wide summary statistics from the UK Biobank at six different p-value thresholds both in terms of associations with skin cancer phenotypes and in terms of secondary phenotype associations. We further applied LDpred, a tool that adjusts GWAS summary statistics for the effects of linkage disequilibrium [7], to explore the performance of PRS incorporating the entire spectrum of available genetic information across the genome. There is extensive literature on constructing genome-wide PRS using random effects, shrinkage methods, or thresholding (our focus) [7, 18, 19], but none of these methods have been evaluated in a PheWAS setting.
In this paper, we focus our attention on skin cancer, but the approaches used in this paper can be applied to other diseases with well-defined molecular subtypes (for example, breast cancer with ER and HER2-defined subtypes). We chose to use skin cancer as a demonstrative example for a variety of reasons. First, our discovery dataset (MGI) is particularly enriched for skin cancer cases due to the strong skin cancer clinical program at Michigan Medicine and due to the high rate of surgery for skin cancer patients. MGI primarily recruits participants undergoing surgery and is therefore enriched for cancers and other medical comorbidities when compared to a general population [4]. Additionally, skin cancer has well-defined subtypes, which allows us to explore performance of subtype-specific PRS. Skin cancer also provides a setting in which there may be genetic factors uniquely related to particular subtypes as well as genetic factors that are shared risk factors for all skin cancer subtypes. The various PRS–phenotype associations identified in this paper demonstrate ways to explore shared and subtype-specific phenotypes, and this joint framework may provide an enhanced understanding of the genome x phenome landscape.
We introduce an online visual web catalog called PRSweb that provides PRS-PheWAS results for melanoma, basal cell carcinoma, and squamous cell carcinoma. PheWAS results are available using four different PRS construction methods explored in this paper: latest GWAS, NHGRI-EBI GWAS catalog, the UK Biobank GWAS summary statistics using different significance thresholds, and LDpred. The weights and the marker list for each PRS method can be downloaded. Furthermore, PheWAS summary statistics can be accessed from PRSweb (see Web resources), providing future investigators with readily available and useful tools to perform further analyses.
Comprehensive phenome-wide and genome-wide analyses of large biobank studies with publicly available summary statistics can be rich resources for PRS construction, especially if the trait-of-interest’s prevalence is high in the biobank. Using PRS, we can synthesize complex genetic information that can then be used to identify these shared genetic components across phenotypes. Compared to prior and existing literature, our contribution is new in four principal directions: (1) comparing various PRS construction methods in terms of their relationships with related EHR-derived phenotypes and subtypes (2) comparing PRS associations with secondary phenotypes across the phenome of MGI (academic medical center) and UK Biobank (population-based), (3) developing PRS-based methods for understanding the shared and unique genetic contribution across disease subtypes both in terms of disease biology and in terms of secondary phenotype associations, and (4) introducing a publicly accessible online visual catalog PRSweb to visually represent the PRS x phenome landscape and access summary data from PheWAS.
For each skin cancer subtype (melanoma, basal cell carcinoma, and squamous cell carcinoma), we generated four different sets of PRS: (1) based on merged summary statistics published in the NHGRI EBI GWAS catalog [1], (2) based on the latest available GWAS meta-analysis [30–32], (3) based on linkage disequilibrium (LD) clumping and p-value thresholding on publicly available GWAS summary statistics from the UK Biobank data [9], and (4) based on reweighting effect sizes of GWAS summary statistics by modeling LD with LDpred [7].
For each of the obtained SNP sets for each trait, we constructed a PRS as the sum of the allele dosages of risk increasing alleles of the SNPs weighted by their reported or reweighted log odds ratios. Restated, the PRS for subject j in MGI was of the form PRSj = ΣiβiGij where i indexes the included loci for that trait, βi is the log odds ratios retrieved from the external GWAS summary statistics for locus i, and Gij is a continuous version of the measured dosage data for the risk allele on locus i in subject j. The PRS variable was created for each MGI and UKB participant. For comparability of effect sizes corresponding to the continuous PRS across cancer traits and PRS construction methods, we transformed each PRS of the corresponding analytical data set to the standard Normal distribution using “ztransform” in the R package “GenABEL” [35].
In this study, we first constructed PRS for skin cancer subtypes using either the latest GWAS or the corresponding entries of the GWAS catalog. To compare the association between PRS and skin cancer phenotypes across different PRS construction methods, we fit the following model for each PRS and skin cancer phenotype:
logit(P(Phenotypeispresent|PRS,Age,Sex,Array,PC))=β0+βPRSPRS+βAgeAge+βSexSex+βArrayArray+βPC,
where the PCs were the first four principal components obtained from the principal component analysis of the genotyped GWAS markers and where “Array” represents the genotyping array. Our primary interest is βPRS, while the other factors (Age, Sex and PC) were included to address potential residual confounding and do not provide interpretable estimates due to the preceding application of case-control matching. Firth’s bias reduction method was used to resolve the problem of separation in logistic regression (Logistf in R package “EHR”) [36–38], a common problem for binary or categorical outcome models when a certain part of the covariate space has only one observed value of the outcome, which often leads to very large parameter estimates and standard errors.
We then evaluated each PRS’s (1) ability to discriminate between cases and controls by determining the area under the receiver-operator characteristics (ROC) curve (AUC) using R package “pROC” [39]; (2) calibration using Hosmer-Lemeshow Goodness of Fit test in the R package “ResourceSelection” [40, 41]; and (3) accuracy with the Brier Score in the R package “DescTools” [42]. These evaluations did not adjust for additional covariates. These metrics were estimated using roughly 2/3 of the matched data as a test set after fitting the above model on the remaining 1/3 of matched data, which we will refer to as the training data. We used these metrics and the logistic regression results to choose a PRS construction method to use for each skin cancer subtype moving forward. We compare these measures for various PRS-phenotype relationships for each phenotype separately, so the comparative performance of these measures is not biased by the different case-control sampling. To explore the impact of incorporating non-significant loci into the PRS construction, we further performed the above analyses with PRS constructed using UK Biobank GWAS summary statistics with different p-value thresholds. Similarly, we compared the LDpred-based PRS that assumed six different fractions of causal variants (non-zero effects) in the prior: 100%, 10%, 1%, 0.1%, 0.01%, and 0.001%. For LDpred comparisons we also report Nagelkerke’s pseudo-R2 to be consistent with the LDpred workflow [7].
Using the chosen PRS for each subtype, we conducted two PheWAS to identify other phenotypes associated with the PRS first for the 1,578 phenotypes in MGI and then for the 1,366 phenotypes from UK Biobank. To evaluate PRS-phenotype associations, we conducted Firth bias-corrected logistic regression by fitting a model of the above form for each phenotype and data source. Age represents the birth year in UK Biobank. To adjust for multiple testing, we applied the conservative phenome-wide Bonferroni correction according to the analyzed PheWAS codes (nMGI = 1,578 or nUK Biobank = 1,366). In Manhattan plots, we present–log10 (p-value) corresponding to tests of H0: βPRS = 0. Directional triangles on the PheWAS plot indicate whether a phenome-wide significant trait was positively (pointing up) or negatively (pointing down) associated with the PRS.
To investigate the possibility of the secondary trait associations with PRS being completely driven by the primary trait association, we performed a second set of PheWAS after excluding individuals affected with the primary or related cancer traits for which the PRS was constructed, referred to as “exclusion PRS PheWAS” as described previously [4]. We then constructed new PRS scores representing shared and subsite-unique genetic components and performed a PheWAS for each.
To evaluate the impact of the matching in the PRS PheWAS and exclusion PRS PheWAS analyses in more concrete terms, we performed sensitivity analyses in which we conducted the PheWAS analyses using the unmatched data.
To evaluate how well prior presence of a secondary diagnosis can identify subjects with increased risk of developing skin cancer, we created a binary variable taking the value 1 if a given subject (1) was diagnosed with the secondary diagnosis and then diagnosed with skin cancer at least 365 days after or (2) was diagnosed with the secondary diagnosis and never diagnosed with skin cancer. We then fit a Firth bias-corrected logistic regression of the following form:
logit(P(Primaryphenotypeispresent|Predictor,Age,Sex,Array,PC))=β0+βPRSI(Secondarynonskincancertrait)+βAgeAge+βSexSex+βArrayArray+βPC
where Array and PC were defined as before. Unless otherwise stated, analyses were performed using R 3.4.4 [43].
The online open access visual catalog PRSweb available at http://statgen.github.io/PRSweb was implemented using “Pandas”, a Data Analysis Library, which offers high level performance for large data structures and data analysis in the Python3 environment [44]. In combination with “Jinja2”, a templating language for Python, and “Bootstrap”, a Cascading Style Sheets (CSS) framework (see Web resources), static HTML files were compiled and allow easy and fast hosting of all PRS-PheWAS results. The interactive plots are drawn with the JavaScript library “LocusZoom.js” (see Web resources) which offers dynamic plotting, automatic plot sizing and label positioning.
Data were collected according to Declaration of Helsinki principles. MGI study participants’ consent forms and protocols were reviewed and approved by the University of Michigan Medical School Institutional Review Board (IRB ID HUM00099605 and HUM00155849). Opt-in written informed consent was obtained.
We first explored the comparative performance of various PRS construction strategies in terms of the resulting PRS associations with related phenotypes in the skin cancer setting. Table 2 provides the results.
Using each of the chosen PRS described above (mPRS, bPRS, and sPRS), we tested the association between each PRS and each of the 1,578 constructed phenotypes in MGI. For each PRS, the strongest associations were observed with dermatologic neoplasms that included overall skin cancer, melanoma, “other non-epithelial cancer of skin” (the PheWAS parent category of basal and squamous cell carcinoma), and carcinoma in situ of skin. In addition, secondary dermatologic traits such as actinic keratosis (with parent category “degenerative skin conditions and other dermatoses”), chronic dermatitis due to solar radiation (with parent category “dermatitis due to solar radiation”), and seborrheic keratosis were found to be associated with all three PRS (Fig 1 and Table K in S1 File). “Diseases of sebaceous glands”, “sebaceous cyst”, and “scar conditions and fibrosis of skin” were associated with bPRS. mPRS was most strongly associated with the melanoma phenotype (OR 1.48, 95% CI [1.41, 1.56]), while bPRS was most strongly associated with basal cell carcinoma (OR 1.65, 95% CI [1.56, 1.75]) followed closely by “other non-epithelial cancer of the skin” (OR 1.39, 95% CI [1.34, 1.44]). sPRS was most strongly associated with overall skin cancer (OR 1.34, 95% CI [1.3, 1.38]). The OR of all these phenotypes indicated an increased risk for primary and secondary traits with increasing PRS.
To substantiate the detected dermatologic associations, we reiterated the association screen of the three PRS in the matched phenome of the population-based UK Biobank data set (Fig 1). In general, stronger evidence for association was found in UKB compared to MGI. This may be driven by the larger sample sizes, e.g. a total of 13,623 skin cancer cases versus 4,503 in MGI. In the UK Biobank phenome, the large majority of the previous associations with dermatologic neoplasms were validated with the exception of the trait “dermatitis due to solar radiation”, which had substantially fewer cases in UKB compared to MGI (390 versus 2,959 cases). Unlike MGI, all three PRS were significantly associated (at the phenome-wide level) with “cancer, suspected or other” and “malignant neoplasm, other.” bPRS and sPRS were both associated with “diseases of the sebaceous glands” and “sebaceous cyst.”
In order to explore whether the identified PRS-phenotype associations were driven by the primary trait used to define the PRS (for example, as a side effect of treatment given after diagnosis with the primary trait), we performed a PheWAS for each PRS in which we excluded subjects who were cases for the primary trait or other skin cancer subtypes [4]. Results are shown in Table K in S1 File and Fig D in S1 Text. Actinic keratosis, a skin condition believed to be a precursor to non-melanoma skin cancers, remained significantly associated with the squamous cell carcinoma PRS in MGI and all three PRS in UK Biobank [46–48]. No other phenotypes were significant for MGI. “Sebaceous cyst” and its parent category “diseases of the sebaceous gland” were significant in the main UK Biobank PheWAS and remained significantly associated with basal cell carcinoma PRS and squamous cell carcinoma PRS in UK Biobank in the Exclusion PheWAS. Appendix 1 in S1 Text provides additional information on a sub-analysis of actinic keratosis as a predictor for future skin cancer.
In the PRS-PheWAS analyses, we note a striking overlap in the secondary dermatological traits significantly associated with each of the three PRS (mPRS, bPRS, sPRS). One potential explanation for this is that subjects may have more screening after an initial skin cancer diagnosis. Indeed, many subjects have multiple skin cancer diagnoses (Fig F in S1 Text). Fig 2 shows the number of risk loci shared by different PRS. Six risk loci are shared between the mPRS, bPRS, and sPRS.
This observation inspired a follow-up exploration in which we defined a PRS for each cancer subtype using the loci unique to that subtype’s chosen PRS. We call these new PRS scores mPRS-u, bPRS-u, and sPRS-u, which reflect the unique loci in the PRS for melanoma, basal cell carcinoma, and squamous cell carcinoma respectively. We also define a PRS consisting of all loci shared across the three skin cancer subtypes, which we call the shared PRS.
Table C in S1 Text shows the association between the various constructed PRS and the skin cancer phenotypes. As with mPRS, mPRS-u was most strongly associated with the melanoma phenotype. The bPRS-u score was similarly most strongly associated with basal cell carcinoma. We note that the melanoma AUC for the mPRS score was 0.61 (95% CI, [0.59, 0.62]) and is only 0.54 (95% CI, [0.52, 0.56]) for the mPRS-u score. Similarly, the basal cell carcinoma AUC for the bPRS score was 0.64 (95% CI, [0.62, 0.66]) and is only 0.57 (95% CI, [0.55, 0.59]) for the bPRS-u score. The sPRS-u score is not more strongly associated with the squamous cell carcinoma phenotype than the other skin cancer subtypes. For this reason, we do not include this PRS in further analyses. The shared PRS constructed as the unweighted sum of risk alleles of loci present in all three PRS scores (mPRS, bPRS, and sPRS) is more strongly associated with all three subtype phenotypes than the overall skin cancer phenotype.
Fig H in S1 Text shows PRS-PheWAS results using mPRS-u and bPRS-u. The scores again reveal their subtype specificity, while no notable secondary associations were observed. Although not shown here, additional exploration into the loci identified uniquely for each subtype, e.g. via pathway enrichment analyses, may provide some insight into subtype-specific biological mechanisms. Fig I in S1 Text shows PRS-PheWAS results for the shared PRS. Most strikingly, the shared skin cancer PRS was associated with the top skin cancer and dermatologic traits that were previously found to be associated with the three partially overlapping PRS constructs, suggesting that a shared genetic risk may be driving many of these secondary associations. These six underlying loci (HERC2 [MIM 605837] /OCA2 [MIM 611409], IRF4 [MIM 601900], MC1R [MIM 155555], RALY [MIM 614663], SLC45A2 [MIM 606202] and TYR [MIM 606933]) were previously found to be associated not only with skin cancer traits, but also with pigmentation traits of skin, eyes and hair (Fig 2; MIM 266300) [31, 49–68].
One of these pigmentation traits, skin tanning ability, the tendency of skin to sunburn rather than to suntan, is a well-known risk factor for all skin cancer traits [68, 69]. A PRS based on the independent risk variants of a recent GWAS meta-analysis on skin tanning ability [69] was strongly associated with overall skin cancer, melanoma, basal cell carcinoma, and squamous cell carcinoma and even outperformed the constructed PRS in some cases (Table C in S1 Text). Furthermore, the skin tanning ability PRS PheWAS identified a very similar set of traits as the shared skin cancer PRS, but in general displayed stronger associations (Fig I in S1 Text).
To explore whether a PRS incorporating non-significant loci will outperform a PRS incorporating only significant loci, we constructed PRS using loci related to the phenotype at six different p-value thresholds based on publicly available GWAS summary statistics from the UK Biobank. Larger p-values indicate greater SNP depth (with more SNPs being incorporated into the PRS).
The ICD-code-based collection of UK Biobank GWAS results did not include basal cell carcinoma or squamous cell carcinoma subtypes; rather, it included only the merged trait ‘non-epithelial cancer of skin’ (Fig B in S1 Text). Thus, we limited our assessment of the summary statistics to the overall skin cancer GWAS and the melanoma GWAS (Table J in S1 File).
Table D in S1 Text provides the results. As with the other PRS construction methods, the melanoma PRS was most strongly associated with the melanoma phenotype for all p-value cutoffs except 5x10-4. For this p-value cutoff, the melanoma PRS had similar AUC and OR for the melanoma and basal cell carcinoma phenotypes. This p-value cutoff represents the least conservative inclusion cutoff with 1,193 included loci, and its results indicated that inclusion of too many suggestive SNPs at lower thresholds may reduce PRS performance. However, we also note that the most conservative cutoff (5x10-9) produced a PRS that was based on only six loci, which had a weaker OR and AUC compared to other PRS created with less stringent cutoffs. The best performance in terms of AUC and OR relating to the melanoma phenotype were observed for p-value thresholds 5x10-7 and 5x10-8, which included 13 and 9 loci respectively. The small number of loci identified by this method at more conservative p-value cutoffs may be driven by the lower sample size for melanoma in the UK Biobank compared to the published melanoma GWAS meta-analyses (n cases = 2,691 and n cases = 6,628 respectively). We note that the melanoma PRS constructed using the UK Biobank summary statistics produced lower AUC across all p-value thresholds than was seen for the latest GWAS and GWAS catalog PRS construction methods.
Among the skin cancer subtypes, the PRS constructed for overall skin cancer was most strongly associated with basal cell carcinoma across all p-value thresholds, with odds ratios ranging from 1.4 (95% CI [1.32, 1.48]) to 1.64 (95% CI [1.55, 1.74]). Among the PRS, the overall skin cancer PRS had the greatest discrimination for the overall skin cancer phenotype. Overall skin cancer and melanoma PRS had similar performance in terms of discrimination for the melanoma phenotype across various depths. The overall skin cancer PRS tended to be more strongly associated with and have similar or slightly better discrimination for the overall skin cancer phenotype compared to the melanoma PRS, indicating that the overall skin cancer PRS was more accurate at predicting the overall skin cancer phenotype than the melanoma PRS. The overall skin cancer PRS had very similar association with and discrimination abilities for the overall skin cancer phenotype across all p-value thresholds except the least conservative (p = 5x10-4), for which the AUC and odds ratio were smaller. Overall, the highest AUCs and strongest OR signals for both PRS and all skin cancer phenotypes were found at depths of 5x10-7 and 5x10-8.
In addition to associations with the primary and related phenotypes, we explored associations between PRS constructed at various UK Biobank summary statistic depths and secondary phenotypes. Fig J (overall skin cancer) and Fig K (melanoma) in S1 Text show PRS-PheWAS results in MGI using PRS constructed at different depths. Depths of 5x10-7 and 5x10-8 produced very similar results, and other depths tended to identify fewer phenotypes associated with the corresponding PRS. Phenotypes that were associated with the PRS at other depths had weaker associations. PRS-PheWAS using the melanoma PRS and the overall skin cancer PRS produced somewhat different results. For example, the melanoma PRS at different depths did not identify strong associations with “diseases of sebaceous glands”, which is similar to previous PRS-PheWAS results for mPRS in MGI and UKB. In contrast, the overall skin cancer PRS did identify associations with “diseases of sebaceous glands” or its subcategories for all depths except 5x10-5 and 5x10-4. Fig L in S1 Text provides some additional information about the impact of depth on p-values for selected secondary associations.
We evaluated the performance of PRS constructed using the LDpred algorithm, which incorporates millions of SNPs into the PRS definition. Table E in S1 Text provides results. For the overall skin cancer PRS, a modelled 1% proportion of causal variants produced the best results in terms of AUC and OR with respect to the overall skin cancer phenotype (OR 1.30, 95% CI [1.26, 1.35]) and AUC 0.58, 95% CI [0.56, 0.60]). For the melanoma PRS, a modelled 0.001% proportion of causal variants produced the best results with respect to melanoma (OR 1.42, 95% CI [1.36, 1.49]) and AUC 0.60, 95% CI [0.58, 0.62]). This LDpred melanoma PRS performed slightly better compared to the melanoma PRS constructed using UK Biobank summary statistics at a 5x10-8 depth in terms of associations with the primary phenotype. Using the PRS with a percentage of assumed causal variants producing the best pseudo R2 statistic from Table E in S1 Text, we performed a PRS-PheWAS as shown in Figure M in S1 Text.
Table 3 summarizes the secondary phenotypes significantly associated with various PRS at the phenome-wide level. Many general skin cancer phenotypes are strongly associated with nearly all PRS. In particular, actinic keratosis and dermatitis due to solar radiation are associated with PRS for all three disease. In contrast, sebaceous cysts and “diseases of sebaceous glands” are strongly associated with PRS for basal cell carcinoma and squamous cell carcinoma but not with PRS for melanoma.
For comparison of the aforementioned PRS-PheWAS results and to provide researchers with resources for future PRS-based analyses, we developed an open access, online visual catalog PRSweb available at https://statgen.github.io/PRSweb that enables interactive exploration of the PheWAS results for each of the skin cancer subtypes and different PRS construction methods explored in this paper, for both the MGI and UK Biobank phenomes. PRSweb shows PRS-PheWAS plots with various choices of PRS in the drop-down menu (example screenshot in Fig 3) and offers downloadable PRS constructs (list of independent risk variants with corresponding weights). Mouse-over boxes offer detailed information about top results if needed, without impeding the overall user experience (grey box in Fig 3). Enrichment of cases in the upper quartiles of the PRS distribution are presented in forest plots.
Polygenic risk scores combine information from a large number of genetic variants to stratify subjects in terms of their risk of developing a particular disease. In the first aim of this paper, we demonstrate that PRS can also be used to explore shared and unique genetic risk profiles and secondary phenotype associations for related disease subtypes. We focus our attention on the setting of skin cancer, but the statistical approach can be applied to study other diseases with well-defined molecular subtypes.
For each skin cancer subtype, we constructed PRS using various PRS construction methods and evaluated their associations to the overall skin cancer phenotype and the three most common skin cancer subtypes: melanoma, basal cell carcinoma, and squamous cell carcinoma. We demonstrated that PRS constructed using EHR-derived phenotypes can sometimes (but not always) have good performance in terms of specificity for the primary phenotype. All PRS were positively associated with all skin cancer phenotypes under consideration. This suggests that each individual PRS may capture some shared genetic contributions for disease development across skin cancer subtypes.
For each skin cancer subtype, we performed a PRS-PheWAS to identify secondary phenotypes that are associated with the corresponding PRS. We identified many dermatological features in addition to the primary phenotype, indicating the ability of PRS to reproduce associations with the primary phenotype even after multiple testing corrections and covariate adjustment. The majority of these associations were replicated in a PRS-PheWAS performed for the UK Biobank phenome. Our analyses identified actinic keratosis, which is believed to be a precursor to squamous cell and basal cell carcinoma, as an independent predictor of basal cell and squamous cell carcinoma, and we demonstrated that incorporating the PRS in addition to clinical information improved discrimination for future skin cancer diagnoses (Appendix 1 in S1 Text) [46–48]. Additionally, some secondary phenotypes (for example, diseases of the sebaceous glands and sebaceous cysts) were identified in PRS-PheWAS only for the non-melanoma subtypes, which may provide some insight into some differentiating features of these subtypes.
In an additional analysis, we identified loci that were shared among all three skin cancer subtypes’ PRS. Loci overlap between the PRS for the three subtypes may indicate factors related to common biology between the subtypes. We noted that all shared loci (HERC2/OCA2, IRF4, MC1R, RALY, SLC45A2 and TYR) were also loci that had been associated with human pigmentation traits and/or harbor key genes of the biochemical pathway of melanogenesis [49, 53–61, 63, 66–70]. To more directly explore secondary associations common to all skin cancers, we constructed PRS using SNPs shared by all three skin cancer subtypes and a PRS for skin tanning ability using results from a recent GWAS meta-analysis.[69] The skin tanning ability PRS PheWAS identified a very similar set of traits to the shared PRS PheWAS, suggesting that the shared genetic component may in part represent genetic factors influencing the skin pigmentation and the skin reaction to sun exposure. In an attempt to more directly identify secondary associations unique to each subtype, we also constructed PRS using SNPs unique to each subtype’s PRS. This analysis did not identify any strong subtype-specific associations, perhaps suggesting that the main genetic drivers of skin cancer are shared across subtypes.
In this paper, we explore strategies for constructing a PRS using markers and weights obtained from various publicly-available sources. We compare three general strategies for PRS construction. In the first strategy, we consider PRS constructed using a small number of markers and weights identified from either the latest GWAS or GWAS meta-analysis or the NHGRI-EBI GWAS catalog. We first compare these two PRS construction methods in terms of their associations with related and unrelated EHR-derived phenotypes. A priori, we may have some belief that the latest (and often the largest) GWAS may provide a better source of evidence to use for PRS construction due to larger sample sizes and (potentially) more carefully curated data. The latest GWAS and GWAS catalog methods produced PRS with similar performance in terms of their associations with and discrimination for the primary phenotype used to construct the PRS for both basal cell carcinoma and melanoma. The latest GWAS method produced better results for squamous cell carcinoma.
In the second PRS construction strategy, we use UK Biobank summary statistics at different p-value depths to construct PRS. We found that incorporating additional loci that did not reach genome-wide significance did improve the PRS’ ability to discriminate cases from controls for the primary phenotype up to a point. In particular, PRS constructed using SNPs with p-values less than 5x10-8 or 5x10-7 resulted in the best performance, but further increasing the p-value threshold resulted in reduced performance. Crucially, we also observed stronger associations between the PRS and secondary phenotypes for PRS constructed using depths of 5x10-8 and 5x10-7. These results suggest that some benefit may be observed by incorporating loci that do not reach significance into the PRS construction but incorporating too many loci with larger p-values may not improve the predictive ability of the PRS (for both primary and secondary phenotypes). However, this gain or reduction in PRS performance may depend on the phenotype of interest and on the prevalence of the phenotype in the analytical sample.
In the third PRS construction strategy, we use the LDpred method to construct PRS using the whole spectrum of observed genetic information. For melanoma, the LDpred PRS which modelled smaller fractions of causal variants were favored over the ones modelling larger fractions. While the PRS construction with LDpred performed similarly to the various depth approach in our particular study, its computational cost was substantially higher. However, recent work indicated that LDpred might outperform pruning and thresholding approaches when larger training data sets are available [71].
In our study, PRS associations with the primary phenotype were generally stronger using the latest GWAS and GWAS catalog-based PRS than for the PRS constructed using LDpred or using UK Biobank summary statistics at different depths. Since the underlying case numbers in the discovery GWAS for the former were substantially larger than the case numbers in the UK Biobank GWAS, no direct comparison between approaches was possible, and also because the required full summary statistics of the contributing and larger skin cancer case-control studies of the latest GWAS and GWAS catalog entries were not made available. Additionally, these simpler PRS construction strategies appeared to more clearly differentiate between related subtypes than the genome-wide PRS construction methods. All three PRS construction strategies produced many similar secondary associations from PRS-PheWAS, as shown in Table 3. Overall, these simpler PRS construction methods worked well in our particular skin cancer setting, and we did not see any improvements to using a much larger number of SNPs in the PRS construction. Future releases of full summary statistics from large skin cancer GWAS meta-analyses will enable more liberal thresholds and consequently may result in better performing PRS [3, 72, 73].
As a product of this study, we provide an online visual catalog PRSweb (see Web resources) that provides PRS-PheWAS results for the various skin cancer phenotypes for PRS constructed using the different methods explored in this paper. PRSweb will provide a routine way to compare different PRS construction methods and to explore PRS-PheWAS results in detail. Additionally, PRSweb provides the PRS construction details, which researchers can download and use in their own analyses. In the future, we plan to extend this online platform to include PheWAS for many other cancer phenotypes, which will make this online platform a general tool for identifying phenotypes related to particular types of cancer.
One limitation of the generalizability of this study comes from the homogeneous race profile of MGI and UK Biobank. UK Biobank consists of subjects of primarily European descent, and we restricted our analyses to subjects of European descent in MGI (excluding about 10% of the subjects in MGI) in order to ensure greater comparability between the two datasets. Additionally, many of the existing GWAS were conducted on European populations, and we wanted to consider similar samples when comparing the performance of PRS constructed using summary statistics from European populations. Unlike UK Biobank, MGI is not a population-based sample; rather, it is a sample of patients recruited from a large academic medical center. Patients were recruited prior to surgery through the anesthesiology department, and therefore they may present a potential for selection bias. PRS-PheWAS results are susceptible to collider bias caused by PRS relationships with both skin cancer diagnosis and other diseases related to sampling. The exclusion PheWAS strategy attempts to overcome this obstacle by evaluating PRS associations with secondary phenotypes only in the subjects without a skin cancer diagnosis. This approach does not remove the possibility of sampling bias, but it may help reduce the impact of sampling on the estimated PRS-phenotype associations. The chosen design of matched case-control studies can reduce bias and detection of false positives compared to the unmatched analysis, but it is typically less powerful than the unmatched analysis (Fig N in S1 Text).
An additional limitation for all EHR-based phenome-wide studies is the potential for bias due to phenotype misclassification. In S1 Section 1, we discuss this issue in more detail, and we provide a sensitivity analysis exploring the impact of misclassification on study results in Fig O in S1 Text. In spite of these limitations, a principled comparison of the various methods explored in this paper may provide researchers with a sense of the robustness of their PheWAS inference to the PRS construction method and an analytical framework for the exploration of shared genetic architecture of related traits.
PRSweb; https://statgen.github.io/PRSweb University of Michigan Medical School Central Biorepository; https://research.medicine.umich.edu/our-units/central-biorepository UK Biobank; http://www.ukbiobank.ac.uk/ UK Biobank GWAS summary statistics; https://tinyurl.com/UKB-SAIGE TOPMed variant browser, https://bravo.sph.umich.edu/freeze5/hg38/ TOPMed program, https://www.nhlbi.nih.gov/science/trans-omics-precision-medicine-topmed-program Minimac4; https://genome.sph.umich.edu/wiki/Minimac4 BCFtools; https://samtools.github.io/bcftools/bcftools.html KING; http://people.virginia.edu/~wc9c/KING/ FASTINDEP; https://github.com/endrebak/fastindep PLINK; https://www.cog-genomics.org/plink2/ Eagle; https://data.broadinstitute.org/alkesgroup/Eagle/ UCSC Genome Browser; http://genome.ucsc.edu/ R; https://cran.r-project.org/ NHGRI-EBI GWAS Catalog; https://www.ebi.ac.uk/gwas/ dbSNP; https://www.ncbi.nlm.nih.gov/projects/SNP/ Imputation server; https://imputationserver.sph.umich.edu/ Jinja, https://github.com/pallets/jinja Locuszoom, https://github.com/statgen/locuszoom.
|
10.1371/journal.pntd.0005349 | Distinct antibody responses of patients with mild and severe leptospirosis determined by whole proteome microarray analysis | Leptospirosis is an important zoonotic disease worldwide. Humans usually present a mild non-specific febrile illness, but a proportion of them develop more severe outcomes, such as multi-organ failure, lung hemorrhage and death. Such complications are thought to depend on several factors, including the host immunity. Protective immunity is associated with humoral immune response, but little is known about the immune response mounted during naturally-acquired Leptospira infection.
Here, we used protein microarray chip to profile the antibody responses of patients with severe and mild leptospirosis against the complete Leptospira interrogans serovar Copenhageni predicted ORFeome. We discovered a limited number of immunodominant antigens, with 36 antigens specific to patients, of which 11 were potential serodiagnostic antigens, identified at acute phase, and 33 were potential subunit vaccine targets, detected after recovery. Moreover, we found distinct antibody profiles in patients with different clinical outcomes: in the severe group, overall IgM responses do not change and IgG responses increase over time, while both IgM and IgG responses remain stable in the mild patient group. Analyses of individual patients’ responses showed that >74% of patients in the severe group had significant IgG increases over time compared to 29% of patients in the mild group. Additionally, 90% of IgM responses did not change over time in the mild group, compared to ~51% in the severe group.
In the present study, we detected antibody profiles associated with disease severity and speculate that patients with mild disease were protected from severe outcomes due to pre-existing antibodies, while patients with severe leptospirosis demonstrated an antibody profile typical of first exposure. Our findings represent a significant advance in the understanding of the humoral immune response to Leptospira infection, and we have identified new targets for the development of subunit vaccines and diagnostic tests.
| Leptospirosis is zoonotic disease of global importance, with over a million cases and nearly 60,000 deaths annually. Symptomatic disease presentation ranges from a mild febrile disease with non-specific symptoms to severe forms, characterized by multi-organ failure, lung hemorrhage, and death. Factors driving severe outcomes remain unclear, but the host immune response likely plays an important role. In the present study, we applied high throughput techniques to identify the antibody profiles of patients with severe and mild leptospirosis. We discovered a limited number of immunodominant antigens, specific to patients. Surprisingly, we found the antibody repertoire varies in patients with different clinical outcomes and hypothesized that patients with mild symptoms were protected from severe disease due to pre-existing antibodies, while the profile of patients with severe outcomes was representative of a first exposure. These findings represent a substantial step forward in the knowledge of the humoral immune response to Leptospira infection, and we have identified new targets for vaccine and diagnostic test development.
| Leptospirosis causes over one million cases and nearly 60,000 deaths annually, with the greatest disease burden in urban slums in tropical and subtropical countries [1–3]. Ten pathogenic Leptospira species, over 200 serovars, and a large number of mammalian reservoirs, including rats, have facilitated the emergence of leptospirosis as a major, global public health problem. Humans typically become infected through direct contact with reservoir urine-contaminated soil or water, and develop a broad spectrum of clinical manifestations, including hepato-renal failure and pulmonary hemorrhage syndrome in severe cases, which have high mortality rates [2, 4–6]. The factors contributing to disease severity remain poorly understood, but bacterial virulence, inoculum dose and the host immune response are thought to play important roles in development of severe outcomes [2, 4].
Experimental animal models of Leptospira infection have provided a majority of evidence that antibodies play a key role in protection against and clearance of Leptospira infection [7–9]. Passive transfer of whole cell leptospiral vaccine and specific anti-leptospiral antibodies (Ligs) are protective against homologous infection in animal models, demonstrating antibodies are sufficient for immunity against experimental homologous infection [10–13]. Additionally, antibodies against LPS are serovar-specific, are correlated with agglutinating antibody titers, and confer limited cross-protection against other serovars [14, 15]. Several studies have shown that leptospirosis patients develop a robust antibody response during infection, especially anti-LPS antibodies, which correspond to the majority of the antibodies produced [12, 16, 17].
The large number of pathogenic Leptospira serovars and poor cross-protection observed for anti-LPS antibodies, have made the identification of anti-Leptospira protein antibodies a high priority for vaccine and diagnostic test development [18, 19]. In support of this, immunization with an LPS-deficient Leptospira strain in experimental animal models conferred cross-protection, implicating anti-protein and other immune responses in protection against infection. [19] Additionally, our group has applied a protein microarray methodology to evaluate the antibody repertoire generated in natural Leptospira infection and identified strong antibody responses in healthy exposed individuals as well as several IgG serodiagnostic antigens specific to patients [20, 21].
Analyses of antibody immune responses against infectious agents are essential not only for diagnostic and vaccine development, but also to providing insight in the mechanisms involved in pathogenicity [22]. Protein arrays are an excellent platform that allow for the screening of antibody protein targets in a high-throughput manner, with high sensitivity and high specificity [22–24]. These elements facilitate the assessment of many analytes simultaneously and allow for the identification, quantification and comparison of individual antigenic responses following exposure to microorganisms. Our group has efficiently employed high-density proteome arrays in the characterization of antibody signatures against several infectious agents of human and veterinary importance [25–30], including Leptospira interrogans and other spirochetes [21, 31].
In the current study, we used a whole genome proteome microarray approach to describe the first comprehensive profile of the human antibody response to symptomatic Leptospira infection. We probed 192 serum samples including patients with different clinical outcomes and healthy controls, and compared their antibody profiles against L. interrogans serovar Copenhageni proteins, the serovar associated with >90% of the urban leptospirosis cases in Salvador, Brazil [32, 33]. We identified promising candidates for the development of new diagnostic tests and subunit vaccines and discovered different antibody profiles, which associated with disease severity. Lastly, the antibody kinetics suggest a majority of patients with severe leptospirosis likely have a primary infection, while those with milder disease have evidence of a secondary infection. Our results provide novel insights into the complexity of the immunity in naturally-acquired leptospirosis as well as new diagnostic test candidates.
The study protocol was approved by the institutional review board committees of Yale University and Oswaldo Cruz Foundation prior to study initiation. All participants provided written informed consent in their native language prior to sample and data collection. All samples were anonymized before research use.
All 61 patient samples were collected during active surveillance for acute leptospirosis at the Hospital Couto Maia (31 severe group patients) and the São Marcos Emergency Clinic (30 mild group outpatients) in Salvador, Brazil between years 2005–2011. Laboratory confirmation was defined as positive microagglutination test (seroconversion, four-four rise in titer, or single titer ≥ 1:800) and/or positive ELISA and/or positive PCR for Leptospira DNA as previously described [32]. Serum samples from patients with mild or severe leptospirosis were collected twice: (i) acute sample, collected at patient admittance at the health care unit and (ii) convalescent sample, collected 5–276 days after the first sampling. Controls consisted of (i) 37 sera from healthy Leptospira-unexposed (naïve) volunteers from California/US and (ii) 37 sera from healthy participants enrolled in a cohort study in a high risk urban slum community in Salvador, endemic for leptospirosis.
The entire ORFeome of Leptospira interrogans serovar Copenhageni strain Fiocruz L1-130 was amplified by PCR and cloned into pXI vector using a high-throughput PCR recombination cloning method developed by our group [34]. In this strategy, cloned ORFs were expressed with C-terminal hemaglutinin (HA) tag and N-terminal poly-histidine (His) tag. Genes larger than 3kb were cloned as smaller segments as described previously [20, 21] and the ligA and ligB genes (LIC10465 and LIC10464, respectively) were fragmented according to the repeated Big domains present in the structure of each protein (LigB Repeats 7–12, LigA Repeats 7–13 and LigA/B Repeats 1–6) [35]. After identifying the seroreactive antigens on the microarrays, the inserts in the corresponding plasmids were confirmed by nucleotide sequencing by the Sanger method.
Microarray fabrication was performed as described previously [20, 21]. Briefly, purified mini-preparations of DNA were used for expression in E. coli in vitro based transcription-translation (IVTT) reaction system (RTS Kit, Roche), following the manufacturer´s instructions. Negative control reactions were those performed in the absence of DNA template (“NoDNA” controls). Protease inhibitor mixture (Complete, Roche) and Tween-20 (0.5% v/v final concentration) were added to the reactions, which were then printed onto nitrocellulose coated glass FAST slides using an Omni Grid 100 microarray printer (Genomic Solutions). Multiple negative control reactions and positive control spots of an IgG mix containing mouse, rat and human IgG and IgM (Jackson Immuno Research) were added to the arrays. Protein expression was verified by probing the array with monoclonal anti-polyhistidine (Sigma Aldrich) and anti-hemaglutinin (Roche Applied Science) as previously described [20, 21].
Human sera samples were diluted 1/100 in Protein Array Blocking Buffer (Whatman) supplemented with 10% v/v E. coli lysate 10mg/mL (McLab) and incubated 30 min at room temperature (RT) with constant mixing prior to addition to the microarray. Arrays were blocked for 30 min with Protein Array Blocking Buffer and then incubated with diluted samples overnight at 4°C, with gentle rocking. Washes and incubation with conjugate antibodies were performed as described previously [20, 21]. Slides were scanned in a Perkin Elmer ScanArray confocal laser and intensities were quantified using QuantArray package.
Selected ORFs were cloned into pET100-TOPO plasmid (Invitrogen) for His-tagged recombinant protein expression in BL21 (DE3) Star E. coli cells, according to the manufacturer’s recommendations. Recombinant protein expression was performed with EnPresso B system (Biosilta). Briefly, pre-cultured cells were inoculated 1/100 into 3.5 mL of EnPresso B medium supplemented with Ampicilin 100 μg/mL Reagent A 1.5 U/μL and grown shaking (160 rpm) at 30°C for 16–18 hs in 24-well culture blocks. Expression was induced by the addition of 350 μL of the booster reagent supplemented with 15U/μL Reagent A and 100 mM IPTG, for 24 h at 30°C under 160 rpm shaking. Cells were then harvested and lysed with 0.05 g of Cellytic Express (Sigma) for each mL of final culture, for 30 min at RT. Lysates were applied to a Ni2+-charged resin (Qiagen) and recombinant proteins were manually purified using 20mM Tris (pH 8.0) buffers with increasing concentrations of Imidazole. Washes varied from 5 mM to 40 mM Imidazole, depending on the protein, and elution was performed with 500 mM or 1M Imidazole. Imidazole was removed by dialysis (Thermo Scientific dialysis cassettes) and the purified proteins were checked for homogeneity in 12.5% SDS-PAGE. Protein concentration was determined by the BCA method (Thermo Scientific) according to the manufacturer's recommendations.
The assay was performed as described previously [23]. Briefly, 100 ng of each purified protein was immobilized on a nitrocellulose membrane strip. A semi-automatic micro-aerolization device was used to generate parallel bands with no visible marks. The membrane was cut into 0.5 cm wide strips perpendicularly to the antigen bands. The strips were blocked for 90 min with 4% reduced-fat bovine milk diluted in PBST (PBS + 0.5% Tween 20) and then incubated for 1 h at RT with individual serum samples diluted 1:200 in PBST 0.25% BSA and 5% v/v E. coli lysate 20mg/mL. After 3 washes with PBST, the strips were incubated for 1 hour with alkaline phosphatase–labeled anti-human IgG antibody (Sigma-Aldrich) diluted 1:30.000 in PBST 0.25% BSA. The strips were then washed 3 times with PBST and revealed with Western Blue Stabilized Substrate for Alkaline Phosphatase (Bio-Rad) for 10 min. The reaction was stopped with distilled water. Strips were air-dried and scanned images were converted to gray scale before band intensity quantification with ImageJ software (found at http://rsbweb.nih.gov/ij/).
Array signal intensity was quantified using QuantArray software. Spots intensity raw data were obtained as the mean pixel signal intensity with automatic correction for spot-specific background. Data was normalized by dividing the raw signal for each IVTT protein spot by the median of the sample-specific IVTT control spots (fold-over control [FOC]) and then taking the base-2 logarithm of the ratio (log2 FOC). Conceptually, a normalized signal of 0.0 is equal to control spot signal, and a normalized signal of 1.0 is 2-fold higher than control spot signal.
When evaluating a protein spot as reactive or non-reactive, normalized signals >1.0 were considered reactive. These designations were used to evaluate response frequency and to identify a subset of sero-reactive proteins for further analysis. A given protein on the array was considered sero-reactive if it was reactive in at least 60% of the samples in one or more of the following groups: severe disease, acute sample (n = 30); severe disease, convalescent sample (n = 30); mild disease, acute sample (n = 30); mild disease, convalescent sample (n = 30); endemic controls (n = 30); naïve controls (n = 30). Sero-reactive proteins were identified separately using IgG and IgM responses.
For each sero-reactive protein, sample groups were compared using t-tests [R stats package] and the area under receiver operator characteristic curve (AUC) [R rocr package]. Proteins with t-test p-value < 0.05 after correction for false discovery [36] and AUC > 0.70 were identified as differentially reactive.
Clinical features of the leptospirosis patients participating in this study were described using frequencies and medians with interquartile (IQR) ranges calculated in Excel (Table 1). The Fisher Exact test or the Mann-Whitney test were used to compare clinical presentations of patients with mild or severe disease using GraphPad Prism 5.02 software.
The raw and normalized array data used in this study have been deposited in the Gene Expression Omnibus archive (www.ncbi.nlm.nih.gov/geo/), accession number GSE86630.
To identify antigens associated with symptomatic leptospirosis and severe disease (requiring hospitalization), we enrolled 31 patients hospitalized with suspected leptospirosis, 30 individuals treated at an urgent care facility for suspected leptospirosis, 30 individuals living in the same communities as enrolled patients (hyperendemic controls), and 30 unexposed controls (naïve controls). All patients survived and provided paired acute and convalescent sera samples. Table 1 describes patient characteristics for clinical and biochemical tests performed during hospitalization or outpatient treatment. Hospitalized patients presented with more severe disease: 77.4% had oliguric renal failure, 6.5% had respiratory failure, and 22.5% required ICU admission, while none of these outcomes were observed in outpatients. Additionally, the agglutinating antibody titers for hospitalized patients were significantly higher during acute illness and convalescence compared to patients with mild leptospirosis (p = 0.011; p = <0.0001). However, while hospitalized patients (severe disease) were older (p = 0.039) and predominantly male (p = 0.03), there were no significant differences in days of symptoms at acute or convalescent sample collections between patients with mild and severe leptospirosis (acute p = 0.085; convalescent p = 0.681). Therefore, any differences observed in outcomes were not due to duration of illness or sampling times.
In order to determine whether there is an antibody signature specific to symptomatic disease, we probed the protein arrays with a collection of 192 sera samples, including leptospirosis patients and healthy individuals living in areas with or without endemic transmission of leptospirosis. IgM and IgG probing revealed a set of 478 reactive antigens for both acute and convalescent phases, corresponding to 12.5% of all 3819 proteins and segments included on the arrays. Of these, 255 were specific for IgM, 128 were specific for IgG and 95 were recognized by both antibodies (Fig 1A). Interestingly, we detected a majority of the IgM and IgG antigens in patients with mild disease (Fig 1B and 1C). To identify antigens specific to patients with confirmed leptospirosis (serodiagnostic antigens), we then compared antigens from the sera of patients with those from healthy individuals and found 36 antigens with significantly higher IgG reactivity in leptospirosis patients than in healthy volunteers from United States or healthy individuals living in a highly endemic area in Brazil. Of these, 12 (33%) were identified during acute leptospirosis (S2 Table) and 33 (92%) during convalescence (S3 Table).
Early antigen detection during infection is critical for the development of a new diagnostic test for leptospirosis. Therefore, we first focused on serodiagnostic antigens identified during acute phase in patients with mild or severe disease. Surprisingly, we found only a limited subset of all the seroreactive antigens were significantly recognized by IgGs in patients relative to endemic and naïve control volunteers: 11 of the 128 in the mild patient group and 28 of the 55 in the severe group (Fig 2A). Of these only 5 of the 11 and 9 of the 28 were present during acute illnesss. For the mild group, the Lig proteins were the antigens with highest accuracy, especially LigA/B 1–6, with 90% sensitivity, 86% specificity and AUC of 0.916. To determine whether we could increase both sensitivity and specificity by combining the antigens, we constructed Receiver Operating Characteristic (ROC) curves for combinations of the 5 antigens to assess antigens diagnostic performance (Fig 2A). We found that combining the top two antigens LigA/B 1–6 and LigA 8–13 yielded slightly higher sensitivity (86%) and specificity (91%) than the other combinations (Fig 2A). We performed similar analyses for the 9 antigens specific to the severe group. Again, the best diagnostic accuracy was achieved with LigA/B 1–6 (AUC = 0.935, 87% sensitivity, 100% specificity) followed by LIC20276 (AUC = 0.901, 84% sensitivity, 92% specificity). When we combined both antigens, sensitivity reached 94%, and specificity was 100% (Fig 2B). For the remaining antigens, sensitivity ranged from 77% to 90% and specificity ranged from 77% to 92%. Again, other combinations did not yield better combined sensitivity and specificity (Fig 2B). Our results indicate that we have identified candidates for new leptospirosis diagnostic tests and have discovered that there may be a limited dominant antigen antibody response to Leptospira infection.
We analyzed the responses from convalescent sera to determine whether there were major shifts in antibody responses to specific antigens with time. Patients recovering from mild disease had significantly higher IgG titers for 10 antigens compared to endemic controls, while the number of antigens nearly tripled for patients with severe clinical presentation (S3 Table). Antigens identified at convalescent phase accounted for ~92% of all diagnostic antigens (33 in 36 total IgG antigens) and LigA/B 1–6 and LigB 8–12 were the antigens with best diagnostic performance for patients with severe and mild disease, respectively. While these antigens do not have diagnostic potential, they do represent possible subunit vaccine candidates as robust antibody responses were generated over the duration of illness.
To confirm the diagnostic and subunit vaccine potential of the sero-reactive antigens detected on the microarray chips, we purified six proteins from E. coli BL21 in vitro (Fig 3B), and printed onto nitrocellulose membranes. We probed the immunostrips with serum from 8 endemic controls and 20 acute-phase patients, of which 10 had mild disease and 10 had severe disease. Serum from leptospirosis patients showed greater reactivity than serum from controls, especially serum from severe patients at convalescent phase (Fig 3A). To assess the ability of these six antigens to distinguish between patients and controls, a multi-antigen ROC curve was generated (Fig 3C), and demonstrated that the six selected antigens yielded a specificity of 100% and a sensitivity of 60% for acute mild group and 90% for the remaining groups.
As there is limited knowledge of the factors contributing to leptospirosis severe disease outcomes, we compared the antibody kinetics of patients to determine whether there are differences in antibody responses based on disease severity. We first compared the global IgG and IgM reactivities against all 478 reactive antigens identified in the microarrays by comparing the summed average signal intensities for each antigen during acute illness with that at convalescence. We detected a trending increase in IgG reactivity in patients with severe leptospirosis, which reached statistical significance when we analyzed the signals from the 36 patient-specific antigens (p<0.05) (S2 Fig). We did not observe this trend in patients with mild disease. For IgM-specific antigens, we observed no significant differences for either patient group or antigen set (S2 Fig). Thus, we identified significant IgG responses increases only in the severe patient group over time.
To understand the differences in antibody kinetics in patients in more detail, we next compared the antibody responses to the 36 differentially reactive antigens at the acute and convalescent time points for each individual by two way t-test. Based on the results of each t-test the individuals were categorized as: (i) increasing, when average response to the 36 differentially reactive antigens was higher at convalescent time point than acute, and p-value < 0.05), (ii) no change (p-value > 0.05) or (iii) decreasing, when average response to the 36 differentially reactive antigens was lower at convalescent time point than acute, and p-value < 0.05. This comparison yielded vastly different profiles for patients with mild disease and severe disease. When analyzing IgG responses, we categorized 74.4% of patients with in the severe group as “increasing” versus only 29.6% of patients in the mild group (Fig 4A). When analyzing IgM responses, we categorized 32.3% of patients in the severe group as “increasing” versus only 3.3% in the mild group (Fig 4B). Additionally, 90.0% of IgM responses did not change over time in the mild group, compared to 51.6% in the severe group. Altogether, these data clearly demonstrate that leptospirosis patients with different clinical presentations generate distinct antibody profiles.
In our kinetic antibody analyses, we enrolled five patients with mild leptospirosis, which had antibody profiles that resembled those of patients with severe leptospirosis: all had increases in IgG levels over time for 10 antigens (Fig 4C and S2). Though these five patients clearly developed an antibody response more representative of patients with severe disease (S3 Fig), including a higher convalescent agglutinating antibody titer (400–12800), they did not present with any severe clinical outcomes we measured. All other clinical and laboratory features were similar to the 25 patients with mild leptospirosis (S4 Table).
Leptospirosis is a disease with a broad spectrum of clinical manifestations ranging from asymptomatic and nonspecific acute febrile illnesses to life-threatening renal failure or pulmonary hemorrhage syndrome [2, 37]. Over a million cases of severe leptospirosis occur every year. This figure represents only a faction (potentially 5–15%) of the total mild leptospirosis cases, which usually are not identified by surveillance systems. The mechanisms involved in poor disease progression remain poorly defined, but pathogen related and host factors likely contribute to this heterogeneity [2, 4]. Here, we identified 12 specific IgG antigens that differentiate acute symptomatic disease from uninfected individuals in endemic regions and therefore represent promising diagnostic candidates for an early laboratory test for the diagnosis of leptospirosis. We also identified patient-specific antigens during convalescence, which are putative subunit vaccine candidates. Lastly, we showed that patients with different clinical presentations generate distinct antibody kinetic profiles, and we hypothesize that since antibodies are protective, disease severity and the antibody signatures may indicate primary and secondary infections.
We identified 12 IgG serodiganostic antigens for acute leptospirosis. Among them are the well-known sero-reactive proteins LigA/B 1–6, LigA 8–13, LigB 8–12 and LIC10973 (OmpL1). Several published studies used the Ligs as diagnostic markers for leptospirosis [35, 38–41] as well as OmpL1, especially in combination with LipL21, LipL32 or LipL41 [42]. Our group has previously identified LIC10486 (hypothetical protein) and LIC12544 (DNA binding protein) using the protein microarray platform [21]. The remaining 6 proteins LIC10024 (adenylate/guanylate cyclase), LIC11591 (exodeoxyribonuclease VII large subunit), LIC20077 (polysaccharide deacetylase) and the hypothetical proteins LIC11274, LIC20276 and LIC12731 are promising newly identified serodiagnostic antigens, especially LIC20276, which improved diagnostic performance for severe disease in combination with LigA/B 1–6. Interestingly, patients showed antibody reactivity against several proteins annotated as hypothetical proteins, not only at acute disease, but also during convalescence. These results indicate that even though these proteins have not been assigned any function, they are indeed expressed by the bacteria and might play an important role in host infection. Further studies should be done in order to evaluate these antigens performance in different diagnostic platforms, such as ELISA and rapid tests. For diagnostic purposes, a complete validation study needs to be performed, including the probing of a more extensive sample collection, comprising more leptospirosis patients as well as healthy controls and patients with other febrile illness, such as dengue, sifilis and hepatitis A.
The results presented here are consistent with our previous findings [21]. We detected 13 of the 24 IgG antigens previously found in hospitalized patients, strengthening the diagnostic potential of those antigens and validating the protein microarray antigen discovery platform. The inclusion of 39% of L. interrogans predicted ORFeome, however, did not provide significant advantage in diagnostic antigen discovery, since only 3 out of the 1489 proteins and segments added to the microarray were serodiagnostic, indicating that the algorithm used by our group to select the proteins included in the partial microarray was effective. Indeed, 32 out of the 36 diagnostic antigens identified here fall in at least one of the enrichment categories described by our group for antibody recognition [20, 43, 44].
Leptospirosis patients and healthy controls reacted against 12% of the L. interrogans predicted ORFeome. The majority of the imunodominant antigens were IgM specific, which corresponded to >50% of the sero-reactive proteins. The high number of IgM antigens may reflect the broad and low-affinity antigen-antibody interaction typical of IgM antibodies [45, 46]. These features usually make IgM a hard indicator of reliable diagnostic tests and might have hindered the identification of IgM diagnostic targets, as they usually account for lower specificity in IgM-based serological tests and high background reactivity in negative samples [45]. Here, we had great success in detecting IgG antigens with potential use as diagnostic or vaccine targets, but further studies are needed to identify IgM antigens.
In our previous work, we have shown that healthy individuals who live in areas with endemic transmission of leptospirosis have a background IgG reactivity against leptospiral protein antigens, possibly due to the constant exposure to the pathogen [21]. As it is well known that antibodies are one of the main immune mechanisms in naturally-acquired leptospirosis [16], the presence of high IgG levels in such individuals suggests that those antibodies might play an important role in protection against the development of clinical leptospirosis. Despite this background IgG reactivity, we were able to identify antigens for which IgG levels were even higher among hospitalized leptospirosis patients, especially at the patient's convalescent sample [21]. Indeed, most of the 36 serodiagnostic antigens identified in the present study were detected in the convalescent sample of patients with severe disease. A considerably smaller number of antigens was detected in patients with the mild form, suggesting that their IgG antibody response is more similar to healthy individuals living in the same area.
The distinct antibody profiles associated with each group were not due to differences in days of symptoms. We hypothesize that patients with mild leptospirosis had a background IgG reactivity that protected them from severe clinical manifestations while the lack of such IgG response might have favored the development of severe outcomes in hospitalized patients. In general, the first contact with an infectious agent is serologically characterized by a gradual increase in IgM, with a peak on days 7–10 after pathogen exposure, followed by an increase in IgG on days 10–14. In a secondary infection, however, a robust IgG response is rapidly mounted as a consequence of the activation of memory B cells generated during the primary infection [47–49]. In the light of this, the fact that patients with mild leptospirosis maintained their IgG levels at acute serum sample, collected approximately 5 days after the onset of symptoms, and at convalescent sample, collected at least 13 days later, suggests that they mounted an anamnestic response due to a secondary leptospiral infection. In contrast, patients with the severe form showed an antibody response typical of a primary infection, with an increase in IgG levels from acute to convalescent phases.
Our results indicate that the presence of antibodies anti-leptospiral proteins may be protective against clinical severe leptospirosis and that patients with mild disease might have had previous leptospiral infection(s). However, numerous aspects can affect the host immune response against an infectious agent, including the inoculum size. Patients with severe clinical presentations might have been infected with a higher bacterial load than patients who presented the mild form, developing thereby a more intense immune response. In addition, we can't affirm that any of the patients enrolled in the present study had never been exposed to leptospira before since leptospirosis is highly endemic in their community. Nonetheless, there is a need of studies of this kind to help elucidate the immune response associated with naturally-acquired leptospirosis and we believe our work brings relevant information to the field.
|
10.1371/journal.pgen.1004464 | Arabidopsis DELLA Protein Degradation Is Controlled by a Type-One Protein Phosphatase, TOPP4 | Gibberellins (GAs) are a class of important phytohormones regulating a variety of physiological processes during normal plant growth and development. One of the major events during GA-mediated growth is the degradation of DELLA proteins, key negative regulators of GA signaling pathway. The stability of DELLA proteins is thought to be controlled by protein phosphorylation and dephosphorylation. Up to date, no phosphatase involved in this process has been identified. We have identified a dwarfed dominant-negative Arabidopsis mutant, named topp4-1. Reduced expression of TOPP4 using an artificial microRNA strategy also resulted in a dwarfed phenotype. Genetic and biochemical analyses indicated that TOPP4 regulates GA signal transduction mainly via promoting DELLA protein degradation. The severely dwarfed topp4-1 phenotypes were partially rescued by the DELLA deficient mutants rga-t2 and gai-t6, suggesting that the DELLA proteins RGA and GAI are required for the biological function of TOPP4. Both RGA and GAI were greatly accumulated in topp4-1 but significantly decreased in 35S-TOPP4 transgenic plants compared to wild-type plants. Further analyses demonstrated that TOPP4 is able to directly bind and dephosphorylate RGA and GAI, confirming that the TOPP4-controlled phosphorylation status of DELLAs is associated with their stability. These studies provide direct evidence for a crucial role of protein dephosphorylation mediated by TOPP4 in the GA signaling pathway.
| Gibberellins (GAs) are essential regulators of plant growth and development. They are tightly related to crop productivity in the first “green revolution.” GA triggers its responses by targeting DELLA proteins, the important repressors, for degradation. This process is believed to be regulated by protein phosphorylation and dephosphorylation, but there are not any reports describing the identification of phosphatases regulating this critical event. By screening an ethyl methane sulfonate (EMS)-mutagenized Arabidopsis thaliana population, we identified a protein phosphatase TOPP4, a member of protein phosphatase 1 (PP1), that acts as a positive regulator in the GA signaling pathway. TOPP4 promotes the GA-induced degradation of DELLA proteins by directly dephosphorylating these proteins. This study provides an important insight for the switch role of protein phosphorylation and dephosphorylation in GA signal transduction and sheds light on PP1 protein phosphatases in regulating plant growth and development.
| Gibberellins (GAs) are a class of major plant hormones mediating almost all physiological events during normal plant lifespan, including seed germination, leaf formation, cell elongation and flowering time control, etc [1]–[3]. In recent decades, several molecular components essential for GA signal transduction have been characterized using genetic and biochemical approaches [4]. One group of these components is nuclear-localized DELLA proteins. These proteins belong to a subset of GRAS family of putative transcriptional regulators that contain specific DELLA motifs at their N-termini and conserved GRAS domains at their C-termini. They are key repressors of the GA signaling pathway [2], [5]. In Arabidopsis genome, there are five DELLA proteins, designated as GA INSENSITIVE (GAI), REPRESSOR OF ga1-3 (RGA), REPRESSOR OF ga1-3-LIKE protein (RGL)1, RGL2, and RGL3, respectively [6]–[9]. Genetic analyses indicated that these DELLAs have overlapping and sometimes distinctive roles in regulating plant growth and development. For example, GAI and RGA are important for stem elongation [10], [11]; RGL2 regulates seed germination [9]; whereas RGA, RGL1, and RGL2 are involved in floral development [12]–[14]. A major GA signaling cascade has recently been elucidated. In the nucleus, GA is perceived by its receptor, GIBBERELLIN INSENSITIVE DWARF 1 (GID1) [15], [16]. The formation of the ligand-receptor complex enhances the interaction of GID1 with the DELLA domain of DELLA proteins [17], leading either to their direct inactivation [18] or ubiquitination by SCFSLY1/GID2 (Skp1-Cullin-F-box protein complex) E3 ligase [19]–[22]. Ubiquitinated DELLA proteins are subsequently degraded by the 26S proteasome system, triggering GA responses.
In the absence of GA, on the other hand, DELLAs are stably localized in the nucleus where they interact with other transcription factors to inhibit the transcription of GA-responsive genes [23]–[27], restraining growth and development processes in Arabidopsis [6], [9], [11], [13]. DELLAs also promote the transcription of GID1b by interacting with other transcription factors, and maintain GA homeostasis by up-regulating the expression of GA biosynthetic genes GA 20-oxidases 2 (GA20ox2) and GA 3-oxidases 1 (GA3ox1) [28]. Moreover, DELLAs are important integrators of other phytohormones, including auxin, ethylene, abscisic acid (ABA), brassinosteroid (BR), and jasmonate (JA) [29]–[31], and environmental factors, such as light [24], [25], cold [32], and salt [33]. Very recently, these proteins were found to regulate cortical microtubule organization [34]. Besides ubiquitination and glycosylation [19], [22], [35], limited evidence also suggested that DELLA proteins are regulated by reversible protein phosphorylation and dephosphorylation [36], [37]. The detailed molecular mechanisms, however, are poorly understood and the protein phosphatases involved in this process have not been reported.
Protein phosphatases 1 (PP1s) are a major group of serine/threonine (Ser/Thr) protein phosphatases. They are expressed ubiquitously in eukaryotes [38], regulating diverse cellular processes in animals [39], although their functions in plants are uncertain. In Arabidopsis, PP1s are referred to as type-one protein phosphatases (TOPPs) [40]. Previous studies indicated that they regulate embryonic development and blue light-dependent stomatal opening [41], [42]. In general, molecular mechanisms of TOPPs in regulating plant growth and development are not well studied.
Using a forward genetic approach, we identified that one of the nine TOPPs in Arabidopsis, TOPP4, is involved in GA signal transduction. Biochemical analyses revealed that TOPP4 directly interacts with and dephosphorylates DELLA proteins RGA and GAI, promoting the GA-induced destabilization of these two proteins. A novel regulatory mechanism for protein dephosphorylation in the GA signaling pathway via TOPP4 is proposed.
We isolated an extremely dwarfed mutant from a 2000 M2 ethyl methane sulfonate (EMS)-mutagenized Arabidopsis population. The dwarfed plant was back-crossed three times with wild-type Col-0 and the resulting mutant was used in all studies presented. The mutant exhibits lack of apical dominance and aberrant leaf phyllotaxy (Figure 1A–C). Compared to wild-type plants, the mutant has tiny, curled, and dark-green rosette leaves (Figure 1A–C), delayed flowering (Figure S1), smaller flowers with irregular and narrow sepals (Figure 1D), partially twisted petals and siliques (Figure 1D–E), reduced mature pollen grains in anthers (Figure 1F), and fewer seeds in mature siliques (Figure 1G). This mutant resembles GA deficient or signaling mutants, since the dwarfism, reduced rosette radius, delayed flowering time, and high chlorophyll content of the mutant are similar to those of ga1-3, ga4, gai-1, and gid1a/b/c [16], [43], [44] (Figure S1). The mature heterozygous mutant plants were semi-dwarfed with clustered siliques and no apical dominance, suggesting that the traits were inherited in a semi-dominant manner (Figure 1B). When the mutant was back-crossed to Col-0, the F2 population resulting from self pollination had a segregation ratio of 97∶257∶118 (normal plants∶semi-dwarf plants∶dwarf plants), close to the expected 1∶2∶1 segregation ratio for a semi-dominant single locus.
Map-based cloning was employed to identify the gene responsible for the mutant phenotype (Figure 1H). The mutant in Col-0 ecotype background was crossed to Landsberg erecta-0 (Ler-0). The extremely dwarfed homozygous plants were selected from the F2 population for mapping. The locus was roughly mapped to a site close to a known marker nga168 on chromosome 2. Eight newly developed insertion/deletion (In/Del) and cleaved amplified polymorphic sequence (CAPS) markers were then used for fine mapping (Table S1). The corresponding locus was eventually mapped to a 90-Kb region between markers T5I7-29008 and T28M21-47168, with three and two recombinants for each marker, respectively, in the population of about 1000 individuals. This 90-Kb region contains 32 gene loci according to the gene annotation data obtained from the Arabidopsis genome database. We sequenced all 32 genes, and found a G to A single-nucleotide substitution in At2g39840 that resulted in the conversion of threonine (Thr) to methionine (Met) in amino acid 246 which is near the C terminus of the predicted protein sequence (Figure 1I). At2g39840 encodes a protein previously named TOPP4 [40]. The mutant was therefore designated as topp4-1. The single nucleotide substitution of topp4-1 did not influence TOPP4 gene transcription and its protein level (Figure S2).
Our genetic result suggested that topp4-1 protein should have a dominant-negative effect. To confirm that, we made transgenic plants by introducing a construct containing the full-length cDNA of topp4-1 driven by a cauliflower mosaic virus (CaMV) 35S promoter (35S-topp4-1) into wild-type plants. More than 20 independent transgenic lines were obtained. All of them showed dwarfed phenotypes similar to those of the topp4-1 mutant plants, such as lacking apical dominance, abnormal leaf phyllotaxy and curled rosette leaves, and reduced sterility (Figure 2A). The severity of the defective phenotypes appeared to be positively correlated with the expression levels of the topp4-1 gene (Figure 2A–B). Most of the transgenic seedlings died before flowering. We also generated a construct containing the cDNA of topp4-1 driven by its own promoter (pTOPP4-topp4-1) and transformed it into wild-type plants. The pTOPP4-topp4-1 transgenic plants showed topp4-1 mutant-like phenotypes (Figure 2A). Furthermore, we constructed two different artificial microRNA (amiRNA) vectors that specifically target the TOPP4/topp4-1 gene (amiR-TOPP4-1 and amiR-TOPP4-2) [45]. The topp4-1 mutant could be partially rescued regarding to inflorescence height, rosette leaves, and flowering time when the mutated topp4-1 gene was knocked down by amiRNA (Figure 2C). The recovery effect was positively correlated with the knocked down level of the topp4-1 gene (Figure 2D). These results clearly indicated that the mutated topp4-1 protein caused a dominant-negative effect on plant growth.
To confirm whether the single nucleotide substitution of TOPP4 was responsible for the defective phenotypes, we transformed topp4-1 plants with a 1.6-Kb TOPP4 genomic fragment under the control of CaMV 35S promoter (35S-TOPP4) using an Agrobacterium tumefaciens-mediated floral dipping method [46]. Ten independent T1 transgenic lines were obtained, all of which showed semi-dwarfed phenotypes. Five independent lines apparently containing a single insertion were selected for further studies. Approximately 25% of the plants from each of the four lines (#2–#5) had obviously complemented phenotypes in the T2 generation: they were significantly taller than the topp4-1 background but still a little shorter than wild-type plants (Figure 3A). However, one line (#1) had a weak complemented phenotype (Figure 3A). Subsequent quantitative reverse transcription-polymerase chain reaction (qRT-PCR) revealed that this line had relatively low expression level of TOPP4 compared to the other four lines (Figure 3B). Therefore, it seemed that the inflorescence heights of the transgenic plants were positively correlated with the expression levels of TOPP4 (Figure 3A–B). The point mutation of TOPP4 was responsible for the extremely dwarfed phenotype of topp4-1. In addition, we transformed TOPP4 gene under the control of its own promoter (pTOPP4-TOPP4) into the topp4-1 mutant. The T2 transgenic lines showed increased rosette width than the topp4-1 mutant, but the height and the curled leaf phenotypes were not altered (Figure S3).
In order to analyze the phenotype of the loss-of-function mutants, we searched the SALK and GABI-Kat T-DNA insertion databases for T-DNA insertion alleles of AT2g39840. Two independent T-DNA lines, SALK_090980 and N466328, were identified by PCR-based analyses (Figure S4A–C). Neither of them had obvious mutant phenotypes. In SALK_090980, the T-DNA is inserted 92 nucleotides upstream of the initiation codon ATG (Figure 1I). This insertion did not alter the transcription level of TOPP4 (Figure S4D). In N466328, the T-DNA is located in the 3′ untranslated region, 23 bp after the stop codon TGA (Figure 1I). qRT-PCR analysis showed that, the expression level of TOPP4 is decreased in this mutant (Figure S4D). N466328 did not show obvious phenotype, possibly because it is a knock-down rather than a null mutant. It still expresses about 40% of the wild-type level of TOPP4.
To test whether further reducing the expression level of TOPP4 can finally result in a dwarfed phenotype, we transformed the two amiR-TOPP4 vectors into wild-type plants. Four plants from 86 T2 transgenic lines of amiR-TOPP4-1 and five plants from 128 T2 transgenic lines of amiR-TOPP4-2 exhibited dwarfed phenotypes. Three representative plants, amiR-TOPP4 #1-1, #2–7 and #2–38, were selected for subsequent analyses. They showed shorter inflorescences, curled leaves, decreased fertility, and retarded growth (Figures 3C–D and S5). Compared to wild type, the expression level of the TOPP4 gene in these amiRNA transgenic lines was decreased to about 30% of their wild type counterpart (Figure 3E). Correspondingly, TOPP4 protein also dramatically declined in these lines (Figure 3F).
To investigate the effect of TOPP4 on plant growth and development, we transformed wild-type plants with the 35S-TOPP4 construct. More than 20 independent lines were obtained. The constitutive expression of TOPP4 in these lines was confirmed by qRT-PCR (Figure 2E). Interestingly, all of them had enlarged organs compared to wild-type plants (Figure S6A). One of the representative transgenic lines (#7) with the highest expression level of TOPP4 was selected for subsequent analyses (Figure 2F). Overexpression of TOPP4 in wild-type plants resulted in elongated hypocotyls, increased plant height, thickened stems, and enlarged rosette leaves, inflorescences, flowers and siliques (Figures 2F and S6B–H).
To understand the expression patterns of TOPP4, a 2-Kb fragment upstream of the translation initiation codon ATG of the TOPP4 gene was fused to the β-glucuronidase (GUS) reporter gene in the binary vector pCAMBIA 1300-GUS (pTOPP4-GUS) and this vector was transformed into wild-type plants. More than 20 transgenic lines were obtained, and T2 generation transgenic plants were used for GUS staining analyses. Because all transgenic lines showed consistent expression patterns, only one representative line, pTOPP4-GUS #8, was used for further analyses. In young seedlings, GUS staining was detected in the stele of roots and hypocotyls (Figure 4A–C), the vascular bundles of cotyledons (Figure 4B), and newly emerging leaves (Figure 4C). In mature leaves of 3-week-old plants, GUS expression was observed mainly in tips, blades (Figure 4D–E), stomata (Figure 4F), and the base of trichomes (Figure 4G). Cross sections of the rosette leaves also revealed GUS activity in vascular bundles and mesophyll cells (Figure 4H). Furthermore, GUS staining was observed in pistil and stamen filaments of flowers (Figure 4I–J), as well as the apex and the base of elongating siliques (Figure 4K). In conclusion, TOPP4 is ubiquitously expressed in various organs throughout development, suggesting its diverse and crucial functions in plant developmental processes. qRT-PCR analyses were consistent with the GUS staining results and showed relatively higher expression levels of TOPP4 in stems, rosette leaves, and young siliques (Figure 4L).
To reveal subcellular localization of the TOPP4 protein, a yellow fluorescent protein (YFP)-tagged TOPP4 was transiently expressed in mesophyll protoplasts from wild-type plants. TOPP4-YFP protein was ubiquitously distributed in cells. It was mainly localized in the nucleus and at the plasma membrane (Figure 5A). TOPP4-YFP signals were also found in cytoplasm (Figure 5A). Analyses of the transient expression pattern of a green fluorescent protein (GFP)-tagged TOPP4 in Nicotiana benthamiana leaves confirmed the subcellular distribution of the TOPP4 protein (Figure 5B). The localization of TOPP4 to the plasma membrane was also verified by plasmolyzing roots of 10-day-old 35S-TOPP4-GFP plants with 0.8 M mannitol for 1 h. Confocal analysis of these roots revealed that TOPP4-GFP was associated with the plasma membrane (Figure 5C). This result was further confirmed by an immunoblotting assay using purified plasma membrane fraction (Figure 5D).
The morphological similarity of topp4-1 plants to GA deficiency or signaling mutants prompted us to examine whether the GA signal transduction is altered in topp4-1. Therefore, responses of the mutant to exogenously applied GA3 were analyzed. Although GA3 rescued the dwarfed phenotype of ga1-3 as previously reported [47], it did not affect the stem elongation of topp4-1 (Figure 6A). Moreover, GA3 increased the transcription level of GA responsive genes EXPANSIN A8 (EXP8) and PACLUBUTRAZOL RESISTANCE 1 (PRE1) in wild type and N466328 backgrounds, but almost had no effect on these genes in topp4-1 and amiR-TOPP4 #1-1 (Figure 6B). These results indicated that topp4-1 is insensitive to GA and that the GA signaling pathway is blocked in the mutant. Thus, TOPP4 is likely involved in GA signal transduction.
The level of TOPP4 protein after GA3 treatment was also examined by immunoblotting using an anti-TOPP4 antibody. A band with a molecular mass between 37 and 55KD was detected, and it was weak in N466328 (Figure 6C), demonstrating that it is the corresponding band of TOPP4 and the anti-TOPP4 antibody is specific. After exogenous GA3 treatment, TOPP4 protein was increased in wild-type plants but was almost unaltered in topp4-1 (Figures 6C and S7). This result suggested that GA may promote TOPP4 protein accumulation in wild-type plants, and this effect is apparently disturbed in topp4-1. This might be one of the reasons that topp4-1 is insensitive to GA. However, the GA3-induced TOPP4 protein accumulation was not caused by the increased transcription level of TOPP4 (Figure 6D). We also detected the TOPP4 protein level in the gid1a/b/c mutant treated with or without GA3. Compared to wild type, TOPP4 protein was significantly decreased in the gid1a/b/c mutant (Figure 6E), probably attributed to that GA could not be perceived in it. After GA3 treatment, the TOPP4 protein level was not obviously changed in the gid1a/b/c mutant (Figure 6E). These results indicated that GA enhances the TOPP4 protein level through a GA-GID1 pathway.
RGA and GAI are the main repressors of GA signaling and their accumulation causes severe dwarf phenotypes in plants [44]. Given the fact that topp4-1 also showed severe dwarfism, genetic interactions between TOPP4 and RGA or GAI were studied. We screened rga-t2 topp4-1, gai-t6 topp4-1, and rga-t2 gai-t6 topp4-1 from offsprings of a cross between topp4-1 and the DELLA penta mutant (gai-t6 rga-t2 rgl1-1 rgl3-1 SGT625-5-2, Ler background). Then, the rga-t2 topp4-1, gai-t6 topp4-1, and rga-t2 gai-t6 topp4-1 double and triple mutants were back-crossed with Col six times for subsequent analyses. At the same time, the rga-t2 and gai-t6 single mutants and the rga-t2 gai-t6 double mutant were screened from the same genetic cross as control (Figure 7A). Phenotypic analyses revealed that the loss-of-function mutants rga-t2 and gai-t6 could partially reverse the defective phenotype of topp4-1 (Figure 7A). The topp4-1 mutant had almost no inflorescences, but both rga-t2 topp4-1 and gai-t6 topp4-1 double mutants had 2–3 cm inflorescences, relatively longer than those of topp4-1 (Figure 7A–B). And the triple mutant rga-t2 gai-t6 topp4-1 was taller than the double mutants (Figure 7A–B). These results suggested that RGA and GAI are repressors of TOPP4-mediated stem elongation. TOPP4 therefore may genetically associate with RGA and GAI in the GA signaling pathway.
We also analyzed the relationship between TOPP4 and three other DELLA proteins, RGL1, RGL2, and RGL3. The rgl mutations failed to rescue the dwarfism of topp4-1 (Figure S8A). Previous studies provided evidence that RGL1, RGL2, and RGL3 may not be required for the repression of stem elongation, but may be mainly involved in seed germination and floral development [9], [12]–[14]. We next examined the flower development and seed germination in those double mutants. The defective flower morphology of topp4-1 was also observed in rgl1-1 topp4-1, rgl2-1 topp4-1, and rgl3-1 topp4-1 (Figure S8B–C). The seed germination of rgl1-1 topp4-1 and rgl3-1 topp4-1 was slightly resistant to paclobutrazol (PAC, a specific inhibitor blocking the kaurene oxidase reaction in GA biosynthetic pathway), similar to that of topp4-1, whereas rgl2-1 topp4-1 had more resistance to PAC, similar to the single mutant rgl2-1 (Figure S8D) [9]. Therefore, it seemed that TOPP4 has no genetic interaction with RGL1, RGL2, and RGL3, regarding stem elongation, flower development, or seed germination.
From the genetic results, we hypothesized that RGA and GAI could be over-accumulated in topp4-1. To test this hypothesis, the levels of GFP-RGA and GAI in wild type, topp4-1, and TOPP4 overexpressing transgenic plants were assessed by immunoblotting. The topp4-1 plants had more RGA and GAI than wild type, whereas these proteins were significantly lower in TOPP4 overexpressing plants than in wild type (Figure 8A). At the same time, the RGA protein level was also significantly increased in three amiRNA lines (Figure 8B). Thus, TOPP4 may positively regulate the degradation of RGA and GAI and the dwarfed phenotype of topp4-1 may be caused by overaccumulation of these two proteins.
The degradation of DELLA proteins required dephosphorylation at Ser or Thr [36], but the phosphatase responsible for this activity had not been identified. We considered that TOPP4 is likely a candidate for this process. This was tested by comparing the degradation of GFP-RGA in wild-type plants and the topp4-1 mutant. First, a previously reported cell-free system [36] was used in which total proteins were solubilized in a degradation buffer and incubated at 22°C for 0, 15, 30, and 60 min followed by determination of the GFP-RGA abundance by immunoblotting. GFP-RGA protein from wild-type plants was rapidly degraded after 15 min of incubation and little was detected after 60 min (Figure 8C). Degradation rate was clearly slower in the topp4-1 mutant (Figure 8C). Addition of TOPP4 protein immunoprecipitated from wild-type plants to total protein extracts of the topp4-1 mutant increased the rate of GFP-RGA degradation, although not to the level of wild type (Figure 8C). Mutated topp4-1 protein from the topp4-1 mutant did not reverse the delayed degradation (Figure 8C). These results demonstrated that TOPP4 regulates the stability of DELLA protein.
Carbobenzoxy-Leu-Leu-leucinal (MG132), a specific 26S proteasome inhibitor, is reported to block the degradation of DELLA proteins [24]. We used this inhibitor in this cell-free system to determine if the TOPP4-mediated degradation of DELLA protein is dependent on the ubiquitin-proteasome pathway. Supplementation with 100 µM MG132 strongly blocked the degradation of GFP-RGA both in wild type and topp4-1 (Figure 8C), suggesting that the 26S proteasome acts downstream of TOPP4 in the degradation of DELLA protein.
GA can rapidly induce DELLA protein degradation [19], so we tested this effect in the topp4-1 background. Seedlings were incubated in Murashige-Skoog (MS) liquid medium supplemented with 100 µM GA3 for 0, 45, and 90 min. Total proteins were extracted from these seedlings and assessed by immunoblotting analyses. GFP-RGA accumulation was rapidly reduced after 45 min treatment with GA3 in wild-type seedlings (Figure 8D), but this process was apparently delayed in the topp4-1 seedlings (Figure 8D), confirming that TOPP4 is critical for the GA-induced degradation of DELLA proteins. When seedlings were treated with both GA3 and cycloheximide (CHX, a chemical blocking protein synthesis), the degradation of GFP-RGA was still delayed in topp4-1 (Figure 8D), suggesting that this phenomenon was not affected by de novo protein synthesis. Taken together, we concluded that TOPP4 facilitates the GA-induced degradation of DELLA proteins through a 26S proteasome pathway.
The nuclear localization of TOPP4 suggests a role for regulating nuclear proteins such as transcription factors. DELLA proteins also function in the nucleus and therefore they may physically interact to each other. To test this possibility, we performed a protein-protein interaction assay. Because only RGA and GAI showed genetic relevance with TOPP4, we then only examined interactions of TOPP4 with RGA or GAI proteins both in vitro and in vivo using a number of different biochemical approaches. Recombinant Histidine (HIS)-RGA, HIS-GAI, and glutathione S-transferase (GST)-TOPP4 were purified from E. coli and an in vitro pull-down experiment was carried out. HIS-RGA or HIS-GAI was pulled down together with GST-TOPP4 using a glutathione sepharose 4B resin (Figure 9A). However, this GST bound GAI protein was gradually reduced when the amount of FLAG-topp4-1 was increased in the same reaction system (Figure 9B), indicating that mutated topp4-1 protein and TOPP4 can competitively interact with DELLA proteins. Next, TOPP4 or topp4-1 was expressed as DNA binding domain (BD) protein fusions, and RGA and GAI were expressed as transactivation domain (AD) protein fusions in yeast strain Y190. Interactions of TOPP4-BD and RGA-AD or GAI-AD were confirmed by β-galactosidase (β-gal) activity (Figure 9C). Mutated topp4-1 seemed to interact with RGA or GAI slightly more than did the wild-type TOPP4 in yeast two-hybrid assay (Figure 9C). Further, to determine the interaction of TOPP4 and RGA or GAI in planta, we performed co-immunoprecipitation (co-IP) and bimolecular fluorescence complementation (BiFC) assays. We used 35S-TOPP4-GFP plants as materials and Col as a negative control in co-IP assay. TOPP4 protein was immunoprecipitated with anti-GFP antibody and TOPP4-bound proteins were subjected to immunoblotting analysis. Both RGA and GAI were co-immunoprecipitated with TOPP4 in 35S-TOPP4-GFP, but could not be detected in immunoprecipitated complexes of Col (Figure 9D). Finally, when TOPP4-YFPN and RGA-YFPC or GAI-YFPC were transiently co-expressed in leaves of Nicotiana benthamiana, YFP fluorescence was clearly detected in nuclei, which were confirmed by 4,6-diamidino-2-phenylindole (DAPI) staining (Figure 9E). These results strongly supported the existence of in vitro and in vivo interactions between TOPP4 and the two DELLA proteins RGA and GAI.
We next asked whether phosphorylated RGA and GAI could be dephosphorylated by TOPP4. Immunoblotting analysis using anti-GFP antibody resulted in a single band of RGA (Figure 10A). After treatment with active calf intestinal phosphatase (CIP), the band had greater electrophoretic mobility, which is representative of the dephosphorylated form (Figure 10A). Similar results were obtained for GAI protein (Figure 10A). This finding was consistent with the GAI phosphorylation status reported by Fu et al. [22]. To further demonstrate that phosphorylated RGA and GAI are the substrates of TOPP4, we incubated total protein extractions from the topp4-1 mutant with GST-TOPP4 or mutated GST-topp4-1 produced by E. coli. Both GFP-RGA and GAI treated with GST-TOPP4, but not GST-topp4-1, showed increased electrophoretic mobility (Figures 10B and S9A). These results are consistent with those obtained by CIP treatment, indicating that TOPP4, but not topp4-1, can dephosphorylate phosphorylated RGA and GAI directly.
Moreover, we analyzed the post-translational modification of GFP-RGA in wild type and the topp4-1 mutant by two-dimensional gel electrophoresis (2-DE). Protein gel blots showed several spots with different isoelectric point (pI) values in wild-type, topp4-1, and 35S-TOPP4 transgenic plants. In wild type, the basic forms of GFP-RGA, which represent the dephosphorylated status, were dominant (Figures 10C and S9B), while in topp4-1, the basic forms were decreased, and the acidic forms, which represent the phosphorylated status, were increased compared to wild type (Figures 10C and S9B). In TOPP4 overexpressing plants, the basic forms of GFP-RGA were increased significantly. The spot at the more acidic side (indicated by open arrowhead) was decreased while that at the more basic side (indicated by solid arrowhead) was increased compared to wild type (Figure 10C). These results provided further confirmation that TOPP4 can dephosphorylate DELLA proteins. The phosphorylated forms of RGA were increased in topp4-1, along with its slow degradation and high accumulation in the mutant (Figure 8), suggesting that the dephosphorylated forms of DELLA proteins may facilitate their destruction.
In addition, to investigate whether TOPP4 can dephosphorylate DELLA protein in the absence of GA, we transformed 35S-TOPP4 into gai-1, a mutant in which GAI cannot be degraded through GA-GID1 for the deletion of its DELLA domain [44]. The result showed that overexpression of TOPP4 could not rescue the dwarfed phenotype of gai-1 (Figure S10), suggesting that TOPP4-mediated DELLA dephosphorylation is dependent on the formation of the GA-GID1-DELLA complex.
Reversible phosphorylation and dephosphorylation, controlled by protein kinases and protein phosphatases, respectively, is one of the most important mechanisms of the post-translational modifications of proteins. Previous studies revealed the crucial role of dephosphorylation in plant development, mediated by a protein phosphatase 2A (PP2A), a protein phosphatase 2C (PP2C), and a protein phosphatase 6 (PP6) [48]–[51]. However, the regulatory functions of PP1s in plant development are poorly understood. In this study, we report the isolation and phenotypic characterization of a topp4-1 mutant identified from EMS-mutagenized Arabidopsis plants. Our results suggested a positive role of TOPP4 in the GA signaling pathway, through regulating the stability of DELLA proteins.
Protein kinases and phosphatases play important roles in several phytohormonal signaling pathways in plants. For example, in BR signaling, brassinosteroid-insensitive 2 (BIN2) phosphorylates and inactivates the transcription factor brassinazole-resistant 1 (BZR1) to inhibit plant growth, whereas PP2A dephosphorylates BZR1-P and promotes or inhibits the expression of its downstream response genes [49]. In ABA signaling, PP2C dephosphorylates SNF1-related protein kinase 2s (SnRK2s) to block ABA-mediated stress responses [50]. A key enzyme in ethylene biosynthesis pathway, 1-aminocyclopropane-1-carboxylic acid synthase (ACS), is stabilized by phosphorylation by mitogen-activated protein kinase 6 (MPK6) and destabilized by dephosphorylation by PP2A [52], [53]. In GA signal transduction, however, the function of protein phosphorylation on the stability of DELLA proteins has remained controversial. Fu et al. [54] revealed that both protein kinases and protein phosphatases were required for the GA-induced degradation of the barley DELLA protein SLENDER (SLN1). Subsequent studies indicated that protein phosphorylation increased the interaction between DELLA proteins and SCF ubiquitin ligase [21], [22], [55]. Conversely, Itoh suggested that phosphorylation of SLENDER RICE1 (SLR1) was independent of its degradation in rice [56]. However, recent work showed that Ser/Thr phosphatase inhibitors suppressed the degradation of RGL2 and RGA in Arabidopsis [36], [57]. More recently, a rice casein kinase I named as early flowering 1 (EL1), was identified and shown to stabilize the rice DELLA protein SLR1 by phosphorylation [37].
Based on our genetic and biochemical data and previous studies [36], [57], we concluded that the stability of RGA and GAI is regulated by protein phosphorylation and dephosphorylation (Figure 10D). The phosphorylated forms of RGA and GAI are stable and active, inhibiting the GA signaling pathway in Arabidopsis, consistent with the action of SLR1 in rice [37]. TOPP4 dephosphorylates the phosphorylated RGA and GAI, targeting them for the GA-induced degradation by the ubiquitin-proteasome pathway to promote stem elongation. Degradation of RGA and GAI relieves their restraint on GA signaling. In this process, GA promotes TOPP4 protein accumulation through GID1 (Figure 6C,E), thereby enhancing the dephosphorylation and degradation of DELLAs. But in the topp4-1 mutant, GA cannot promote the accumulation of mutated topp4-1 protein and topp4-1 cannot dephosphorylate RGA and GAI (Figures 6C and 10B). The phosphorylated RGA and GAI are degraded slowly in response to GA3 (Figure 8D), and their accumulation blocks GA signal transduction, resulting in GA-related mutant phenotypes, especially severe dwarfism. In TOPP4 overexpressing plants, less RGA and GAI accumulate than in wild-type plants due to excessive dephosphorylation, leading to a taller inflorescence (Figures 2F and S6). Therefore, our data indicated that TOPP4 is a positive regulator of the GA signaling pathway and functions in stimulating stem elongation by destabilizing DELLA proteins. This dephosphorylation process is likely dependent on the formation of GA-GID1-DELLA module, because overexpression of TOPP4 could not rescue the dwarfed phenotype of gai-1 (Figure S10).
However, unlike some other phosphorylated proteins [BR-signaling kinase 1 (BSK1)] with many phosphorylation sites in plants [58], RGA protein may have very few phosphorylation sites, which was supported by the results of 2-DE (Figures 10C and S9B). Our dephosphorylation analyses showed that the difference between phosphorylated and dephosphorylated status of DELLA proteins was very little (Figure 10A,B), and their status is therefore difficult to assess. This might be one of the reasons that the protein phosphorylation-dephosphorylation modification of DELLAs has been controversial.
An important finding in this work is that the topp4-1 mutant, with a 246Thr to 246Met amino acid substitution attributed to a G-to-A single nucleotide alteration, displays severe growth defects. The 246Thr is not a conservative site in all TOPPs (Figure S11). Its mutation does not affect the interaction of topp4-1 with DELLA proteins (Figure 9C), and the mutated topp4-1 and TOPP4 can competitively interact with DELLA proteins (Figure 9B). But the mutation impairs the dephosphorylation function of topp4-1 on DELLAs (Figure 10B). This result can explain that the single mutation of TOPP4 causes a dominant-negative effect on plant growth and development, resulting in severe defects in topp4-1. The dominant-negative effect is further confirmed by genetic data: expressing either 35S-topp4-1 or pTOPP4-topp4-1 in wild type could mimic the topp4-1 mutant phenotypes (Figure 2A). pTOPP4-TOPP4 only very slightly reversed the defects of topp4-1 and even 35S-TOPP4 could not completely recover it (Figures S3 and 3A). And knocking down topp4-1 gene in the topp4-1 mutant could partially rescue the deficient phenotypes (Figure 2C). Therefore, this dominant-negative material is crucial for elucidating the distinct functions of TOPPs in Arabidopsis.
There are nine PP1s (TOPP1–TOPP9) in Arabidopsis [40], [59], and they share 90.9–99.7% amino acid similarities (Figure S11) [60]. It seems that there might be a high degree of functional redundancy among them. In this study, we demonstrated that the amiRNA lines of TOPP4 showed dwarfed phenotypes, with overaccumulated RGA protein (Figures 3C,3D,8B). In those amiRNA lines, both TOPP4 gene expression and TOPP4 protein level were dramatically reduced (Figure 3E–F). These results confirmed that TOPP4 is a major protein phosphatase in regulating GA-mediated DELLA protein degradation in Arabidopsis.
The PP1 catalytic subunit often binds the regulatory subunit to form a functional enzyme. These regulatory subunits determine the catalytic activity, target the catalytic subunit to specific subcellular compartment, and modulate the specificity of substrates [61]. There are about 100 predicted PP1-binding regulatory subunits in animals [62]. However, to date, only inhibitor-3 (inh3), Arabidopsis I-2 (AtI-2), PP1 regulatory subunit2-like protein 1 (PRSL1) have been identified in plants [41], [42], [63]. In this study, we found that TOPP4 can directly bind RGA and GAI proteins. However, these two proteins have neither PVxF motif nor SILK motif which are present in PP1 regulatory subunits [64]. It is likely that TOPP4 may require another unknown regulatory subunit for controlling the dephosphorylation of RGA and GAI in vivo. TOPP4 is ubiquitously expressed in various organs throughout different growth stages. Therefore, it may be involved in regulation of many developmental processes. The DELLA deficient mutants rga-t2 and gai-t6 only partially rescued topp4-1 phenotypes, suggesting that TOPP4 may promote plant growth also through other signaling pathways. The plasma membrane-localization of TOPP4 also implies that it may participate in many other signal transductions. Identification of the regulatory subunits of TOPP4 in different subcellular locations, tissues, and development stages of plants may provide significant insights into the molecular mechanism of this protein on plant development.
To conclude, we have identified a key phosphatase that can directly dephosphorylate DELLA proteins in Arabidopsis; and we elucidated a mechanism of TOPP4 in regulating GA-mediated stem elongation by controlling DELLA protein stability. Future work will focus on identification of a protein kinase involved in phosphorylating Arabidopsis DELLA proteins, the specific phosphorylation sites on DELLA proteins regulating their stability, and the roles of TOPP4 in other developmental processes.
EMS-mutagenized Arabidopsis thaliana (L.) Heynh transgenic line E361-1 was screened for mutant topp4-1. After back-crossing three times with the wild-type Col-0, topp4-1 plants were used for subsequent research. The T-DNA insertion mutant lines, N466328 and SALK_090980, were obtained from European Arabidopsis Stock Centre (NASC) and Arabidopsis Biological Resource Center (ABRC), respectively. Primers used for identifying homozygous lines are indicated in Table S2 on line. The single, double and triple mutants rga-t2, gai-t6, rga-t2 gai-t6, rga-t2 topp4-1, gai-t6 topp4-1, rga-t2 gai-t6 topp4-1, rgl1-1 topp4-1, rgl2-1 topp4-1, and rgl3-1 topp4-1 were generated from the cross of topp4-1 with DELLA penta mutant (gai-t6 rga-t2 rgl1-1 rgl3-1 SGT625-5-2, cs16298, from ABRC). Primers used for genotyping are indicated in Table S2 on line. pRGA-GFP-RGA line (cs6942), ga1-3 (cs3104), ga4 (cs62), gai-1 (cs63), and gid1a-2/gid1b-3/gid1c-1 (gid1a/b/c, cs16297) were ordered from ABRC. pRGA-GFP-RGA topp4-1 was generated from the cross of pRGA-GFP-RGA with topp4-1.
The topp4-1 plants from the F2 population of a cross between the topp4-1 mutant in Col-0 ecotype background and Ler-0 (cs20, from ABRC) were selected for mapping. Simple sequence length polymorphism (SSLP) markers were used to define the mutant gene to chromosome 2 [65]–[67]. The markers used in fine mapping are listed in the Table S1 on line, including In/Del and CAPS. All of the eight new markers were developed by our lab. We sequenced the 90-Kb between markers T5I7-29008 and T28M21-47168 to identify the TOPP4 gene finally.
For the complementation experiment and overexpressing transgenic line 35S-TOPP4, a 1696-bp genomic sequence consisting of the entire coding region was PCR-amplified by PCR from the genome of wild-type Col-0 with primer set 5′-GGGGTACCTCTTTGCGCGTAATTTTCT-3′ and 5′-CGAGCTCCTCAAGAAAGACCAAATCCA-3′. Underlined regions introduce Kpn I and Sac I sites, respectively. The amplified fragment was cloned into pCAMBIA 1300. Transgenic lines expressing GFP-tagged TOPP4 (35S-TOPP4-GFP), 35S-topp4-1, and 35S-TOPP4-RFP were generated by amplifying the cDNA of Col-0 or topp4-1 with a primer set 5′-TCTAGAATGGCGACGACGACGAC-3′ and 5′-GGTACCTCCTCCTCCAATCTTTGTGGACATCATGA -3′. Underlined regions introduce Xba I and Kpn I sites, respectively. The amplified fragment was cloned into pCAMBIA 1300-GFP or pCAMBIA 1300-RFP. For pTOPP4-TOPP4, the TOPP4 promoter about 2-Kb upstream of ATG was generated with a primer set 5′-CAAGCTTTTCCGACTTAATCCGGTCCA-3′ and 5′-CTCTAGACCTAATTTTTTCGACCCC-3′. Underlined regions introduce Hind III and Xba I sites, respectively. The 35S promoter of 35S-TOPP4 was replaced by this amplified fragment. For pTOPP4-topp4-1, the 35S promoter of 35S-topp4-1 was replaced by the TOPP4 promoter. For promoter analysis (pTOPP4-GUS), the promoter was generated with a primer set 5′-CAAGCTTTTCCGACTTAATCCGGTCCA-3′ and 5′-CGGGATCCCCTAATTTTTTCGACCCC-3′. Underlined regions introduce Hind III and BamH I sites, respectively. The amplified fragment was cloned into pCAMBIA 1300-GUS that we reconstructed. For transient expression in Arabidopsis protoplasts, TOPP4-YFP were generated by amplifying the cDNA of TOPP4 with a primer set 5′-GTCGACATGGCGACGACGACGAC -3′ and 5′-GGATCCAATCTTTGTGGACATCATGA -3′. Underlined regions introduce Sal I and BamH I sites, respectively. The amplified fragment was cloned into PA7-YFP.
Primers for amiRNA-TOPP4 were designed by Web MicroRNA Designer 3 (http://wmd3.weigelworld.org/cgi-bin/webapp.cgi) oligo design algorithm using the RS300 vector sequence and the amiRNA sequences of TOPP4 gene (amiR-TOPP4-1: 5′-TACCTAATTTTTTCGACGCCA-3′; amiR-TOPP4-2: 5′-TAAAATTACGCGCAAAGACTA-3′). Detailed information using overlapping PCR and primer sets is available at Web MicroRNA Designer 3 web site. The full amiRNA fold-back fragment was subsequently cloned into pCAMBIA 1300 by Xba I/Kpn I. The alignment of the nucleotide sequence for targets of amiR-TOPP4 with the same regions of other TOPPs in Arabidopsis was presented in Figure S12.
Plant transformation plasmids were introduced into Agrobacterium tumefaciens strain GV3101, and then were transformed into Arabidopsis plants using the flower-dipping method [46]. T1 transgenic lines were selected on MS [68] plates with 25 mg/L hygromycin (Solarbio, Beijing, China). Genetic and phenotypic analyses were performed mainly in the T2 generation.
The T2 transgenic plants carrying the pTOPP4-GUS construct were immersed in GUS staining solution [50 mM Na-Phosphate buffer, pH 7.0, 1 mM EDTA, 0.1% Triton X-100, 100 µg/mL chloramphenicol, 1 mg/mL 5-bromo-4-chloro-3-indolyl-β-D-glucuronic acid (X-gluc), 2 mM ferricyanide, 2 mM ferrocyanide] and incubated overnight at 37°C. Then, they were decolorized in chloral hydrate solution [8 g of chloral hydrate, 1 mL of glycerol, and 2 mL of water]. The stained tissues were observed and photographed by light microscopy (80i, Nikon). Subcellular localization of TOPP4-GFP was photographed by confocal microscopy (Olympus FluoView FV1000MPE). For plasmolysis studies, roots of 10-day-old 35S-TOPP4-GFP transgenic plants were analyzed as described previously [69] and observed by confocal microscopy.
All qRT-PCR measurements were performed using a MX 3000 Real-time PCR system (Stratagene, La Jolla, CA) with SYBR Premix Ex Taq (Takara Bio, Inc., Shiga, Japan). Total RNA (E.Z.N.A Plant RNA Kit, OMEGA Bio, tek, Norcross, GA) was extracted from 0.05 g of tissue from 2-week-old seedlings grown on MS medium or MS medium containing 10 µM GA3 (Sigma, St. Louis, MO). The cDNAs were synthesized from 1 µg of total RNA using the PrimeScript RT reagent Kit (Perfect Real Time) (Takara Bio, Inc., Shiga, Japan). We used the housekeeping gene GLYCERALDEHYDE-3-PHOSPHATE DEHYDROGENASE C SUBUNIT (GAPC) as a normalization control. All experiments were performed with three replicates. The primers used for qRT-PCR are listed in Table S2 on line.
For transient expression in Arabidopsis protoplasts, mesophyll protoplast separation and PEG4000-mediated transfection were performed as previously described [70]. TOPP4-YFP was used in this process. YFP signals were observed with a confocal microscopy. For transient expression in Nicotiana benthamiana leaves, the Agrobacterium strain containing 35S-TOPP4-GFP construct was infiltrated into leaves of 4-week-old tobacco plants, and GFP and RFP were observed 2 days after transformation by a confocal microscopy.
For inflorescence analyses, 7-day-old plants grown in soil were sprayed with 100 µM GA3 or water for control every 3 days for 3 weeks. For seed germination, seeds were grown on the MS plates contained 0, 5, and 10 µM PAC. Seed germination was scored 6 days after vernalization. Thirty to fifty seedlings were measured each time. These experiments were repeated three times independently.
The primary antibodies used in this study were anti-RGA (Agrisera, Vännäs, Sweden), anti-GFP (Invitrogen, Carlsbad, CA), anti-GST (ZSGB-Bio, Beijing, China), anti-HIS (ZSGB-Bio, Beijing, China), anti-TOPP4, and anti-GAI. Anti-TOPP4 and anti-GAI were made in our lab. Anti-GAI was generated in rabbits by immunizations with full-length protein sequence. The anti-TOPP4 antibodies were generated in rabbits using 150 amino acids at the N terminal of TOPP4. Both anti-GAI and anti-TOPP4 antibodies were affinity purified, and the specificity of them were determined using mutants gai-t6 and N466328, respectively. Immunoblotting analysis was performed as previously described [71]. For immunoblotting detection of TOPP4 in plasma membrane, plasma membrane extraction was isolated from 2-week-old wild-type seedlings and anti-GFP and anti-PIN1 antibodies (N782245, NASC) were used for detecting TOPP4 and PIN1 protein, respectively. For DELLA protein level assay, 20-day-old wild-type, topp4-1, 35S-TOPP4, and three amiRNA transgenic plants were used. For DELLA protein degradation assay, total protein extracts from 20-day-old pRGA-GFP-RGA and pRGA-GFP-RGA topp4-1 plants were prepared as previously described [36] and treated with TOPP4 immunoprecipitated from wild-type plants using anti-TOPP4 antibody, topp4-1 from the topp4-1 mutant, or 100 µM MG132 (Calbiochem, Darmstadt, Germany). For GA3 treatment, 20-day-old pRGA-GFP-RGA and pRGA-GFP-RGA topp4-1 plants were incubated in MS liquid medium containing 100 µM GA3 or 100 µM GA3 together with 50 µM CHX for the indicated time periods. For TOPP4 protein level analysis, 2-week-old wild-type, topp4-1, gid1a/b/c or 35S-TOPP4-GFP plants were treated with 10 or 100 µM GA3 for 4 h. The coomassie brilliant blue-stained rubisco small subunit (RbcS) protein was used as a loading control as indicated. Each experiment was repeated at least three times. Relative band intensities were then calculated for each immunoblot panel by Emage J.
The yeast strain Y190 was used in our experiments. Yeast transformations were performed according to the MATCHMAKER two-hybrid system 3 (Clontech, Shiga, Japan). Full-length cDNA of TOPP4 or topp4-1 fused to the DNA-binding domain of GAL4 was used as the bait protein and GAI or RGA fused to the transcriptional activation domain of GAL4 was used as the prey protein. Yeast clones containing the GAL4-BD-TOPP4/GAL4-BD-topp4-1 and GAL4-AD-GAI or GAL4-AD-RGA constructs were plated on synthetic dextrose (SD)-His-Trp-Leu medium for 5 days at 30°C to assay for interaction. The results of β-gal filter assay were observed in one hour. β-gal activity were detected according to the manufactures protocol (Clontech, Shiga, Japan). This experiment was repeated at least three times.
The GST-TOPP4, HIS-GAI, HIS-RGA proteins were expressed in E. coli BL21. The recombinant proteins were co-incubated in the presence of glutathione sepharose 4B resin (GE, Fairfield, CT), which was used to selectively bind the GST fusion proteins with PBS buffer [140 mM NaCl, 2.7 mM KCl, 10 mM Na2HPO4, 1.8 mM KH2PO4]. The bound proteins were eluted with 1× sodium dodecyl sulfate (SDS) loading buffer and analyzed with anti-GST and anti-HIS antibodies. For competitive pull-down assay, FLAG-topp4-1 was expressed in E. coli BL21. Pull-down reactions were performed in the presence of 5 µg GST-TOPP4, 2.5 µg HIS-GAI, and 1.25 µg or 2.5 µg FLAG-topp4-1. HIS-GAI and FLAG-topp4-1 were mixed, pulled-down with GST-TOPP4 and detected by anti-HIS antibody [72]. These experiments were repeated at least three times.
Co-immunoprecipitation studies of TOPP4 and RGA or GAI were performed on 10-day-old seedlings of Col and 35S-TOPP4-GFP. RGA and GAI proteins from these two materials were adjusted to the same amount. Immunoprecipitation of TOPP4 protein used anti-GFP antibody. Protein G agarose (GE, Fairfield, CT) was used to precipitate the immunoprotein complexes with IP buffer [150 mM NaCl, 50 mM Tris-HCl, pH 7.5, 1% NP-40, 1% protease inhibitors]. After immunoprecipitation, beads were washed four times with IP buffer. Proteins were then released and collected by boiling in 2× SDS loading buffer for 5 min. Immunoprecipitation products were detected by immunoblotting with RGA- or GAI-antibody, respectively. This experiment was repeated at least three times.
The full-length open reading frame sequences for the Arabidopsis TOPP4, RGA, and GAI were amplified and cloned into pEearleygate201-YN and pEearleygate202-YC BiFC vectors to generate TOPP4-YFPN, RGA-YFPC, and GAI-YFPC [73], [74]. Co-infiltration of Agrobacterium strains containing the BiFC constructs and the p19 silencing plasmid was carried out at OD 600 of 0.7∶0.7∶1.0 and infiltrated into leaves of 4-week-old Nicotiana benthamiana plants [24]. The BiFC signal was observed 3 days after infiltration using a fluorescence microscope. Leaves were incubated with 0.2 mg/L DAPI for nuclei staining.
For CIP treatment, 70 µg of total protein extracts from 20-day-old topp4-1 seedlings (for GAI assay) and pRGA-GFP-RGA topp4-1 seedlings (for RGA assay) were added with 50 U CIP (NEB, Ipswich, MA) or the same amount of denatured CIP and incubated at 37°C for 3 h and then detected by immunoblotting. For TOPP4 treatment, 70 µg of total protein extracts from 20-day-old topp4-1 seedlings (for GAI assay) and pRGA-GFP-RGA topp4-1 seedlings (for RGA assay) with buffer [50 mM Tris-HCl (pH 7.0), 0.1 mM Na2EDTA, 5 mM DTT, 0.01% (w/v) Brij 35, 1 mM MnCl2, 1 µM protease inhibitors] were added with 5 µg GST-TOPP4 or GST-topp4-1 and incubated at 30°C for 1 h, after which they were subjected to immunoblotting. An equal amount of extracts was used as the control for each treatment. The reaction was terminated by adding loading buffer. Each experiment was repeated at least seven times.
Total proteins were extracted from 20-day-old pRGA-GFP-RGA, pRGA-GFP-RGA topp4-1, and pRGA-GFP-RGA 35S-TOPP4 plants and separated by 2-DE using a 13-cm, pH 4–7 immobilized pH gradient gel (IPG) strip in Ettan IPGphor 3 Isoelectric Focusing System (GE, Fairfield, CT), and the second-dimensional separation was performed on 8% SDS PAGE gel. The same amount of RGA protein was loaded for each sample. The proteins were detected using anti-GFP antibody. This experiment was repeated three times.
Sequence data from this article can be found in the Arabidopsis Genome Initiative or GenBank/EMBL databases under the following accession numbers: TOPP4 (At2g39840), RGA (At2g01570), GAI (At1g14920), RGL1 (At1g66350), RGL2 (At3g03450), RGL3 (At5g17490), TOPP1 (At2g29400), TOPP2 (At5g59160), TOPP3 (At1g64040), TOPP5 (At3g46820), TOPP6 (At5g43380), TOPP7 (At4g11240), TOPP8 (At5g27840), TOPP9 (At3g05580), TOPP1 (Vicia faba, AB038648), PP1 (Oryza sativa, OSU31773), ser/thr PP1 (Zea mays, ZEAMMB73_175230), PPP1CC (Homo sapiens, HGNC:9283), PPP1CC (Mus musculus, MGI:104872), PP2A (At1g69960).
|
10.1371/journal.pntd.0005659 | Challenges in developing methods for quantifying the effects of weather and climate on water-associated diseases: A systematic review | Infectious diseases attributable to unsafe water supply, sanitation and hygiene (e.g. Cholera, Leptospirosis, Giardiasis) remain an important cause of morbidity and mortality, especially in low-income countries. Climate and weather factors are known to affect the transmission and distribution of infectious diseases and statistical and mathematical modelling are continuously developing to investigate the impact of weather and climate on water-associated diseases. There have been little critical analyses of the methodological approaches. Our objective is to review and summarize statistical and modelling methods used to investigate the effects of weather and climate on infectious diseases associated with water, in order to identify limitations and knowledge gaps in developing of new methods. We conducted a systematic review of English-language papers published from 2000 to 2015. Search terms included concepts related to water-associated diseases, weather and climate, statistical, epidemiological and modelling methods. We found 102 full text papers that met our criteria and were included in the analysis. The most commonly used methods were grouped in two clusters: process-based models (PBM) and time series and spatial epidemiology (TS-SE). In general, PBM methods were employed when the bio-physical mechanism of the pathogen under study was relatively well known (e.g. Vibrio cholerae); TS-SE tended to be used when the specific environmental mechanisms were unclear (e.g. Campylobacter). Important data and methodological challenges emerged, with implications for surveillance and control of water-associated infections. The most common limitations comprised: non-inclusion of key factors (e.g. biological mechanism, demographic heterogeneity, human behavior), reporting bias, poor data quality, and collinearity in exposures. Furthermore, the methods often did not distinguish among the multiple sources of time-lags (e.g. patient physiology, reporting bias, healthcare access) between environmental drivers/exposures and disease detection. Key areas of future research include: disentangling the complex effects of weather/climate on each exposure-health outcome pathway (e.g. person-to-person vs environment-to-person), and linking weather data to individual cases longitudinally.
| Unsafe water supplies, limited sanitation and poor hygiene are still important causes of infectious disease (e.g. Cholera, Leptospirosis, Giardiasis), especially in low-income countries. Climate and weather affect the transmission and distribution of infectious diseases. Therefore, scientists are continuously developing new analysis methods to investigate the impacts of weather and climate on infectious disease, and particularly, on those associated with water. As these methods are based on an imperfect representation of the real world, they are inevitably subjected to many challenges. Based on a systematic review of the literature, we identified seven important challenges for scientists who develop new analysis methods.
| The seasonal and geographic distributions of infectious diseases are currently among the best indications of an association with weather and climate. The literature on climate effects is expanding in response to concerns about global climate change. The significance of the methods and data available is not only confined to the technical procedural aspects; methods and data also impact on the formulation of the specific scientific questions, their selection, and the development of hypotheses. Although our understanding of how weather and climate affect diseases has improved, the wide range of research methods applied make it difficult to get a robust overview of the state of research.
The relationship between climate/weather and infectious diseases is complex (e.g.[1]), as shown in the example illustrated in Fig 1. Investigating the effects of weather and climate on infectious diseases requires the ability to: i) disentangle concurrent modes of transmissions (e.g. environmental from human-to-human transmission); ii) tease apart the individual effects of multiple exposures at different temporal and spatial scales; iii) identify and separate socio-economic drivers and behavioural causes; iv) integrate all these different processes into a unified perspective; v) attribute changes in disease to observed environmental changes (such as climate change); and vi) quantify infectious disease burden resulting from current social, economic and environmental conditions which can help to project the future disease burden resulting from these changes. These are difficult methodological and conceptual demands, and the scientific and public health community could benefit from a critical overview of the available research methods and the challenges ahead.
In this paper, we focus on the particularly important, especially in developing countries, class of infectious diseases associated with water (including those classified as neglected tropical diseases (NTD) according to the World Health Organisation (WHO) [2] the US Centers for Disease Control and Prevention (CDC) [3]) and the journal Plos NTD [4]) (Table 1). According to WHO estimates, 1.1 billion people globally drink water that is of at least ‘moderate’ risk of faecal contamination [5], and 842,000 annual deaths are attributable to unsafe water supply, sanitation and hygiene (including 361,000 deaths of children under age five), mostly in lower income countries [6,7].
Infectious diseases associated with water are classified as follows: “water-system-related” infections (i.e. via aerosols from poorly managed cooling systems, e.g. Legionellosis), “water-based” infections (i.e. via aquatic vectors or intermediate in hosts, e.g. Schistosomiasis), “water-borne” infections (i.e. via bacterial, parasitic and viral oral-faecal infection through ingestion, e.g. cholera), and “water-washed” infections (i.e. infections arising from poor hygiene due to insufficient water, these can also include oral-faecal infection, e.g. hookworm) [8] (see Table 1). Here and throughout, we use the expression “water-associated” to refer to these latter classes of diseases. Of note, we excluded diseases arising from ingestion/contact with inorganic and other chemical compounds (e.g. arsenic) and vector-borne infections linked with water (e.g. malaria, rift valley fever, river blindness) from the “water-associated” diseases.
This Review is not a prescriptive guideline of available methods for a range of problems. We reviewed and summarised the methods used to investigate the effects of weather and climate on infectious diseases associated with water, with the objective of identifying the challenges that scientists are facing when develop new analysis methods. We focused on quantitative analytical approaches, such as: mathematical models, statistical analysis, computational techniques, numerical simulations, epidemiological models and computer-generated agent based models. We excluded purely descriptive observational studies [9].
Discriminating between studies which build explanatory models versus create predictive models is particularly important in statistical modelling [10]. We however avoided this way of grouping. The dichotomy explanatory vs predictive models might be clear from an epistemological point of view [10], nevertheless, we have found it really challenging to rigorously separate papers according to this classification. For most papers, a formal distinction is often impossible as the causal relationships are inferred/discussed from the patterns captured from predictive models, and vice versa the hypothetical-deductive models (e.g. driven by causal relationships), could both be able to predict a range of future scenarios.
The methods for the systematic review followed the Guidelines developed by the Cochrane Collaboration [11]. We searched for English language articles published from 2000 to April 2015. The following databases were searched: Scopus, Medline, EMBASE, CINAHL, Cochrane Library, Global Health and LILACS bibliographic databases. The literature after April 2015 was also monitored using a daily email alert tool provided by Google Scholar (searching for “water borne disease” and “water related disease”) to identify potential papers adopting newly-developed methods not covered by the initial search.
We used search terms related to water-associated diseases (e.g. “water transmission” OR “contaminated fresh water” OR “unsafe water supply” etc.) and quantitative methodologies (e.g. “mathematical epidemiology” OR “simulation”) and weather and climate. The full list of search terms is in the Supporting Information, S2 Text. Papers were reviewed by two people (GL and GN). As the pool of returned papers was quite large, we decided to not use additional specific search terms for pathogens (e.g. `cholera’, `rotavirus’) or diagnosis/symptoms (e.g. 'diarrhoea', 'gastroenteritis'), as this would require a subjective list of potential pathogens and introduce unnecessary bias in the selection of the papers. We included articles that: i) were published in peer reviewed journals; ii) included an infectious disease in human beings; and iii) developed new methods and/or applied established methods to investigate the effects of weather and/or climate on infectious diseases (including papers for which weather and climate variables were among other equally important factors driving disease transmission).
The final set of papers was archived in EndNote (see Supporting Information, S1 Table). We identified specific questions related to the nature of the methods, their range, applicability and limitations (Table 2) that we wanted the Review to address. We then created a spreadsheet consisting of records (rows) corresponding to each paper in our final database, and columns to address the specific questions. Analysis was done in R open source analytic software [12].
Papers were clustered according to the methodology used. More precisely, for each paper we identified the list of technical keywords associated with the methods, including both general concepts (e.g. “time series analysis”) and sub-analysis terms (e.g. “partial autocorrelation function”); the full list of technical keywords is presented in the Supporting Information, S1 Table. Papers that share the same keyword are often connected. Consequently, analytical methods that are likely to be used together in the same papers tend to cluster. Analysis was done by using the “igraph” package in R [13].
Overall, 102 papers were included in the analysis (Fig 2). Analysis of the findings and synthesis of the challenges in formulating new methods are summarised below.
A range of diverse methods are used to study the effect of weather and climate on water-associated diseases. Most of these methods can be connected to two main groups: (i) process based models (PBM), and (ii) time series and spatial epidemiology. In general, PBM were employed when the bio-physical mechanism of transmission of the pathogen was known (as with Vibrio cholerae, the most studied pathogen). PBMs describe the progression of the infectious diseases by mimicking the bio-physical processes, typically, in terms of non-linear differential equations [14]. In contrast, statistically-oriented approaches (such as regression analysis over space or over time) tended to be used when the roles of specific environmental drivers were unclear and the methods needed to find potential correlates between environmental and/or socio-economic variables and patterns of infections.
The two clusters resemble the groups of explanatory and predictive models. Although we recognize the importance of the debate in the philosophy of science about these two groups of models [10], we avoided a formal grouping of the studies according to this classification. Applying the guidelines provided by Shmueli [10] in a rigorous and objective manner was particularly challenging in our contest. For example, many modelling studies (e.g. [78]) provide maps of the risk of a disease, i.e. they make predictions, by calculating the Basic Reproductive Number, which is a typical tool in explanatory compartmental models as built on “First Principles”. Should these models be classified as predictive or explanatory?
The use, or potential use, of the reviewed methodologies to investigate the effects of weather and/or climate is discussed in Table 3. Most of the studies focused on the short-term effects of weather as these time series of health data are more readily available.
Teasing apart the individual effects of multiple abiotic factors (e.g. weather, climate, environmental, demographic, and socio-economic) on the incidence of water-associated infectious diseases is the main challenge. Additional challenges include many other methodological problems arising from the limited understanding of the complex bio-physical mechanisms, concurrent effects of many correlated factors, poor data quality and reporting bias, and uncertainty in the knowledge of relevant parameters; these are all intrinsic limitations of the methods and the data available. Addressing these challenges would enable the formulation of a framework to understand the overall effects of weather, climate, and possibly other environmental and socio-economic factors on water-associated infectious disease.
This research has the common limitations of any systematic review. An important one is the possibility we have missed relevant studies, for example due to the failure of the search engine or studies written in languages other than English. This problem could be overcome, at least in part, by the application of snowballing procedures, i.e. recursively pursuing relevant references cited in the retrieved literature and then adding them to the search results [79]. This technique is less feasible when the initial pool of documents is as large as in the current case.
Nevertheless, considering the high number of studies included in this Review, we expect that the general patterns in our findings are robust. We recognize that there are methods developed and applied to other different classes of infectious diseases (e.g. vector-borne diseases) potentially relevant to our context which were not included in our analysis. The protocol of our review was chosen to ensure the objectivity and reproducibility of the results and the research question we identified.
The choice of a particular method in a study is driven by many factors, including the scientific background of the scientists involved (anecdotally, we noticed that most PBM are employed by scientists working in engineering or physics departments); the ease in implementing the methods (e.g. using freely available statistical packages); and probably the tendency to use already widely used methods, a phenomenon known as the Matthew Effect [80,81]. The choice of the methods ought to be driven by many factors, i.e. the scientific questions, the availability of data, the transmission pathways, and the bio-physical and socio-economic mechanism, as well as the state of the art of the methods. These should be critically assessed on a case-by-case basis, and not based on oversimplified, prescriptive guidelines. The findings of this Review can assist scientists in the critical selection and development of methods for quantifying the effects of weather and climate on water-associated diseases.
|
10.1371/journal.pntd.0005858 | Assessment of risk of dengue and yellow fever virus transmission in three major Kenyan cities based on Stegomyia indices | Dengue (DEN) and yellow fever (YF) are re-emerging in East Africa, with contributing drivers to this trend being unplanned urbanization and increasingly adaptable anthropophilic Aedes (Stegomyia) vectors. Entomological risk assessment of these diseases remains scarce for much of East Africa and Kenya even in the dengue fever-prone urban coastal areas. Focusing on major cities of Kenya, we compared DEN and YF risk in Kilifi County (DEN-outbreak-prone), and Kisumu and Nairobi Counties (no documented DEN outbreaks). We surveyed water-holding containers for mosquito immature (larvae/pupae) indoors and outdoors from selected houses during the long rains, short rains and dry seasons (100 houses/season) in each County from October 2014-June 2016. House index (HI), Breteau index (BI) and Container index (CI) estimates based on Aedes (Stegomyia) immature infestations were compared by city and season. Aedes aegypti and Aedes bromeliae were the main Stegomyia species with significantly more positive houses outdoors (212) than indoors (88) (n = 900) (χ2 = 60.52, P < 0.0001). Overall, Ae. aegypti estimates of HI (17.3 vs 11.3) and BI (81.6 vs 87.7) were higher in Kilifi and Kisumu, respectively, than in Nairobi (HI, 0.3; BI,13). However, CI was highest in Kisumu (33.1), followed by Kilifi (15.1) then Nairobi (5.1). Aedes bromeliae indices were highest in Kilifi, followed by Kisumu, then Nairobi with HI (4.3, 0.3, 0); BI (21.3, 7, 0.7) and CI (3.3, 3.3, 0.3), at the respective sites. HI and BI for both species were highest in the long rains, compared to the short rains and dry seasons. We found strong positive correlations between the BI and CI, and BI and HI for Ae. aegypti, with the most productive container types being jerricans, drums, used/discarded containers and tyres. On the basis of established vector index thresholds, our findings suggest low-to-medium risk levels for urban YF and high DEN risk for Kilifi and Kisumu, whereas for Nairobi YF risk was low while DEN risk levels were low-to-medium. The study provides a baseline for future vector studies needed to further characterise the observed differential risk patterns by vector potential evaluation. Identified productive containers should be made the focus of community-based targeted vector control programs.
| Despite the growing problem of dengue (DEN) and yellow fever (YF) evidenced from recent outbreaks in East Africa, risk assessment for their transmission and establishment through surveys of populations of the Aedes mosquito vectors, remain scarce. By estimating standard indices for the potential vectors, Aedes aegypti and Aedes bromeliae we partly could deduce the risk of transmission of these diseases in three major cities of Kenya, namely Kilifi (DEN-prone) and Kisumu and Nairobi (without any DEN outbreak reports). When compared to established threshold risk levels by WHO and PAHO, our findings suggest low-to-medium risk of urban YF, and high risk of DEN transmission for Kilifi and Kisumu but not Nairobi (low risk level for YF and low-to-medium risk for DEN). The observed seasonal risk patterns, higher Aedes infestation outdoors than indoors and productive container types (jerricans, drums, discarded containers and tyres), provide insights into the disease epidemiology and are valuable for targeted vector control, respectively.
| Dengue (DEN) and yellow fever (YF) are re-emerging diseases of public health importance caused by arboviral pathogens [1–4]. Both diseases share a common ecological niche including non-human primates as reservoir hosts and are vectored primarily by Aedes (Stegomyia) species [5]. Dengue fever is caused by one of the four serotypes of the dengue virus (DENV 1–4) with about 390 million infections reported worldwide each year, 16% of which are from Africa [6,7]. Additionally, an estimated 900 million people are living in YF endemic areas with about 90% of the global infections reported from Africa [8,9].
The rapid geographic spread of these diseases in recent times in Africa and especially in East Africa represents a worrying new trend with occurrence of major epidemics affecting urban human populations [10,11]. This is exemplified by recent DEN outbreaks in Somalia 2011, 2013 [12], Tanzania 2013, 2014 [4,13], Sudan 2010, 2015 [14,15] and various parts of Kenya 2011, 2013, 2015 [1,2]. An outbreak of YF was reported in Kenya in 1992–93 [16], in Sudan 2003, 2005, 2012 [17–19] and neighboring Uganda 2011, 2016 [20,21]. Despite the fact that the last YF outbreak in Kenya occurred over two decades ago, the country is still classified among countries with medium to high risk of YF transmission in Africa [22], and a number of YF cases have recently been imported from Angola where there was an ongoing outbreak [21]. There are currently no antiviral drugs available for either DEN or YF. However, there is a safe efficacious vaccine against YF, and a new, partially approved vaccine for DEN, for use only in geographical settings where epidemiological data indicate a high burden of the disease [23]. Unfortunately, the costs and availability of these vaccines have proved to be challenging for effective disease prevention. While the recent DEN and YF outbreaks in Africa have attracted renewed public health and research attention, effective monitoring and risk assessment for their occurrence remains limited.
Dengue virus (DENV) is known to be transmitted primarily by Aedes furcifer in Africa and Ae. aegypti aegypti in Asia and the Americas [5]. Aedes aegypti aegypti is highly anthropophilic and its larvae develop mostly in artificial containers in and around human habitations, compared to the more sylvatic Ae. aegypti formosus subspecies which develop mostly in tree holes hence linking the emergence of DEN in tropical urban areas to Ae. aegypti aegypti [24,25]. Although the role of Ae. aegypti in the transmission of yellow fever virus (YFV) in East Africa is poorly understood, it plays an important role in YFV transmission in West Africa, driving human-to-human transmission and resulting in dreaded urban outbreaks [26,27]. Yellow fever outbreaks in East and Central Africa have so far been associated with Ae. bromeliae, a member of the Ae. simpsoni species complex [28–30]. Aedes bromeliae is a peri-domestic mosquito species capable of biting humans and monkeys, thereby driving small scale outbreaks in rural populations, with potential to move virus across species from primates to humans [5]. Other species such as Ae. africanus and Ae. luteocephalus, feed on forest monkeys and sustain the sylvatic cycle of YF [31]. Although Ae. albopictus a secondary DEN vector is not known to be present in Kenya, Ae. aegypti and Ae. bromeliae are present in the major cities [32], hence the need to assess the risk of arboviral disease emergence associated with these vectors.
Risk assessment through surveillance of abundance and distribution of Aedes mosquitoes, which are key players in transmission of the pathogens that cause these diseases is critical. This largely relies on estimation of traditional Stegomyia indices (House Index-HI, Container Index-CI and Breteau Index-BI) of immature mosquito populations in households [33–36]. Estimation of such indices may be of operational value and can facilitate the determination of local vector densities and measurement of the potential impact of container-specific vector control interventions such as systematically eliminating or treating larval habitats with chemicals [37]. Surprisingly, estimations of these indices as a means of assessing risk of DEN and YF in Kenya are scarce and/or exclusive to Ae. aegypti in outbreak situations [31]. Moreover, similar investigations on other Stegomyia species such as Ae. bromeliae are completely lacking, in spite of its’ potential role in YFV transmission in Africa [5].
Unplanned urbanization remains an important risk factor that has contributed to the resurgence of these diseases by providing abundant larval habitats from water-retaining waste products and storage facilities in the presence of susceptible human populations [38–40]. A better epidemiologic understanding of entomological thresholds relating to risk can help to prevent a severe outbreak in urban settings. Potential exists for emergence of these diseases, especially YF from proximal sylvan areas, and subsequent introduction into urban areas where dense susceptible populations and competent domestic vectors abound [41], as demonstrated by the recent YF outbreak in Angola and the Democratic Republic of Congo [11,21].
To assess the potential risk of urban transmission of these diseases we estimated HI, CI and BI in the three major cities of Kenya, namely Kilifi (DEN-prone) and Kisumu and Nairobi (DEN-free) in the light of known differential outbreak reports of DEN. These cities, which serve as major tourism, trade and shipping hubs for much of eastern Africa, have high levels of human population movement and potential for heightened risk of importation of viruses. We also investigated possible seasonal patterns and associated risk indices for Ae. aegypti and Ae. bromeliae, as the two vector species implicated in disease transmission in East Africa, inclusive of Kenya. We further characterized the most productive container types based on the number of immature mosquitoes surveyed, reared to adults, and identified; information, which can be used to guide targeted source reduction/control operations.
The study was carried out on the outskirts of the major cities of Kenya; Nairobi and Kisumu (with no history of DEN outbreak) and Mombasa (DEN endemic and outbreak prone). While the phenomenon of DEN expansion is associated with urban human settlement, incidence of the disease in rural areas is also on the rise and is sometimes even higher than in urban and semi-urban areas/communities [40,42,43]. Therefore, our study targeted the cities, where we specifically selected sites in peri-urban suburbs around the main cities, Githogoro (Nairobi County), Kisumu (Kisumu County) and Rabai (suburb within Kilifi County, at the outskirts of Mombasa city), mainly for logistical reasons, including ease of access to homesteads and households.
Githogoro is located about 13.1 km from the Central Business District (CBD) on the outskirts of Nairobi (01°17'S 36°48'E), the largest city and capital of Kenya (Fig 1). Nairobi has a total surface area of 696 km2, a population of 3.1 million people [44], and is situated at an altitude of 1,661 m above sea level (asl). Githogoro is an urban informal settlement with most of the houses made of iron sheeting and consisting of a single room. A few houses have more than one room and some yard space.
In Kisumu (00°03′S 34°45′E), the study sites included Nyalenda B, Kanyakwar and Kajulu located on the outskirts of Kisumu CBD at a distance of approximately 6.5 km, 5.8 km and 27.8 km, respectively. Kisumu is the third largest city in Kenya and the second most important city after Kampala in the greater Lake Victoria basin (Fig 1). It has a human population of >400,000 [44] and is situated at an altitude of 1,131 m asl. The houses in this area mostly have cemented walls and roofs made of iron sheeting. Water storage in containers is a common practice by the communities.
The study sites included Bengo, Changombe, Kibarani, and Mbarakani, in Rabai, which is located on the outskirts of Mombasa, though administratively it belongs to Kilifi County (Fig 1). Rabai is situated about 24.5km to the north-west of Mombasa CBD, the second largest city in Kenya, which is situated on an island (4°03'S 39°40'E). Mombasa has a total surface area of 294.7 km2, a population of 1.2 million people [44] and is situated at an attitude of 50 m asl. The houses in Rabai have walls that are either cemented, made of stones, or mud. The roofing system consists of iron sheeting or grass thatch. Water storage in containers is an equally common practice in these communities.
All three-study cities generally experience two rainy seasons, the long rains season (April-June) and the short rains season (October-December), interspersed by two dry seasons (January-March and July-September).
We conducted a cross-sectional survey of water holding containers situated both indoors and outdoors for presence of immature mosquito stages (larvae at all instars and pupae). The inspections and entomological surveys were conducted by a team of four trained personnel in houses that were selected at random for the initial survey. An interval of one house was applied during the first sampling and unique numbers assigned to each house for ease of identification in subsequent surveys during the next season. In cases where a house could not be sampled in subsequent surveys, either due to absence of the inhabitants or the owners declining entry, it was substituted for the next closest available house. Each survey was conducted over five consecutive days and 100 houses from the selected sites were targeted, within each of the three main urban areas (Nairobi, Kilifi, Kisumu). Repeat sampling of the same 100 houses / city was conducted for the dry season (July-September 2015 in Nairobi; January-March 2016 in Kilifi and Kisumu) and for the long rains (April-June 2015 in Kilifi, and Kisumu; April-June 2016 in Nairobi) and short rains (October-December 2014 in Kilifi, and Kisumu, October-December 2015 in Nairobi) seasons. As such, there was a total of three sampling occasions (with 100 houses being sampled per study city and per season, corresponding to 900 sampling points), for the survey conducted from October 2014 to June 2016. Sampling in Nairobi was limited to Githogoro, whereas in Kilifi (Rabai) and Kisumu, operational surveys were conducted to reflect the proportionate size of each site in terms of the number of houses present. These sites were Bengo, Kibarani, Changombe and Mbarakani in Kilifi and Kajulu, Kanyakwar and Nyalenda B in Kisumu.
The survey of immature stages of Aedes Stegomyia mosquito species targeted artificial water-holding containers (indoors and outdoors) of any size and natural breeding sites (tree holes, banana axils, flower axils and colocasia) in peri-domestic areas of selected houses. Sampling was carried out using standardized sampling tools based on the type of water holding container encountered [45]. For small discarded containers (mostly found around the house, holding water which is not for household use), the water was emptied into a white tray and a plastic Pasteur pipette was used to collect the immatures. Jerrican (small plastic containers, 5-40L holding water for household use) surveys entailed pouring the water through a sieve into a bowl with a good contrast and collecting all immatures from the sieve with an aspirator. In large containers such as metal and plastic drums (50-210L containers used to store water for household use), the immatures were collected using ladles and aspirators when less than 20 were present or by emptying the water through a sieve when there were more than 20. Ladles, aspirators and pipettes were used to collect immatures from tyres as well as from tree holes and leaf axils. Flashlights were used where necessary. We captured information on each container sampled including: indoor or outdoor, natural or artificial, and the capacity of the container (>70L, 20L-70L, <20L). Immatures collected from containers were placed in whirlpaks (Nasco, FortAtkinson, WI) labeled with the pertinent information and transported to the field laboratory.
Larval samples were placed in individual rearing trays for each container types. All pupae collected for the separate container types were transferred to individual adult cages. Larvae were fed fish food (Tetramin) daily and the trays were inspected twice a day and pupae transferred to adult cages as well. This was done until all collected larvae/pupae had emerged to adults. During rearing, male and female Aedes mosquitoes were left together in a cage (small plastic buckets covered with fine netting materials and secured with rubber bands) and supplied with a 6% glucose solution on cotton wool. At the end of each sampling session, all adults were knocked down using triethylamine, placed in cryotubes and preserved in liquid nitrogen for transportation to the laboratory at the International Centre of Insect Physiology and Ecology in Nairobi. In the laboratory the resulting adult mosquitoes were morphologically identified using available taxonomic keys [46–48] and counted and data on the species and number collected from the different container types were captured in Excel.
A container was considered positive when at least one Ae. aegypti or Ae. bromeliae larva or pupa was found [45], and a house positive if at least one container type indoor was found infested with Ae. aegypti and/or Ae. bromeliae larvae. We estimated the classical Stegomyia indices: HI (percentage of houses infested with Ae. aegypti or bromeliae immatures), CI (percentage of water-holding containers infested with Ae. aegypti or bromeliae immatures), and BI [number of Ae. aegypti or bromeliae positive containers (indoor and outdoor) per 100 houses inspected].
We tested for significance of area/site and for seasonal effects in the patterns of observed indices (BI, HI, CI) using analysis of variance (ANOVA) followed by mean separation using the Tukey test (P = 0.05). Data for the different seasons were also pooled in each area to estimate the overall Stegomyia indices, and similarly compared for the different seasons and areas. Correlation analysis was performed to test for significant correlations between the indices- BI, HI, and CI.
The density of Ae. aegypti (total number of mosquitoes collected per total number of positive containers) indoors and outdoors was established and the difference compared within each area using a t-test.
The inspected containers were further categorized into 9 types based on similarity in certain features (e.g. size, natural or artificial, etc). The productivity of each of these container types was calculated per season and area as the percentage of the total number of immatures (larvae or pupae) determined by the adults reared from the container types (Productivity = 100 x (total number of immatures) / number of positive containers). We also applied ANOVA to test for significant differences in the proportion of positive containers (positivity) and compared the productivity among the container types after angular transformation. Container positivity for the different seasons was compared within an area using the Chi-Square test.
All analyses were carried out in R version 3.3.1 [49] at α = 0.05 level of significance. Based on estimated indices we classified the areas/sites in terms of epidemic risk levels for YF or DEN as low, medium or high with reference to established epidemic thresholds [50,51]. HI values for Ae. aegypti and Ae. bromeliae were used to estimate risk of YFV transmission for the individual species with values of HI > 35%, BI > 50 and CI > 20% considered as high risk of urban transmission of YFV; HI < 4% BI < 5 and CI < 3% considered as unlikely or low risk of the disease transmission [50]. Similarly, the Pan American Health Organization (PAHO) has established threshold levels for dengue transmission based on HI for Ae. aegypti with low being an HI < 0.1%, medium an HI 0.1%–5% and high an HI > 5% [51].
We sought permission from household heads through oral informed consent to allow water-holding containers in their residences to be surveyed. Household survey of mosquitoes was carried out with ethical approval from Kenya Medical Research Institute Scientific and Ethics Review Unit (KEMRI-SERU) (Project Number SERU 2787).
A total of 11,695 mosquitoes were reared from the larvae and pupae collected from water holding containers, both indoors and outdoors, from all sites and cities. These included Ae. aegypti (63.5%), Ae. bromeliae (2.9%), Eretmapodite chrysogaster (1.9%) and Culex spp. (31.53%). Aedes metallicus, other Aedes species (Ae. tricholabis, Ae. durbanensis) together with Aedeomyia furfurea, Uranotaenia spp, Anopheles gambiae s.l and Toxorhynchites spp. each comprised 0.1% or less of the total collection (Table 1). Focusing on our species of interest, a total of 7,424 Ae. aegypti were collected from all sites comprising 3,342 (45.0%) from Kilifi, 3,733 (50.3%) from Kisumu and 349 (4.7%) from Nairobi with an overall higher proportion (76%) being collected outdoors than indoors (24%). The Ae. aegypti densities recorded indoors and outdoors were not significantly different in the DEN-outbreak prone county of Kilifi (n = 17.5 indoors, n = 15.4 outdoors, P = 0.7). In contrast, counties of Kisumu (n = 8.3 indoors, n = 16.8 outdoors, P = 0.036) and Nairobi (n = 0.7 indoors, n = 14.7 outdoors, P = 0.048) (with no documented records of DEN outbreaks) had significantly higher densities of Ae. aegypti outdoors compared to indoors (Fig 2).
Similarly, a total of 335 Ae. bromeliae were collected mainly outdoors (92%). The highest proportion was sampled in Kilifi (63%, n = 211), followed by Kisumu (32.8%, n = 110) and then Nairobi (4.2%, n = 14) (Table 1).
The rainy seasons recorded the highest proportions of Ae. aegypti in all three areas evaluated in this study. In Kilifi, long rains constituted 1,648 (49.3%) of the total Ae. aegypti collected, followed by short rains 1,172 (35.1%) with the lowest 522 (15.6%) observed during the dry season. An analogous pattern was found in Kisumu and Nairobi. In Kisumu, the long rains, short rains and dry season each accounted for 1,470 (39.4%), 1,441 (38.6%) and 822 (22.0%) of the total Ae. aegypti sampled. Surprisingly, collection of Ae. aegypti in Nairobi was highest during the short rains 152 (43.6%), followed by the long rains 143 (41%) and then the dry season at 54 (15.4%). However, the seasonal difference observed between long and short rains in Nairobi was not statistically significant (χ2 = 0.38, P = 0.5).
Relative to Ae. aegypti, very low numbers of Ae. bromeliae were encountered from containers during our study. However, a seasonal pattern of abundance, with the highest proportion collected during one of the rainy seasons, was observed at all the areas. In Kilifi, Ae. bromeliae collected during the long rains, short rains and dry seasons made up 52.9%, 45.1% and 1.9%, respectively, of the total collection. However, in Kisumu the highest proportion was recorded in the short rains (70.9%), while the long rains and dry seasons recorded 10% and 19.1% respectively of the total collection. In Nairobi, there was no record of Ae. bromeliae in the short rains and dry seasons, and this mosquito species was only recorded in the long rains. In terms of occurrence in container types, Ae. aegypti was mostly encountered in artificial containers such as jerricans, drums, tyres and other discarded containers at all the sites. However, to a lesser extent Ae. aegypti was found in natural container types such as tree holes and leaf axils in Kilifi and Kisumu (Table 2). Natural breeding sites like leaf axils were the most productive site for Ae. bromeliae at all the sites (Table 3). In fact, Ae. bromeliae was not found breeding in artificial containers in Nairobi, although to a minor extent it bred in artificial containers such as Jerricans and other discarded containers (Table 3) in Kilifi and Kisumu, mostly co-habiting with Ae. aegypti.
There was no significant difference in Ae. aegypti immature productivity by season or area. However, the contribution of container types to productivity of this species varied significantly (Df = 9, F = 6.41 P < 0.0001). Significant differences were mostly observed between drums and animal drinking containers (P = 0.0008), drums and basins (P = 0.01), drums and natural breeding sites (P = 0.002), jerricans and animal drinking containers (P = 0.01), jerricans and natural breeding sites (P = 0.02), tyres and animal drinking containers (P = 0.013) and between tyres and natural breeding sites (P = 0.022). Overall, in Kilifi, the most productive container types were jerricans (36.3%) in the long rains, discarded containers (34.7%) in the short rains, and drums (49.0%) in the dry season (Table 4). Similarly in Kisumu, the most productive container types were the jerricans (29.5%) in the long rains, drums (24.5%) and discarded containers (24.1%) in the short rains and drums in the dry (38.1%) season (Table 4). In Nairobi, drums (32.9%) were the most productive container types in the long rains, tyres (84.9%) in the short rains, and tanks (63.0%) in the dry season (Table 4).
The most productive containers for Ae. bromeliae in Kilifi and Kisumu were discarded containers and natural breeding sites, while in Nairobi natural breeding sites were the most productive breeding sites (Table 5).
Based on the number of each container types surveyed and the number positive, we found significant differences in container positivity between the areas (Df = 2, F = 9.6, P = 0.0002) and seasons (Df = 2, F = 84.26, P = 0.018). Significant differences existed in the container type positivity between Kilifi and Kisumu [95% CI, (0.329, 26.392), P = 0.043], Kisumu and Nairobi [95% CI, (-37.214, -11.152), P < 0.0001], but not between Kilifi and Nairobi. Generally, animal drinking containers and tyres were the most positive containers in Kilifi, tanks and discarded containers in Kisumu, and tyres and tanks in Nairobi. Similarly, container positivity was significantly different between the long rains and dry seasons [95% CI, (2.393, 28.456), P = 0.016], long and short rains [95% CI, (-27.122, -1.059), P = 0.03], but not between the short rains and dry season. The proportion of positive containers was significantly different for all three seasons in Kilifi (χ2 = 119.0, P < 0.0001) and Nairobi (χ2 = 31.7, P < 0.0001) but not in Kisumu (χ2 = 4.45, P < 0.1078). Tyres were the most positive containers both in the long and short rains in Kilifi while drums were the most positive containers in the dry season. In Kisumu, tanks constituted the most positive containers in the long rains, basins in the short rains and drums in the dry season. In Nairobi, discarded containers ranked as the highest positive containers in the long rains, tyres in the short rains and tanks in the dry season.
The overall Ae. aegypti CI was higher during the long rains followed by dry season and then short rains in Kilifi. In Kisumu, CI was higher in the dry season, followed by the long rains and then short rains, while in Nairobi, CI was higher in the long rains followed by short rains and then dry season (Fig 3A). The seasonal differences observed in all three cities were not significant (P = 0.14). However, the observed CI values were significantly different among the different cities (Df = 2, F = 16.69, P = 0.012), with differences recorded between Kilifi and Kisumu [95% CI, (0.483, 35.450), P = 0.046], Kisumu and Nairobi [95% CI, (-45.45, -10.48), P = 0.01], but not between Kilifi and Nairobi. CI was equally significantly different even at smaller scale among the sites (Df = 5, F = 3.133, P = 0.037). Overall, CI was highest in Kanyarkwar (Kisumu) and lowest in Kibarani (Kilifi).
The overall Ae. aegypti HI was highest in the long rains (24%, 15% and 0%), compared to the short rains (20%, 12% and 0%) and dry season (8%, 7% and 1%) respectively in Kilifi, Kisumu, and Nairobi (Fig 3B). Our analysis showed that overall HI values varied significantly in the different cities (Df = 2, F = 11.24, P = 0.023) with among area differences recorded between Kilifi and Nairobi [95% CI, (-29.96, -4.04), P = 0.02], but not between Kilifi and Kisumu or Kisumu and Nairobi. Also, the overall HI was highest in Kanyarkwar (Kisumu) and lowest in Githogoro (Nairobi).
Overall BI for Ae. aegypti varied significantly across the seasons (P = 0.044), with highest values observed in the long rains (141, 134 and 28), compared to the short rains (82, 83 and 7) and dry season (22, 46 and 7) in Kilifi, Kisumu and Nairobi, respectively (Fig 3C). Also, significant variation in the overall BI values was evident between areas (BI: Df = 2, F = 8.68, P = 0.035) and seasons (Df = 2, F = 7.52, P = 0.044). Among-area differences were observed between Kisumu and Nairobi [95% CI, (-145.66, -3.68), P = 0.043], but not between Kilifi and Kisumu or Kilifi and Nairobi. Likewise significant seasonal differences in BI values occurred between the long rains and dry seasons [95% CI, (6.01, 147.99), P = 0.0386], but not between the long and short rains, or the short rains and dry seasons in all three areas. Similarly, the overall BI was highest in Kanyarkwar (Kisumu) and lowest in Githogoro (Nairobi).
Based on HI values estimated for Ae. aegypti in reference to threshold levels for DEN transmission (low HI < 0.1%, medium HI 0.1%–5% and high HI > 5%) established by PAHO [51], both Kilifi and Kisumu were classified as being at high-risk for DEN transmission in all three seasons, while Nairobi was classified as being at low risk in both the long and short rains and at medium risk in the dry season (Table 6). Even small-scale differences in DEN risk across sites among the major areas Kilifi and Kisumu were evident, highest in Kanyakwar (Kisumu) and Mbarakani (Kilifi) (Table 6).
Similarly, with reference to the WHO threshold levels for urban YFV transmission (low HI < 4%, Medium 4%-35% and high HI > 35%), our risk level values for Ae. aegypti, show that Kilifi and Kisumu could be classified as being at medium-risk of an urban YF epidemic in all three seasons based on estimated HI values, and Nairobi at low risk in all three seasons (Table 7).
We found no significant difference in overall index values (CI, HI and BI) for Ae. bromeliae (Fig 3D, 3E and 3F), among the three areas in the different seasons (P > 0.05). However, based on the HI estimated for this species, compared to the established threshold levels for urban YFV transmission [50] and assuming that Ae. bromeliae could transmit YFV, only Kilifi could be classified as being at medium risk during the long rains but at low risk in the short rains and dry seasons. Both Kisumu and Nairobi can be classified as being at low risk levels of transmission in all three seasons (Table 8).
Equally strong positive correlations were recorded between the BI and HI (R2 = 0.887, P = 0.001) as well as the BI and CI (R2 = 0.721, P = 0.028) (Table 9).
Aedes aegypti and Ae. bromeliae were the major Stegomyia species recorded at all sites/cities, justifying estimation of indices for the two species considering their potential roles in DENV and YFV transmission [26,27,29,30]. Our findings support the sympatric existence of both species in these growing urban ecologies in Kenya.
Although particular container types were more likely to be positive than others, it was noteworthy that these were not necessarily the most productive, suggesting that positivity did not always translate to productivity. Aedes aegypti in all three areas were mostly found breeding in jerricans, drums (which were particularly productive in all seasons), tyres, and discarded containers. This was equally observed in an earlier study in Mombasa city, during entomologic investigations of a recent DEN outbreak [2]. These containers could be targeted at the community level through awareness creation and public health education for the control of Ae. aegypti mosquitoes. In this way, the local inhabitants can help reduce Ae. aegypti larval sites by reducing these containers in and near their homes or by properly covering them to prevent gravid females from laying their eggs in them [37]. Observations from this study show that Ae. aegypti is also capable of developing in natural sites especially in the water holding axils of banana plants. Aedes aegypti breeding in banana and colocasia plants have also been reported by Philbert and Ijumba (2013) in a study on the preferred breeding habitats of Ae. aegypti in Tanzania [52]. This adaptation should be monitored as it will take away any gains made from targeting control of breeding in artificial water holding containers. Immature stages of Ae. bromeliae, a species which is known to preferentially breed in phytotelmata, the water-holding axils of plants [53], were also found developing in artificial containers indoors and outdoors in this study. Its ability to develop in artificial containers both indoors and outdoors has also been reported in another study in coastal Kenya [54]. Both Ae. aegypti and Ae. bromeliae were also found co-developing in several artificial and natural breeding sites. Utilization of artificial breeding sites may be an indication that Ae. bromeliae is increasingly adapting to the urban environment, bringing it closer to human hosts and increasing the risk of transmission of a range of the arboviruses that cause human disease, including YFV.
Risk values for both Ae. aegypti and Ae. bromeliae were different not only between areas and seasons, but we found finer scale differences between the sites, suggesting spatio-temporal variation with non-uniform risk even within the same general ecology. Although water storage in containers is a common practice in these cities during the rainy and dry seasons, DEN outbreaks that have occurred in Mombasa have mostly been associated with the long and short rains [2]. The estimated HI and BI for Ae. aegypti both showed the same seasonal pattern in all three areas. The strong correlations between the traditional Stegomyia indices observed in this study, clearly indicates that they are all important in determining risk of transmission. It will also be important to investigate how the Stegomyia indices correlate with the observed DEN cases, especially in the coastal site of Kilifi County.
Estimated risk values suggested that both Kilifi and Kisumu were at high risk of DEN transmission while Nairobi was at low risk. Based on our findings, risk of DEN in Kilifi is high especially during the long rains (April-June) and short rains (November- December). This correlates with reports of DEN outbreaks in coastal Kenya, with outbreak peaks during the long and short rains in the 2013/2014 outbreaks [1,2]. High indices were also recorded in Mombasa city during this outbreak [2], with HI values comparable to that reported for Kilifi and Kisumu in our study. High indices have also been recorded in neighboring countries of Ethiopia [55] and Tanzania [56], which are prone to DEN outbreaks. Low indices were recorded in Nairobi, and this may partially explain the absence of reports of epidemic DEN in this part of the country, in spite of people arriving with infection from endemic areas during outbreaks [57]. Surprisingly, this study recorded high DEN risk indices in Kisumu yet there has been no reported outbreak in the region. This finding suggests that the mere presence of high abundance of Ae. aegypti as observed in Kisumu, may not be sufficient in estimating the risk of DEN transmission and that other factors should be considered including susceptibility of the Ae. aegypti population to the DENV, as well as their feeding behavior. All of these can affect vectorial capacity as has been demonstrated for Ae. albopictus [58].
We also observed significantly higher numbers of Ae. aegypti immatures outdoors compared to indoors in Kisumu and Nairobi. There is reason to believe that immatures will eventually emerge to adults posing biting risk to humans both indoors and outdoors in Kilifi compared to the outdoor risk in Kisumu and Nairobi, thereby leading to an increased risk of exposure to DEN transmission. This differential proximity of Ae. aegypti to human dwelling/activity may be a contributing factor to the differential epidemiology and outbreak pattern of DEN in the different cities. Earlier studies on the ecology of Ae. aegypti in the Kenyan coast suggested that the larvae of the domestic form Ae. aegypti aegypti develops indoors as opposed to the sylvatic form Ae. aegypti formosus which develops outdoors especially in forest tree holes and a polymorphic population which develops either indoors or outdoors in tree holes, steps cut into coconut palm trees, discarded tires, or tins [24]. Based on our observation, it is likely that the vector population in Kisumu and Nairobi is predominantly Ae. aegypti formosus, which has been described in other studies as a less efficient DEN vector when compared to Ae. aegypti aegypti [59,60]. A study to correlate the indoor vs outdoor larval habitats to possible genetic diversity among the species and susceptibility to DEN viruses is warranted.
Aside from the aforementioned biological factors which can impact occurrence of DEN outbreaks, temperature is by far the most important climatic variable that can modulate this pattern [61] and should also be considered. Generally, the different study areas have different average monthly temperatures, 22°C to 28°C in Nairobi, 28°C to 30°C in Kisumu and 27°C to 31°C in the coastal area of Kenya where DEN is endemic. We are not sure how well the observed differences in the risk indices relate to the prevailing environmental temperature among the different areas. Higher temperatures have been shown to increase the ability of Ae. aegypti to transmit DENV by reducing the extrinsic incubation period [62–64]. However, it is important to note that the diurnal temperature fluctuations may be more important in modulating the transmission dynamics.
This study only inferred risk from infestation patterns of Ae. aegypti. How these risks relate to actual prevalence in the human population is deserving of further consideration. There is evidence to suggest that some silent DEN transmission goes unreported in Kisumu, as a serological survey carried out by Blaylock et al. (2011) in this part of the country reported DEN seroprevalence levels of 1.1%. This value is similar to that reported by Morrill et al. (1991) for DEN in the coastal area of Kenya during non-epidemic periods [65]. Dengue is known to manifest clinically like malaria and diagnostic tools for DEN detection are unavailable in most health centers in the East African region, including Kenya [57]. It is therefore very important to confirm undiagnosed malaria cases, as it is possible some of these cases may actually be DEN.
Generally, the risk of an urban YF epidemic occurring in Kenya based on vector abundance data observed in this study was classified as low to medium, with the risk due to Ae. aegypti being higher as compared to Ae. bromeliae. However, the role of Ae. aegypti in the transmission of YFV in East Africa has not been fully evaluated and in the documented outbreak that occurred in Kenya in 1992/93, it was observed that this was driven by sylvatic vectors mainly Ae. africanus and Ae. keniensis and that Ae. aegypti was not at all associated with the outbreak [31]. Aedes bromeliae has also been described as a YFV vector in this region, as it was the principal vector in the largest YF outbreak that occurred in Omo River in Ethiopia [29], as well as in outbreaks in Uganda [30]. Aedes simpsoni is a complex of at least three sister species of which only Ae. bromeliae has been implicated as a YFV vector [66]. To understand better the risk due to this species, it will be important to differentiate the sub-species occurring in these urban areas in parallel with vector competence status, which was outside the scope of this study.
In Kilifi and Kisumu the high abundance of Ae. aegypti especially in the rainy season is considered sufficient to allow YFV transmission in association with other YFV vectors species such as Ae. bromeliae, Aedes metallicus and Er. chrysogaster found at some of the sites. However, their ability to act as efficient YFV vectors in urban areas in Kenya needs to be evaluated as data on their vectorial capacity is completely lacking. It is important to note that high numbers of Ae. bromeliae were recorded in our study area in Kilifi, and that clarification of the role of this species in the transmission of endemic arboviruses, such as DENV and chikungunya virus is needed, as it may be acting as a potential secondary vector.
In conclusion, Ae. aegypti remains the only known DEN vector in Kenya with sufficient abundance in the major cities to sustain transmission. It is highly abundant and the risk values are indicative of high risk of DEN transmission in Kilifi and Kisumu. The key containers that are utilized by this species for oviposition are water storage containers that can be effectively targeted to reduce vector numbers and, consequently, the risk of virus transmission through community mobilization and public health education. The oviposition site preference, indoor vs outdoor containers, between the study areas is suggestive of behavioral and/or genetic variation occurring in the different vector populations, calling for further studies. Overall, our findings provide a baseline for future studies to understand further the observed differential risk patterns especially with respect to the vectorial capacity of the different populations of Ae. aegypti and Ae. bromeliae for DENV and YFV transmission.
|
10.1371/journal.pgen.1003723 | Deep Resequencing of GWAS Loci Identifies Rare Variants in CARD9, IL23R and RNF186 That Are Associated with Ulcerative Colitis | Genome-wide association studies and follow-up meta-analyses in Crohn's disease (CD) and ulcerative colitis (UC) have recently identified 163 disease-associated loci that meet genome-wide significance for these two inflammatory bowel diseases (IBD). These discoveries have already had a tremendous impact on our understanding of the genetic architecture of these diseases and have directed functional studies that have revealed some of the biological functions that are important to IBD (e.g. autophagy). Nonetheless, these loci can only explain a small proportion of disease variance (∼14% in CD and 7.5% in UC), suggesting that not only are additional loci to be found but that the known loci may contain high effect rare risk variants that have gone undetected by GWAS. To test this, we have used a targeted sequencing approach in 200 UC cases and 150 healthy controls (HC), all of French Canadian descent, to study 55 genes in regions associated with UC. We performed follow-up genotyping of 42 rare non-synonymous variants in independent case-control cohorts (totaling 14,435 UC cases and 20,204 HC). Our results confirmed significant association to rare non-synonymous coding variants in both IL23R and CARD9, previously identified from sequencing of CD loci, as well as identified a novel association in RNF186. With the exception of CARD9 (OR = 0.39), the rare non-synonymous variants identified were of moderate effect (OR = 1.49 for RNF186 and OR = 0.79 for IL23R). RNF186 encodes a protein with a RING domain having predicted E3 ubiquitin-protein ligase activity and two transmembrane domains. Importantly, the disease-coding variant is located in the ubiquitin ligase domain. Finally, our results suggest that rare variants in genes identified by genome-wide association in UC are unlikely to contribute significantly to the overall variance for the disease. Rather, these are expected to help focus functional studies of the corresponding disease loci.
| Genetic studies of common diseases have seen tremendous progress in the last half-decade primarily due to recent technologies that enable a systematic examination of genetic markers across the entire genome in large numbers of patients and healthy controls. The studies, while identifying genomic regions that influence a person's risk for developing disease, often do not pinpoint the actual gene or gene variants that account for this risk (called a causal gene/variant). A prime example of this can be seen with the 163 genetic risk factors that have recently been associated with the chronic inflammatory bowel diseases known as Crohn's disease and ulcerative colitis. For less than a handful of these 163 is the causative change in the genetic code known. The current study used an approach to directly look at the genetic code for a subset of these and identified a causative change in the genetic code for eight risk factors for ulcerative colitis. This finding is particularly important because it directs biological studies to understand the mechanisms that lead to this chronic life-long inflammatory disease.
| Inflammatory bowel diseases (IBDs) are classified as chronic relapsing inflammatory diseases of the gastrointestinal tract. The two major forms of IBDs are Crohn's disease (CD, OMIM 266600) and ulcerative colitis (UC, OMIM 191390). Both genetic and environment factors play a central role in the pathogenesis of the inflammatory response of IBDs [1].
Recent genome-wide association (GWA) studies and meta-analyses in IBD have shown great success, with the identification of 163 independent IBD risk loci. While some loci were shown to be specific to either CD or UC risk, most have been shown to impact on both diseases, supporting earlier claims that these diseases share genetic risk factors [2]. These recent studies have identified important disease pathways but the common SNPs identified, with generally modest effects, explain only 14% and 7.5% of disease variance for CD and UC, respectively [3].
Due to linkage disequilibrium in the genome and limitations of GWAS chip designs to date, genome-wide scans typically identify common variants that tag regions of variable sizes containing multiple candidate genes for disease susceptibility. Although there have been a few notable exceptions, most of the common associated SNPs do not clearly identify causal variants, and further studies are needed to highlight the causal gene in many associated regions [4]–[6]. Sequencing of exons within associated regions in order to identify rare variants with strong effect on disease has been proposed as a means to help identify the causal genes and to help explain a further portion of disease variance. We have recently performed a pooled next-generation sequencing study in Crohn's disease, and identified association to novel low-frequency and rare protein altering variants in NOD2, IL23R, and CARD9, as well as IL18RAP, CUL2, C1orf106, PTPN22 and MUC19 [7]. We opted to use a similar targeted pooled next-generation sequencing approach to study UC-associated regions from our recent meta-analysis of 3 independent genome-wide scans for UC [8]. Using this approach we identified putative causal variants significantly associated to UC in three of the 22 loci examined and identified variants of interest for an additional six loci.
We selected 200 ulcerative colitis cases and 150 healthy controls of French Canadian ancestry from among samples collected by the NIDDK IBD Genetics Consortium. Samples were pooled in batches of 50 cases or 50 controls and normalized in order for the DNA pool to reflect sample allele frequencies. We targeted 55 genes from 14 UC-associated regions, as well as 7 regions identified in CD showing nominal replication in our UC GWAS study and an additional candidate gene (ECM1) reported in recent literature [6], [8]–[10]. PCR amplification primers were successfully designed to capture a total of 508 amplicons for a total of 305 Kb or 70% of our original target sequences. Of these 508 PCR reactions, 472 (93%) successfully amplified in each of the 7 sample pools and we used these to construct libraries for high-throughput sequencing on an Illumina Genome Analyzer II. This sequencing yielded large amounts of high-quality data for each pool, that captured 99% of our amplified target regions (283 Kb total; 117 Kb exonic sequences) and achieved 1575× median coverage per pool (corresponding to 31.5× per sample).
We used the previously described variant calling method Syzygy, designed to accommodate pooled study designs, to identify rare and low-frequency single nucleotide variants in our pooled samples [7]. Syzygy detected 1590 high confidence variants in our target regions, including 309 coding region variants (189 missense, 114 synonymous, 2 nonsense and 4 essential splice junction variants) with 56% of these already reported in dbSNP version 132, a non-synonymous/synonymous ratio of 1.7 and a transition/transversion ratio of 2.38 (Table S1). These results are similar to those obtained from our recent re-sequencing study in CD, as well as those reported by the 1000 Genomes Project, and are indicative of a relatively high true-positive rate for our dataset. This was confirmed by genotyping the 350 discovery DNA samples for a random subset of 237 variants from the total of 1590 high quality variants (Table S2).
After removal of variants that did not validate, variants observed only once in our sequencing dataset (singletons) and variants from the MHC region, 84 non-synonymous coding variants (missense, non-sense and splicing variants), were used for subsequent analyses. Following removal of common variants (frequency >5%) and variants that did not design in our genotyping assays, we carried out follow-up genotyping for 42 of these variants. Genotyping was performed in 6 independent case-control cohorts totaling 7,292 UC cases and 8,018 HC (Table S3), and additional data was obtained for 7,143 UC cases and 12,186 HC from the International IBD Genetics Consortium (IIBDGC) Immunochip project for 14 of these variants [3].
Since our study focuses on infrequent and rare variants, we expect few non-reference alleles for these variants in each subcohort studied, which precludes the use of asymptotic statistics utilized in typical association studies of common variants. Also, given the low frequencies of the variants tested, population structure is likely to be a more substantial problem and thus requires a stratified analysis with strict population case-control matching. We used a previously described mega-analysis of rare variants (MARV) approach that provides a permutation-based estimate of significance, within each sub-cohort, and accommodates variable numbers of case-control samples in each independent population for single-marker analysis [7].
With a target set of 42 variants we can define a traditional corrected significance level of P = 0.0012 for our study. Three variants, located in the CARD9, IL23R and RNF186 genes, reach this significance threshold suggesting that these could possibly be the causal genes/variants within these two loci (Table 1). Specifically, our results show that the c.IVS11+1G>C CARD9 splice variant confers significant protection to UC (P = 1.47×10−11; OR = 0.39 [0.30–0.53]). We previously identified this splice variant in a sequencing project of CD loci and demonstrated that it leads to an alternatively spliced transcript that is missing exon 11 [7]. Our results also identify significant association to the valine to isoleucine substitution at position 362 (Val362Ile) in IL23R (P = 1.18×10−03; OR = 0.79 [0.68–0.91]) previously reported by a recent re-sequencing of positional candidates in Crohn's disease [7], [11]. The significantly associated rare variant that we identified in RNF186 (P = 8.69×10−4; OR = 1.49 [1.17–1.90]) encodes an alanine to threonine substitution at position 64 (Ala64Thr). RNF186 encodes a protein with a RING domain and two transmembrane domains. Importantly, the disease-coding variant is located in the RING domain, a domain with a predicted E3 ubiquitin-protein ligase activity (Fig. 1).
Independence of effect between rare variants in IL23R and CARD9 and the reported common association signals in these genes has previously been shown [7], [11]. For RNF186, the Ala64Thr variant is mostly found on the protective haplotype background from the previously identified common variant, indicating that the reported association is not likely due to partial LD with the common variant. In addition, reciprocal conditional logistic regression analysis, using a subset of samples where both variants were genotyped (3548 UC cases and 3607 healthy controls) shows that these are independent association signals (data not shown).
Given the challenge inherent in achieving corrected significance thresholds for rare variants, even with large cohorts, we expect that some of the other variants that we identified and found to have nominal significance (0.0012<P<0.05) are truly associated with UC. In fact with a target set of 42 variants included in follow-up genotyping, and supposing these are independent and under the null, we would expect <1 SNP to exceed P<0.01 (with a probability of less than 1% to observe 3 or more associations at this level) and ∼2 SNPs to exceed P<0.05 by chance alone (with a probability less than 0.0001 to observe 9 or more association at this level), whereas we observe 3 SNPs with P<0.01 and 9 SNPs with P<0.05, suggesting that there are additional true positives that have not met the more stringent threshold. Indeed, within the group of SNPs that we found to have nominal significance are two non-synonymous coding variants (Gly149Arg and Val362Ile) in IL23R that we and others have shown to be associated with protection from IBD (Table 1) [7], [11]. In addition to these previously-validated variants in IL23R, we have found variants that are nominally associated with UC in the genes encoding CEP72, LAMB1, CCR6, JAK2, and STAC2 (Table 1). Specifically, we identified two nominally associated rare variants in CEP72 (Lys314Arg and Asp316Asn) in perfect LD with each other that appear to protect from UC (Table 1). As we also sequenced the only other gene in this locus (TPPP), but did not find any associated variants in it, this suggests that CEP72 is potentially causal. Similarly, we sequenced both genes in the LAMB1-DLD locus on chromosome 7, with the nominally associated rare variant in LAMB1 (Ile154Thr) suggesting a role for this gene in risk to UC, especially as the associated allele is located in its DUF287 domain and is predicted to have a damaging effect [12]. All genes within the CCR6-FGFR1OP-RNASET2 locus were sequenced, with a single nominally-associated variant (Ala369Val) in CCR6, consistent with this gene's probable role in the migration and recruitment of dendritic and T cells during inflammatory and immunological responses [13]. Within the JAK2-INSL6-LHX3 locus, we only sequenced JAK2 given its key role in signaling from the IL12R/IL23R, a biological pathway proven to be associated with IBD, and identified a nominally associated variant (Arg1063His) within its catalytic domain. STAC2 is within a locus with 16 other genes including ORMDL3, which has been suggested to be the most likely causal gene based previous genetic and functional studies in IBD and asthma [8], [14]. Although we find a nominally associated variant in STAC2 (Lys302Arg) and none in ORMDL3, we have only sequenced 10 of the 17 genes within this locus (Table S4). Studies of each of these variants to determine their functional impact will be essential to prove causality.
Genome-wide association studies in IBD have been very successful in identifying genomic regions associated with CD, UC or both. Only infrequently have these GWA studies also directly identified the causal genes/variants, with NOD2, IL23R and ATG16L1 being the few known examples. A recent targeted (exons and exon-intron boundaries) sequencing approach of known CD loci resulted in the identification of potentially causal variants in eight of the 36 loci examined [7]. The primary objective of the current study therefore was to use the same approach to identify likely causal variants within genes that were located in genomic regions associated with UC. While there are over 100 UC loci that have been identified and validated to date, we examined 22 UC loci that were known at the time of the initiation of this project. Of these 22 loci, the current study identified potentially causal variation in three of the loci: two protective alleles in CARD9 and IL23R, and an allele increasing risk in RNF186.
The identification of a rare variant (Ala64Thr) in RNF186 that shows significant association to UC strongly suggests that this is the causal gene within this locus. Importantly, the disease-coding variant is located in the RING domain, a domain with a predicted E3 ubiquitin-protein ligase activity. Ubiquitin ligases have been shown to regulate key adaptors of proinflammatory pathways [15]–[17]. We previously reported that RNF186 expression was higher in human intestinal tissues than in immune tissues [8]. We showed by immunostaining that the RNF186 protein was expressed at the basal pole of epithelial cells and lamina propria within colonic tissues. Using GEO public microarray datasets, we pursued a systematic follow-up analysis of expression profiles of epithelial cells in response to bacterial products, PAMPs/pathogens. We found that RNF186 gene expression was significantly up-regulated in small intestine epithelium and induced by Shigella infection in mice (P = 4.21×10−8) (Figure 1, Panel A) [18], [19]. Both invasive (INV+) and non-invasive (INV−) strains of Shigella induced significant overexpression of RNF186 in intestinal tissues of 4-day- and 7-day-old mice infected for 2 or 4 hours. To further identify putative transcriptional regulators of RNF186 expression, we employed a text-mining and network-generating analysis of human protein-protein, protein-DNA, protein-RNA and protein-compound interactions. Specifically, from our analyses we hypothesize that RNF186 is transcriptionally regulated in a two-step process by the transcription factor Hepatocyte Nuclear Factor 4, alpha (HNF4A) (Figure 1, Panels B,C). Several studies have shown that HNF4A binds to the promoter region and up-regulates the expression of yet another transcription factor HNF1A [20]–[22]. Knockdown of HNF4A has been shown to down-regulate HNF1A gene expression [23], [24]. HNF1A, in turn, regulates RNF186 and this interaction has been confirmed by chromatin immunoprecipitation and chip-on-chip assay [25]–[27]. Our own analysis of transcriptional profiles of HNF4A-Null colons recovered from HNF4AloxP/loxPFoxa3Cre and HNF4AloxP/−Foxa3Cre mice uncovered a significant up-regulation of RNF186 transcript [28]. Expression profiling of human tissues also supports this hypothesis, as HNF4A and RNF186 are clearly co-expressed in the small intestine and the colon (Figure S1). This putative interaction is particularly relevant given that HNF4A has previously been shown to be associated, with genome-wide significance, with risk to developing UC [9]. Our analysis now indicates a direct genetic interaction between two IBD susceptibility genes namely, HNF4A and RNF186. While a singular loss-of-function mutation in HNF4A has already been shown to be associated with susceptibility to abnormal intestinal permeability, inflammation and oxidative stress, we speculate that a dual loss-of-function with additional mutation in RNF186 would further exacerbate one's susceptibility to develop chronic inflammation in the gut [29], [30].
In addition to the variants in IL23R, CARD9, and RNF186, we also identified variants of interest in an additional five loci (specifically within the CEP72, LAMB1, CCR6, JAK2, and STAC2 genes). While these latter six still require confirmation, we estimate that many will validate given that we observed an excess of nominally-associated variants. Examining the data from the current study along with the data derived from prior association and sequencing studies suggests that at a minimum, there currently is strong evidence of association to causal variation in IBD (i.e. missense, nonsense or splice junction variants) in the NOD2, ATG16L1, IL23R, MST1, CARD9, IL18RAP and RNF186 genes, and at least suggestive evidence for causal variation in the CUL2, C1orf106, PTPN22, MUC19, CEP72, LAMB1, CCR6, JAK2, and STAC2 genes (Current study and references [4], [5], [7], [11], [31]). While only a small fraction of the recently identified 163 IBD loci have been sequenced (36 CD, 22 UC for total of 42 independent loci) in IBD patients and controls, this would suggest that from ∼10% (15 of 163 total loci) to ∼35% (15 of 42 loci sequenced) of IBD loci have causal variation affecting the protein-coding or splice junctions. There are an additional 5 loci (ITLN1, GSDMB, YDGL, SLC22A4, and FCGR2A) for which there are non-synonymous coding or splice variants present in public databases (dbSNP, 1KG) that are correlated with the index SNP identified in the GWA studies that have yet been tested directly, thus potentially increasing the estimated number of IBD loci with causal variation within the coding and splice regions [3], [32].
Furthermore, it should be noted that with the exception of a small number of variants with significant effect (e.g. R702W, G908R, fs107insC in NOD2; R381Q in IL23R; IVS11+1G>C in CARD9; V527L in IL18RAP – all of which had 0.5>OR>2) most of the rare variants identified by targeted sequencing of loci from GWAS regions have relatively modest effect sizes that are comparable to those observed for the common variants identified by GWA studies. Consequently, very large sample sizes are required to detect statistically significant association. In the current study, for the majority (93%) of variants with an observed minor allele frequency greater than 0.3%, we had more than 80% power to detect significant association if the OR is 2 or greater with the number of samples typed (up to ∼14,000 cases and ∼20,000 controls) (see Table S5). Moreover, should this observation not be limited to risk loci identified by GWA studies, this has implications with respect to future efforts for discovering risk loci. Specifically, if the occurrence of rare variants with large effects sizes is relatively infrequent, then this may favor the current paradigm of locus discovery by GWA followed by targeted sequencing rather than whole-exome or whole-genome sequencing for locus discovery as this would require even larger sample sizes. Alternatively, given the ever- growing size of public databases of common and rare variants, targeted genotyping of known variants within risk loci identified by GWA may prove to be an efficient approach. For example, all but two of the 22 candidate causal variants identified in the current study or that of Rivas and colleagues are now found in the Exome Sequencing Project database.
Regardless of the study design, these results suggest that a significant proportion of IBD loci contain causal variants within exons or exon-intron boundaries. While these rare/infrequent variants may not account for what has been termed “the missing heritability” of common traits, discovering these variants certainly can provide focus for follow-up functional studies. For example, the current sequencing and follow-up genotyping of the chromosome 1p36 locus, which was first identified in a GWA study of UC, identified significant association to the Ala64Thr variant within RNF186. While further studies will be required, the initial bioinformatics and experimental studies described above suggest that this ring finger protein with an ubiquitin-ligase domain may have an important role in the response to microbes/microbial products. Going forward, systematic evaluation of genes within risk loci via expression-driven functional studies in cellular models (i.e. knock-down or over expression) with sensitive high throughput/high content readouts may very well be a complementary approach given that at least a third of IBD risk loci appear to act via gene expression [3].
We selected 200 ulcerative colitis patients and 150 healthy control of French-Canadian descent from the NIDDK IBD Genetics Consortium repository samples. The NIDDK IBDGC samples were collected under rigorous clinical phenotyping and control matching for the purpose of genetic studies [33]. Genomic DNA concentrations were measured by Quant-iT PicoGreen dsDNA reagent (Invitrogen) and detected on the Biotek Synergy 2 plate reader. All DNAs were normalized with at least two round of dilution and quantification down to a concentration of 10 ng/µl as described previously [7]. Equimolar amounts of samples were pooled together in batches of 50 cases and 50 controls for a total of 7 pooled groups.
Target exonic sequences were selected based on the coding exons of 55 genes in 14 UC-associated regions and 7 regions identified in CD with nominal replication in our recent UC GWAS study, as well ECM1 identified from recent candidate-gene study in UC [6], [8]–[10], [34]. Specifically, amplicons were designed from genome build Hg18 using a web-base automated pipeline (Optimus primer: Website (http://op.pgx.ca)) that uses the Primer 3 design software and user defined parameters [35]. Design parameters included amplicon sizes between 400 and 600 base pairs, as well as the inclusion of Not1 tails for subsequent concatenation and shearing steps in library construction. PCR amplification reactions contained 40 ng of pooled genomic DNA, 1× HotStar buffer, 0.8 mM dNTPs, 2 mM MgCl2, 0.4 units of HotStar Enzyme (Qiagen), and 0.25 µM forward and reverse primers in a 10-µl reaction volume. PCR cycling parameters were as follows: one cycle of 95°C for 5 min; 30 or 35 cycles of 94°C for 30 s, 60°C for 30 s, and 72°C for 1 min; followed by one cycle of 72°C for 5 min. Each DNA pools were amplified for 508 PCR reactions; amplification products were then dosed by Quant-iT PicoGreen dsDNA reagent (Invitrogen) quantification and amplification specificity was validated by agarose gel electrophoresis. In total, 472 PCR amplicons (93% amplification success rate, capturing 283 Kb including 117 Kb of target exonic sequences) (Table S6) for each DNA pool were combined in equimolar amounts to obtain equal representation of all target in library construction.
The combined PCR products from each pooled DNA group were concatenated using the NotI adapters and sheared into fragments as previously described [36]. Libraries were constructed according to Illumina single-end library protocol, with 150–200 bp gel size selection and PCR enrichment using 10 cycles of PCR, and then single-end sequenced with 36 cycles on an Illumina Genome Analyzer II. Each sample pool was sequenced using a single lane of Illumina GAII analyzer flowcell; 36-base pair reads were aligned to the genome using MAQ algorithm [37] and base qualities were recalibrated using GATK (Genome Analysis ToolKit) [38]. Finally, variant discovery was performed using the previously described Syzygy software, designed to analyze sequencing data from pooled DNA sequencing [7].
We randomly selected 237 high quality variants for validation in our 350 discovery DNAs samples using Sequenom MassARRAY iPlex200 chemistry. Genotyping assay designs were obtained from the Assay Designer v.3.1 software, and genotyping oligonucleotides were synthesized at Integrated DNA Technologies. The correlation coefficient between observed minor allele frequencies and frequencies estimated from Syzygy for validated variants was calculated in order to evaluate the overall quality of our dataset (Figure S2). Eighty-four high quality non-synonymous coding variants (missense, nonsense and splicing variants (within 2 bp of a splice site)) remained after the exclusion of singletons from our sequencing results, variants that did not validate and variants within the MHC region. We then evaluated these variants in an independent cohort of North-American individual of European descent from the NIDDK IBD genetics consortium (754 cases and 1008 controls); only variants detected in this independent cohort were kept for follow-up genotyping. Following assay design, 42 SNPs were genotyped using Sequenom MassARRAY iPlex200 chemistry in 6 independent follow-up case-control cohorts (7292 cases and 8018 controls) (Table S3). Because of design constraints and assay failures, not all markers were examined in all follow-up sample sets. For a subset of these variants, further genotyping data was obtained from the International IBD Genetics Consortium Immunochip data (7143 UC, 12186 controls)
For all cohorts, UC was diagnosed according to accepted clinical, endoscopic, radiological and histological findings.
Genotyping of the NIDDK IBDGC cohort, as well as the Italian and Dutch cohorts was performed at the Laboratory for Genetics and Genomic Medicine of Inflammation (www.inflammgen.org) of the Université de Montréal.
NIDDK IBD Genetics Consortium (IBDGC) samples were recruited by the centers included in the NIDDK IBDGC: Cedars Sinai, Johns Hopkins University, University of Chicago and Yale, University of Montreal, University of Pittsburgh and University of Toronto. Additional samples were obtained from the Queensland Institute for Medical Research, Emory University and the University of Utah. Medical history was collected with standardized NIDDK IBDGC phenotype forms. Healthy controls are defined as those with no personal or family history of IBD.
The Italian samples were collected at the S. Giovanni Rotondo “CSS” (SGRC) Hospital in Italy.
The Dutch cohort is composed of ulcerative colitis cases recruited through the Inflammatory Bowel Disease unit of the University Medical Center Groningen (Groningen), the Academic Medical Center (Amsterdam), the Leiden University Medical Center (Leiden) and the Radboud University Medical Center (Nijmegen), and of healthy controls (n = 804) of self-declared European ancestry from volunteers at the University Medical Center (Utrecht).
Genotyping of the German cohort was performed at the Institute for Clinical Molecular Biology
Christian-Albrechts-University in Kiel. German patients were recruited either at the Department of General Internal Medicine of the Christian-Albrechts-University Kiel, the Charité University Hospital Berlin, through local outpatient services, or nationwide with the support of the German Crohn and Colitis Foundation. German healthy control individuals were obtained from the popgen biobank.
Genotyping of Swedish UC cases and controls was performed at Karolinska Institutet's Mutational Analysis core facility (MAF). Swedish ulcerative colitis patients and controls were recruited at the Karolinska University Hospital, Stockholm, and at the Örebro University Hospital, Örebro, Sweden.
Genotyping of the Belgian cohort was performed at the Genomics Core Facility at UZ Leuven, using a MassARRAY iPLEX (Sequenom). Belgian patients were all recruited at the IBD unit of the University Hospital Leuven, Belgium; control samples are all unrelated, and without family history of IBD or other immune related disorders.
All patients and control subjects provided informed consent. Recruitment protocols and consent forms were approved by Institutional Review Boards at each participating institutions. All DNA samples and data in this study were denominalized.
Association analysis of follow-up genotyping data was performed using the previously described mega-analysis of rare variants (MARV) approach [7]. Briefly, this method evaluates significance of association from stratified sample, using within sub-cohort permutation of individual phenotypes to provide the test statistic. This approach is robust to population stratification and to deviation from Hardy-Weinburg equilibrium.
We downloaded and analyzed several Gene Expression Omnibus (GEO) public microarray datasets including: (a) Expression data from newborn mice infected with Shigella flexneri; GSE9785 (b) Transcription profiles of colon biopsies from UC patients and healthy controls; GSE11223 (c) Steady-state gene expression data of Tuberculosis infected human primary dendritic cells; GSE34151 (d) PBMC transcriptional profiles in healthy subjects, patients with Crohn's Disease, and patients with Ulcerative Colitis; GSE3365, (e) Transcription profiles of colon biopsies from Crohn's patients and healthy controls; GSE20881, (f) Transcription profile of mouse small intestine epithelium vs. mesenchyme; GSE6383, (g) Gene expression in HNF4 null mouse colons compared to control colons; GSE3116, and (h) Microarray profiles of mouse epithelial colon harboring conditional knock out of HFN4A; GSE11759. Each of these datasets was normalized using quantile normalization routine in MATLAB. Genes were tested for significant differences between pairs of control and stimulated/treated samples within each experiment. After selecting genes with nominal P<0.05, estimated using an unpaired T-test, expression of RNF186 was evaluated whether it passed the significance threshold or not. The results of processing all these datasets are shown in Table S7 and Figures S3, S4, S5, S6, S7, S8, S9, S10, S11, S12, S13, S14. For transcriptional network analysis, we used Metacore's suite of network building algorithms to expand the sub-network around RNF186. The algorithm searches through a manually curated knowledgebase of molecular interaction to identify bidirectional connectivity with genes, proteins and small molecules. The search was constrained to expand the overall network size up to 50 components. Given that the bioinformatic analyses suggested that HNF4A controlled the expression of RNF186, we directly tested for their co-expression in a panel of RNAs from a variety of human tissues. Specifically, expression levels of RNF186 and HNF4A were evaluated using a custom expression array from Agilent, which was designed to include an independent probe for each exon of the genes tested (Figure S1). Briefly, total RNA from bone marrow, heart, skeletal muscle, uterus, liver, fetal liver, spleen, thymus, thyroid, prostate, brain, lung, small intestine and colon were purchased from Clontech Laboratories. A reference RNA sample was also included that consisted of an equal mix from 10 different human tissues (adrenal gland, cerebellum, whole brain, heart, liver, prostate, spleen, thymus, colon, bone marrow). With the exception of the small intestine (RIN = 7.6), all RNAs had a RNA Integrity Value (RIN) value ≥8 (range 8.0–9.3) as measured by Agilent 2100 Bioanalyzer using the RNA Nano 6000 kit (Agilent Technologies). Labeled cRNA was then synthesized from 50 ng of each RNA sample using the Low Input Quick Amp WT labeling kit (Agilent Technologies) according to the manufacturer's protocol. Quantity and quality of labeled cRNA samples were assessed by NanoDrop UV-VIS Spectrophotometer. Sample hybridization was performed according to the manufacturer's standard protocol and microarrays were scanned using the Sure Scan Microarray Scanner (Agilent technologies). An expression value was obtained for each gene in each replicate by calculating the geometric mean of all probes within the gene, followed by a median normalization across all genes on the array. A geometric mean and geometric standard deviation was calculated from at least 3 independent measurements for each tissue.
|
10.1371/journal.pbio.1000047 | Two Separate Interfaces between the Voltage Sensor and Pore Are Required for the Function of Voltage-Dependent K+ Channels | Voltage-dependent K+ (Kv) channels gate open in response to the membrane voltage. To further our understanding of how cell membrane voltage regulates the opening of a Kv channel, we have studied the protein interfaces that attach the voltage-sensor domains to the pore. In the crystal structure, three physical interfaces exist. Only two of these consist of amino acids that are co-evolved across the interface between voltage sensor and pore according to statistical coupling analysis of 360 Kv channel sequences. A first co-evolved interface is formed by the S4-S5 linkers (one from each of four voltage sensors), which form a cuff surrounding the S6-lined pore opening at the intracellular surface. The crystal structure and published mutational studies support the hypothesis that the S4-S5 linkers convert voltage-sensor motions directly into gate opening and closing. A second co-evolved interface forms a small contact surface between S1 of the voltage sensor and the pore helix near the extracellular surface. We demonstrate through mutagenesis that this interface is necessary for the function and/or structure of two different Kv channels. This second interface is well positioned to act as a second anchor point between the voltage sensor and the pore, thus allowing efficient transmission of conformational changes to the pore's gate.
| Voltage-dependent ion channels open with a voltage dependence that is remarkably steep. This steep voltage dependence, which is essential to the propagation of nerve impulses, originates in the interaction between voltage-sensor domains of the ion channel and its pore. The voltage-sensor domains transmit voltage-driven conformational changes to the pore. To understand how this “electromechanical coupling” mechanism works, we have studied the protein–protein interfaces that connect the voltage sensors to the pore using bioinformatics, electrophysiological recordings, site-directed mutagenesis, and chemical cross-linking. We identify two functionally important interfaces: one links the mobile “voltage-sensor paddle” to the pore's gate near the intracellular membrane surface, while the other links an immobile region of the voltage sensor to the pore near the extracellular membrane surface. The two interfaces encompass only a small fraction of the voltage-sensor surface area, but appear to operate in unison to enable voltage-driven conformational changes within the voltage sensor so as to efficiently regulate the pore's gate.
| Voltage-dependent ion channels mediate electrical impulses and thus enable the rapid transfer of information along the cell surface. These impulses underlie information processing by the nervous system, muscle contraction, and many other important biological processes [1]. Members of the large family of voltage-dependent cation channels—including K+, Na+, and Ca2+ selective channels—all share a common architecture consisting of a central ion-conduction pore surrounded by four voltage sensors located on the perimeter. The atomic structures of voltage-dependent K+ channels (Kv channels), determined by x-ray crystallography, have provided the first detailed pictures of voltage-dependent ion channels [2–5]. Through the combination of atomic structural, biochemical, and electrophysiological data, we are beginning to decipher the principles by which voltage-dependent ion channels function as molecular-scale electromechanical coupling devices.
The pore entryway near the intracellular membrane surface is able to constrict (close) and dilate (open) through motions of S6 “inner helices” that define the pore entryway [6–8]. S4-S5 “linker helices” form a cuff surrounding the inner helices and connect the voltage sensors to the pore [4,7]. In the atomic structures of Kv1.2 and a mutant known as paddle chimera, the S4-S5 linker helices are positioned in such a manner that conformational changes within the voltage sensors can easily be transmitted to the inner helices in order to facilitate constriction or dilation of the pore [4,7].
The voltage sensors consist of four membrane-spanning helical segments named S1 through S4. S3 is actually two helices referred to as S3a and S3b. In all of the crystal structures determined, S3b forms with S4 a helix-turn-helix called the voltage-sensor paddle [2–5]. The S4 component of this paddle contains arginine residues that are distributed within the membrane electric field: this positioning of charged amino acids enables the transmembrane voltage to exert an electrostatic force on the voltage sensor, which can bring about conformational changes within the sensor. Accessibility studies in lipid membranes indicate that the S4 helix is displaced by approximately 15 Å across the membrane in association with the voltage-dependent conformational changes [9–11].
In this study, we address the issue of how conformational changes within the voltage sensor are transmitted to the pore. When the electric field within the membrane exerts force on the charged components of the voltage sensor, how is this force transmitted efficiently to the gate? The S4-S5 linker is essential, because it appears to act structurally as a mechanical lever on the pore's gate. But such an action would seem to require a second interface that would serve to fix the voltage sensor's position relative to the pore. We identified the second interface through statistical coupling analysis (SCA) of 360 Kv channel sequences [12–14], and we showed by experiment the importance of this interface to Kv channel function.
Amino acid sequences of 360 Kv channels representing all Kv subfamilies were chosen by PSI-BLAST [15] (e-score < 0.001) and aligned by CLUSTALW [16] and structure-guided manual adjustment (Figure 1A). The sequences include Kv1 to Kv10, ERG, HCN, BK, bacteria, archea, and plant Kv channel families. A representative sequence alignment is shown (Figure 1A). SCA was performed using this alignment of 360 family members (see Materials and Methods) [14]. Figure 1B displays the degree to which the amino acid at a position in the sequence (vertical axis) is sensitive to constraint on the type of amino acid at another position (referred to as a perturbation, horizontal axis) using a color scale ranging from blue (insensitive) to red (sensitive) [14]. Regions of relative sensitivity (existence of co-evolved residues) and insensitivity (absence of co-evolved residues) are apparent. A cluster analysis was used to identify a self-consistent set of co-evolved residues (Figure 1B–1D) [14]; these are mapped onto the atomic structure of the paddle chimera channel (Figure 2A–2C) (Protein Data Bank [PDB; http://www.rcsb.org/pdb/home/home.do] ID 2R9R) [4]. As observed in other protein families, the co-evolved residues form a physically connected network (shown in van der Waals sphere) [12,14]. Overall, there are two interconnected layers within the co-evolved residue network. A top layer surrounds the pore's selectivity filter and extends into the S1 helix of the voltage sensor where S1 contacts the pore at the extracellular membrane surface (Figure 2A and 2B). A bottom layer includes the pore's S5 and S6 helices and extends into the intracellular half of the voltage sensor via the S4-S5 linker helix (Figure 2B and 2C). The S6 helices, also known as the inner helices, form the “activation” gate by opening and closing the pore at its cytoplasmic entryway. The bottom layer of the connected network defines a solid cuff surrounding this gate and extends to the voltage sensors (Figure 2B and 2C).
A comparison of published mutagenesis data on Kv channels with SCA shows a good correlation over S1, S2, S3a, and the C-terminal extent of S4: when mutated, many residues identified by SCA have a large impact on channel gating (Figure 3). SCA is insensitive to near-absolutely conserved residues (highlighted yellow in Figure 3), because insufficient variation precludes detection of co-variation.
The voltage-sensor paddle (S3b and S4 through the fourth arginine) [2] is interesting in that mutations are known to have large effects on channel gating, but paddle residues are not part of the co-evolved set of amino acids (Figure 3). Involvement of the arginine positions may be undetectable due to a high degree of conservation. However the same is not true for other amino acid positions in the paddle that exhibit variation, are demonstrably important by mutation [17], and yet do not appear in the SCA defined co-evolved set of residues (Figure 3). Based on this analysis, we conclude that although the voltage-sensor paddle is important for the function of Kv channels, its residues (arginine excluded) are not co-evolved with amino acids elsewhere on the channel, either in the S1-S2 half of the voltage sensor or in the pore. The lack of co-evolution is consistent with the experiments of Swartz and colleagues demonstrating transferability of the voltage-sensor paddle among voltage sensors of different origins [18]. These properties of the voltage-sensor paddle—absence of co-evolved interfaces and transferability—seem compatible with the idea that the voltage-sensor paddle undergoes motion during channel gating.
The majority of co-evolved residues within the voltage sensor extend their side chains in toward its core rather than out toward its surface, consistent with the idea that the voltage sensor is largely an independent and self-contained domain [3,5] (Figure 2A–2C). There are however two exceptions: the S4-S5 linker, which appears “attached” to primarily the S6 helices near the cytoplasmic membrane surface (Figure 2C), and the S1 helix, which appears “attached” to the pore helix near the extracellular membrane surface (S1–pore interface) (Figure 2A and 2B). Mutational studies have shown that the S4-S5 linker is important for coupling voltage sensor motions to pore gating [19]. Presumably, this is why the S4-S5 linker interface contains co-evolved amino acids, so that it can form a “correct” interface with the S6 helices. The S1–pore interface in contrast to the S4-S5 linker interface has not been studied to a great extent: it includes three residues from the pore helix (I361, P362, and F365; the numbering is based on rat Kv1.2 unless otherwise stated) and two from the S1 helix (C181 and T184), all belonging to the co-evolved set (Figures 3 and 4A).
We note that the crystal structure shows physical contacts between the voltage sensor and pore that are more extensive than those defined by SCA (Figure 2D). For example, physical contacts exist between the S4 helix of the voltage sensor and the S5 helix of the pore (blue surface, Figure 2D), however, this “interface” does not contain co-evolved amino acids reaching across it.
The presence of co-evolved residues across the S4-S5 linker to S6 interface seems to corroborate previous studies demonstrating the importance of this region to channel gating. The S1–pore interface by contrast has not been studied in a highly systematic manner. The presence of co-evolved residues leads us to suspect that this second interface is important as well. To test this hypothesis, we ask two questions: How is the function of the channel affected when (1) the S1–pore interface is disrupted and (2) is constrained by a covalent cross-link?
A direct albeit crude method of testing the importance of a protein–protein interface is to disrupt the interaction by Trp or Ala mutation on one side of the interaction surface. Previous scanning mutagenesis studies have shown that mutations of the three residues (I361, P362, and F365) on the pore side of the S1–pore interface yield either nonfunctional channels or pronounced effects on gating (a shift in the equilibrium between closed to open states) [20,21]. In a Trp mutagenesis study of Shaker, Miller and co-workers showed that C245 in Shaker had a large impact on channel gating, consistent with the SCA analysis [22] (Figure 3). C245 (C181 in Kv1.2) is one of the two residues on S1 that form the S1–pore interface, the other residue being T248 (T184 in Kv1.2). However, the final C-terminal residue on S1 tested in their Trp scanning mutagenesis was L246 (L182 in Kv1.2). Hence we extended the Trp scanning experiments to include additional amino acids. Shaker RNAs containing L246W, E247W, T248W, and L249W in Shaker (L182, E183, T184, and L185, respectively, in Kv1.2) (Figure 4A) as well as wild-type were made and injected into oocytes, and ionic currents were measured using two-electrode voltage clamp (Figure S1). We used L246W as a control to reproduce the work from Hong and Miller. As shown in Figure 4B, the midpoint of activation (V50) of L246W is similar to that of wild-type as was reported [22] (Table 1). E247W mutation has a large impact on channel gating, as the V50 is shifted by more than 50 mV (Figure 4B and Table 1). This amino acid does not point into the interface but it does form a salt bridge to a gating charge Arg on S4 [4]. The large impact of the E247W mutation likely stems from destabilization of the open state by precluding formation of a salt bridge observed in the open conformation crystal structure. Mutation at the next position along the S1 helix, T248W, which points directly at the interface, leads to no detectable current (Table 1). This is an unusual outcome, because voltage sensors are in general rather tolerant to mutation: in combined experiments from different studies [22,23], only two positions—I237 (I173 in Kv1.2) and R297 (R240 in Kv1.2)—fail to tolerate mutation to Trp (Figure 3), and both of these are positioned in the core of the voltage sensor. The outcome of the T248W mutation could mean either that no channels are targeted to the cell membrane or that channels are present but not functional. Either result supports the importance of the S1–pore interface to channel structure and function. The next position along S1, L249W (L185 in Kv1.2), yielded functional channels in oocytes (Figure S1), and the V50 was similar to wild-type Shaker (Figure 4B and Table 1). Thus, we observe a sharp transition from nonfunctional to wild-type like behavior when we introduce the Trp residue adjacent to but not directly on the interface.
We also studied the S1–pore interface using an approach that would seem to be less disruptive than Trp substitution. The structure of a different 6-transmembrane (6-TM) channel, MlotiK1, was determined by Clayton et al. [24]. Although MlotiK1 is not a voltage-dependent K+ channel, it is related to Kv channels and has a similar architecture. Comparison of the MlotiK1 structure with the paddle chimera structure offers interesting clues concerning potentially important chemical interactions between S1 and the pore. A superposition of the paddle chimera and MlotiK1 structures made by aligning the pores shows that the S1–S4 domains coincide in space at only a single location, which corresponds to the interface between S1 and the pore (Figure 5A). In other words the S1–S4 domains adopt different orientations with respect to the pore, but the S1–pore interface is preserved. A more detailed comparison even shows chemical similarities within the S1–pore interfaces: the hydroxyl group of T184 in the paddle chimera (T29 in MlotiK1) forms a hydrogen bond with the backbone carbonyl oxygen of C181 (A26 in MlotiK1) on S1 (helix capping) and the backbone amide of I361 (numbering based on Kv1.2) (I162 in MlotiK1) on the pore helix (Figure 5B). This dual mode of interaction seems to require stringent specifications of the side chain functional group. Inspired by this observation, we generated further mutations in Shaker at T248 (T184 in Kv1.2) and tested them for channel function to examine the importance of the hydroxyl group at this position. Because T248W did not yield functional channels, we reasoned that if a tryptophan mutant introduces too severe a steric clash with the closely packed side chains in the surrounding region, replacement with an isosteric valine or a smaller alanine residue should be well tolerated. However, to our surprise, neither T248V nor T248A produced detectable currents (Table 1). This outcome is surprising, because Kv channels have been extensively studied with alanine scanning mutagenesis and very few mutations abolish function altogether [17]. On the other hand, replacement of T248 with serine resulted in functional channels, similar to wild-type Shaker (Figure 5C and Table 1). These data underscore the importance of the hydroxyl group at the S1–pore interface and the importance of the S1–pore interface to channel function.
We further examined the S1–pore interface by covalently linking the two surfaces together through disulfide bridge formation. For these experiments, we turned to KvAP since it has been extensively tested in the bilayer system with mutagenesis and chemical modifications [9,11]. The use of a different Kv channel also allows us to assess whether the applicability of ideas concerning the S1–pore interface applies to Kv channels that are substantially different than Shaker. Five double cysteine mutants of KvAP containing one cysteine in S1 and another on the pore were tested. The channels were expressed in Escherichia coli, purified in the presence of detergent, reconstituted into lipid vesicles under reducing conditions, and then air-oxidized in the vesicles. Among the five combinations of double cysteine mutants (see Text S1 for the list), only one pair—T47C and V183C in KvAP (T184 and I361 in Kv1.2)—showed significant cross-linking of subunits on nonreducing SDS-PAGE (unpublished data), indicating that a disulfide bridge can be formed across the interface between S1 and the pore helix. Oxidized mutant channels in the planar bilayer system (see Materials and Methods) resulted in brief channel openings and small non-inactivating macroscopic currents (Figure 6A and 6B). Internal barium and external charybdotoxin, well-known K+ channel inhibitors, were used to confirm the identity of the channels as KvAP (Figure 6). In the reduced state (achieved by adding DTT to the channels in membrane vesicles prior to fusion with the bilayer), the same double mutant channels exhibited properties more similar to wild-type KvAP channels (Figure 6 C and 6D). Reduced channels are quickly converted back to oxidizing gating behavior by addition of oxidizing agent (Cu2+-phenanthroline) to the bilayer (unpublished data). These data indicate that a disulfide bridge across the interface is associated with an alteration of gating. In other words, channel gating is sensitive to reversible chemical modifications of the S1–pore interface.
This study was ultimately motivated by a puzzling feature of Kv channels revealed by the crystal structures: the voltage sensors exist as appendages without extensive contacts with the pore [4]. This being the case, how do voltage-driven conformational changes within the voltage sensors transmit mechanical forces onto the pore to open and close the gate? One region of contact in the crystal structures, that formed by the S4-S5 linkers and S6, appears to transmit motions of S4 to the gate [4,19]. A second region of contact in the crystal structures, that formed by S1 and the pore helix, was hypothesized to be important [4]. The present study tests this hypothesis with a statistical analysis of protein sequences and systematic experiments. We show that the S1–pore interface is indeed essential to Kv channel function.
We began by using SCA to identify co-evolved amino acids in Kv channels, in particular those crossing the interface between the voltage sensor and the pore. SCA reports information derived solely from sequence data and thus are independent of atomic structural data. We then map the set of co-evolved amino acids onto the atomic structure of the paddle chimera Kv channel in order to inspect their locations. We do not assume a priori that co-evolved amino acids identified by SCA are necessarily important. Instead, we use the SCA results to motivate new experiments and to interpret old experiments. In the end, based on experimental data, we find a strong correlation between co-evolution and importance to structure and/or function in the Kv channel family. Through this approach, we reach what we believe is a new insight into the function of Kv channels.
Many of the co-evolved amino acids identified by SCA have been studied in the past through mutation and are known to influence channel function (Figure 3). Amino acids in the pore surrounding the selectivity filter, when mutated, affect ion conduction and structure [25,26] as well as gating [21]. Mutations in the pore surrounding the S6 helix bundle crossing (gate) appear to influence a late-opening transition in gating [21]. Furthermore, co-evolved amino acids in the pore are coupled to each other in mutant cycle analysis of gating [21,27,28]. Thus, co-evolved amino acids in the pore in some cases appear to be important for the structure of the selectivity filter and in other cases for stabilizing conformations of the pore associated with gating states.
At two locations, the S4-S5 linker and the extracellular extent of S1, co-evolved residues cross the interface between the voltage sensor and the pore. It is well established through mutational studies that the S4-S5 linker plays an important role in coupling voltage-sensor action to pore gating. In their effort to attach functional voltage sensors to the non–voltage-dependent K+ channel KcsA, Lu and colleagues discovered that compatibility across the S4-S5 linker-to-pore interface is required (i.e., amino acids making both sides of this interface must come from the same voltage-dependent channel) [19]. The S4-S5 linker shows up in the co-evolved set of amino acids presumably because it is under selective pressure to link voltage sensor actions to pore gating through the protein–protein interface it makes with the pore (Figure 3).
The S1–pore interface identified by SCA analysis in the present study was not anticipated from past mutational studies, mainly because this region of Kv channels has not been studied in a highly systematic manner. In retrospect, several past mutational studies hinted at the importance of this region but no physical interpretation was provided [20–22]. The crystal structures of Kv1.2 and paddle chimera showed that S1 makes an apparently physically tight contact with the pore over a small area near the extracellular membrane surface [4,5]. Amino acids on both sides of the S1–pore interface turn out to be part of the co-evolved set (Figure 2). Experiments presented here demonstrate that this interface is important for channel function (Figures 4, 5, and 6). Through disruptive Trp substitution, through more subtle mutations of a specific hydrogen-bonding side chain hydroxyl in the Shaker K+ channel, and through disulfide cross-link oxidation and reduction in the KvAP channel, we demonstrate the functional importance of the S1–pore interface. It is worthwhile to note here that the sole purpose of our experiments with disulfide cross-linking at the S1–pore interface is to address the binary question whether this is a functionally important interface or not. Disulfide bond formation has been successfully used to probe three-dimensional proximity of different parts of proteins that might be distal in primary sequence. However, it is a fact that disulfide bonds can perturb local protein structure [29,30] and affect function. Given the chemically complicated nature of interactions formed at protein–protein interfaces, it is easy to see how placement of an engineered disulfide linkage might not allow the exact recapitulation of the native structural state and interactions. We attribute the difference between the oxidized and reduced versions of the mutant channel to such an effect. The point we wish to make is this: by constraining the S1–pore interface with a disulfide linkage, and by reversing the chemistry with reducing agents, we can affect the function of Kv channels. Thus, we conclude that the S1–pore interface is an important protein–protein contact for normal Kv channel function.
What might be the role of this interface? The crystal structures and associated functional studies assessing motion within voltage sensors have led to the hypothesis that with respect to motion along the transmembrane axis, voltage sensors contain a stationary half (S1, S2, and S3a) and a mobile half (S3b, S4, and S4-S5 linker) [4,7,11,31]. In this hypothesis, the electric field within the membrane exerts force on the charged S4 amino acids and brings about a motion of the mobile half: the voltage-sensor paddle (S3b-S4) is proposed to move about a “hinge” between S3a and S3b and exert a force onto the S4-S5 linker. The linker constricts (closes) or dilates (opens) the inner helix bundle (S6 helices) to gate the pore. In this view, the S1–pore interface might serve to brace the stationary half of the voltage sensor with respect to the pore, thus allowing a more efficient transference of force by the voltage sensor on the gate (Figure 7).
Multiple amino acid sequences representing the Kv family were obtained from the non-redundant database using PSI-BLAST (e-score < 0.001) [15]. KvAP, Shaker, and BK channels were used for initial searches, and multiple iterations of PSI-BLAST were performed. Sequences that contain both the K+ channel selectivity filter (TVGYG or similar) and the voltage-sensor sequences were selected. Non–voltage-dependent channels (i.e., CNG and SK channels) were manually removed based on the annotation in the database. Three hundred and sixty Kv sequences were obtained and they include eukaryotic Kv1 to Kv10, ERG, HCN, BK, bacteria, archea, and plant Kv channel families from 107 species (see Text S2 for the list of species). The full sequence alignments are sufficiently diverse that positions with low conservation show amino acid frequencies near to their mean values found in all natural proteins [12,14]. The sequences were initially aligned using ClustalW [16] and manually adjusted based on the structures of KvAP and the paddle chimera. Only amino acids within the voltage sensor or pore regions (residues 144–417 in rat Kv 1.2) were included for alignment, whereas intracellular domains (i.e., T1, PAS, and CNB domains) were removed since they are not present in all Kv channels. Accessibility studies were used to aid in the alignment of S4 [32–35]. The full sequence alignment is provided as supporting information (Text S3).
The code for the SCA was provided by S. W. Lockless (Rockefeller University). The calculation was performed as described [14]. The subalignment size cutoff value for the choice of perturbation was >0.4. This size cutoff value led us to choose 95 site-specific perturbations to build the statistical coupling matrix. Cutoff values from 0.35 to 0.45 did not change significantly the positions that form the final cluster. Two-dimensional hierarchical clustering of the matrix was carried out with MATLAB Ver. 6.1. (Mathworks) using the city-block distance metric as described [14]. The clustering algorithm is based on coupled two-way clustering analysis developed for gene microarray data [36]. After the first clustering, a sub-matrix (containing 143 positions and 59 perturbations) was extracted, and focused independent clustering was performed to refine the cluster. The second cluster was chosen based on the cutoff of the average perturbation value of 1 kT* in units of “statistical energy” [12].
Mutations were introduced in Shaker-IR [37] cDNA in the pBluescript KS (+) vector by the QuikChange method (Stratagene) and confirmed by sequencing the entire cDNA. These Shaker constructs were linearized with HindIII, and RNA was prepared by in vitro transcription with T7 RNA polymerase (Promega). mRNA was injected into Xenopus laevis oocytes, and K+ currents were recorded using a two-electrode voltage clamp (OC 725C, Werner Instrument Corporation) 1–3 d after injection. Data were filtered at 1 kHz (8-pole Bessel). Microelectrodes typically measured resistances in the range 0.3–0.8 MΩ when filled with 3 M KCl. Bath solution contained (in mM): 96 NaCl, 2 KCl, 0.3 CaCl2, 1 MgCl2, 5 HEPES (pH 7.6). Oocytes were typically held at −80 mV and stepped for 250 ms to different test voltages followed by repolarization. All experiments were carried out at room temperature. Voltage-activation curves were generated using the measured tail currents and fitted to a two-state Boltzmann equation:
where I/Imax is the normalized tail current amplitude, z is the effective charge, V50 is the activation half voltage, F is the Faraday constant, R is the universal gas constant, and T is temperature.
All mutagenesis was performed using the QuikChange method (Stratagene). The single native cysteine in KvAP was mutated to serine (C247S). Double cysteine mutant constructs were made on this cysteineless background. Mutant channels were expressed in E. coli and purified in the presence of detergents as described [38]. The channels were maintained under reducing conditions by the inclusion of 10 mM dithiothreitol (DTT) in the buffers following metal-affinity purification step. The purified channels were reconstituted in POPE : POPG (3:1) lipid vesicles as described elsewhere [39] in the presence of DTT. Following reconstitution, DTT was removed from the buffers and cross-linking was induced by air-oxidation and dialysis at room temperature for 3–5 d, and subsequently verified by non-reducing SDS-PAGE.
Electrophysiology of mutant KvAP channels was performed essentially as described [38], except that planar bilayer membranes were painted with 1,2-diphytanoyl-sn-glycero-3-phosphocoline (DPhPC). The channels were held at −120 mV and repeatedly pulsed to +120 mV test voltage. Experiments with charybdotoxin (CTX) and BaCl2 were carried out using the abovementioned protocol prior to and after the addition of external CTX (4 μM) and internal BaCl2 (2 mM). For testing the effect of reduction of the cross-link, the cross-linked channels were incubated with the presence of 50 mM DTT overnight at room temperature and then studied using the same protocols as described above.
|
10.1371/journal.pgen.1003843 | A Novel Role for Ecdysone in Drosophila Conditioned Behavior: Linking GPCR-Mediated Non-canonical Steroid Action to cAMP Signaling in the Adult Brain | The biological actions of steroid hormones are mediated primarily by their cognate nuclear receptors, which serve as steroid-dependent transcription factors. However, steroids can also execute their functions by modulating intracellular signaling cascades rapidly and independently of transcriptional regulation. Despite the potential significance of such “non-genomic” steroid actions, their biological roles and the underlying molecular mechanisms are not well understood, particularly with regard to their effects on behavioral regulation. The major steroid hormone in the fruit fly Drosophila is 20-hydroxy-ecdysone (20E), which plays a variety of pivotal roles during development via the nuclear ecdysone receptors. Here we report that DopEcR, a G-protein coupled receptor for ecdysteroids, is involved in activity- and experience-dependent plasticity of the adult central nervous system. Remarkably, a courtship memory defect in rutabaga (Ca2+/calmodulin-responsive adenylate cyclase) mutants was rescued by DopEcR overexpression or acute 20E feeding, whereas a memory defect in dunce (cAMP-specific phosphodiestrase) mutants was counteracted when a loss-of-function DopEcR mutation was introduced. A memory defect caused by suppressing dopamine synthesis was also restored through enhanced DopEcR-mediated ecdysone signaling, and rescue and phenocopy experiments revealed that the mushroom body (MB)—a brain region central to learning and memory in Drosophila—is critical for the DopEcR-dependent processing of courtship memory. Consistent with this finding, acute 20E feeding induced a rapid, DopEcR-dependent increase in cAMP levels in the MB. Our multidisciplinary approach demonstrates that DopEcR mediates the non-canonical actions of 20E and rapidly modulates adult conditioned behavior through cAMP signaling, which is universally important for neural plasticity. This study provides novel insights into non-genomic actions of steroids, and opens a new avenue for genetic investigation into an underappreciated mechanism critical to behavioral control by steroids.
| The brain is a prominent target of steroid hormones, which control a variety of neurobiological processes and are critical to the regulation of behavior. Some effects of these hormones involve changes in gene expression and thus emerge slowly, over the course of hours or even days. Other responses to steroids occur rapidly and are independent of transcriptional regulation. Their functions and mechanisms of action are poorly understood, particularly in the context of steroid-mediated control of behavior. Here we show, using the genetic model organism Drosophila melanogaster (the fruit fly), that an unconventional, membrane-bound receptor for the molting hormone ecdysone transmits a novel form of steroid signaling in the adult brain. Our study shows that this novel form of steroid signaling has a robust interface with the classical “memory genes” that encode central components of the so-called cAMP signaling pathway, which is universally important for neuronal and behavioral plasticity. These findings underscore the significance of steroid signaling in memory processing, and provide a foundation for the genetic analysis of rapid, unconventional steroid signaling in behavioral regulation.
| Steroid hormones are essential modulators of a broad range of biological processes in a diversity of organisms across phyla. In the adult nervous system, the functions of steroids such as estrogens and glucocorticoids are of particular interest because they have significant effects on the resilience and adaptability of the brain, playing essential roles in endocrine regulation of behavior. Reflecting their importance in neural functions, steroid hormones are implicated in the etiology and pathophysiology of various neurological and psychiatric disorders, and are thus often targeted in therapies [1]–[7]. The biological actions of steroids are mediated mainly by nuclear hormone receptors—a unique class of transcription factors that activate or repress target genes in a steroid-dependent manner [8]. Substantial evidence suggests, however, that steroid hormones can also exert biological effects quickly and independently of transcriptional regulation, by modulating intracellular signaling pathways [9]. Such “non-genomic” effects might be induced by direct allosteric regulation of ion channels, including receptors for GABA [10] and NMDA [11]. Alternatively, in certain contexts, non-genomic steroid signaling could be mediated by classical nuclear hormone receptors acting as effector molecules in the cytosol [12], [13].
G-protein coupled receptors (GPCRs) that directly interact with steroids have the potential to play an important role in non-genomic steroid signaling. So far, however, only few GPCRs have been identified as bona fide steroid receptors in vertebrates [14], [15]. The G-protein coupled estrogen receptor 1 (GPER, formally known as GPR30) is the best studied GPCR that is responsive to steroids. Pharmacological and gene knockout approaches suggest that this protein has widespread roles in the reproductive, nervous, endocrine, immune and cardiovascular systems [15]. Although other G-protein coupled receptors were predicted to be responsive to steroids (e.g., the Gq-coupled membrane estrogen receptor and estrogen receptor-X), their molecular identity is not known [16], [17]. Overall, the physiological roles of the GPCR-mediated actions of steroids and the underlying molecular mechanisms remain poorly understood, and sometimes controversial, in spite of their importance [18], [19]. In particular, it is unknown how this non-canonical steroid mechanism influences neural functions and complex behaviors.
Drosophila genetics has been extensively used to study the roles and mechanisms of action of steroid hormones in vivo. The major steroid hormone in Drosophila is the molting hormone 20-hydroxy-ecdysone (20E), which orchestrates a wide array of developmental events, including embryogenesis, larval molting and metamorphosis [20]–[22]. Recent studies revealed that 20E also plays important roles in adult flies, regulating: the innate immune response [23], stress resistance, longevity [24], the formation of long-term courtship memory [25] and the active/resting state [26]. In general, the functions of 20E during development and adulthood are thought to be executed by ecdysone receptors (EcRs), members of the evolutionarily conserved nuclear hormone receptor family [21], [27], [28].
In addition to canonical ecdysone signaling via EcRs, Srivastava et al. identified a novel GPCR called DopEcR, and showed that it propagates non-genomic ecdysone signaling in vitro [29]. DopEcR shares a high level of amino-acid sequence similarity with vertebrate β-adrenergic receptors. In situ hybridization [29] and microarray data (FlyAtlas, http://flyatlas.org/) revealed that DopEcR transcripts are preferentially expressed in the nervous system. In heterologous cell culture systems, DopEcR is localized to the plasma membrane and responds to dopamine as well as ecdysteroids (ecdysone and 20E), modulating multiple, intracellular signaling cascades [29]. Furthermore, Inagaki et al. recently detected DopEcR expression in the sugar-sensing gustatory neurons of adult flies, and showed that DopEcR-mediated dopaminergic signaling enhances the proboscis extension reflex during starvation [30]. Nonetheless, little is known about whether DopEcR functions as a steroid receptor in vivo, and about how it drives responses in the central nervous system (CNS) to modulate complex behaviors. Here, we report for the first time that DopEcR mediates non-genomic ecdysone signaling in the adult brain, and that it is critical for memory processing. We also show that, during memory processing, DopEcR transmits information via novel steroid signals that interact with the cAMP pathway, a signaling cascade that is universally important for neuronal and behavioral plasticity. Our genetic study thus uncovers underappreciated GPCR-mediated functions and mechanisms of action that employ non-canonical steroid signaling to regulate the adult nervous system and, thereby, behavior.
PBac(PB)c02142 is a piggyBac transposon insertion in the second intron of the DopEcR gene (Figure 1A). Adult flies homozygous for PBac(PB)c02142 displayed a significant reduction in DopEcR transcript levels (<20% of levels in control), in both the head (Figure 1B) and the body (data not shown). Df(3L)ED4341 is a chromosomal deficiency that removes multiple genes on 3L, including DopEcR (Flybase: http://flybase.org/). Flies trans-heterozygous for PBac(PB)c02142 and Df(3L)ED4341 showed levels of DopEcR transcript comparable to those in PBac(PB)c02142 homozygotes (Figure 1B). PBac(PB)c02142 is therefore a hypomorphic allele of DopEcR, and it was mainly used in this study to investigate the functions of DopEcR in behavioral plasticity. PBac(PB)c02142 is referred to as DopEcRPB1 hereafter. DopEcRPB1 homozygotes reached adulthood and exhibited no gross morphological defects. General motor activity was not significantly impaired, as judged by analysis of reactive climbing behavior (Figure S1).
In order to obtain some insight into the endogenous expression pattern of DopEcR, we generated DopEcR-Gal4, a Gal4 driver that contains the putative enhancer/promoter sequence of DopEcR (a 588-bp DNA fragment upstream of the DopEcR transcription start site). DopEcR-Gal4 was found to induce GFP reporter gene expression preferentially in the nervous system. In the adult brain, DopEcR-Gal4-regulated reporter gene expression was particularly prominent in the mushroom body (MB) (Figure 1C and 1D). It is not likely that the endogenous DopEcR expression is accurately recapitulated by the 588-bp DNA fragment used for DopEcR-Gal4. Nevertheless, the reporter gene expression shown in Figure 1C and 1D implies the presence of the endogenous DopEcR in the MBs of the adult brain (see Discussion). Reporter gene expression driven by DopEcR-Gal4 was also observed in neuronal soma and fibers localized in each segment of the thoracicoabdominal ganglion (Figure 1F and 1G). In addition, a number of fibers connecting the ganglion to the brain, abdomen and appendages were found to be GFP-positive (Figure 1F and 1G).
To investigate the role of DopEcR in the CNS, we tested DopEcR mutations for effects on the electrophysiological properties of the adult giant-fiber (GF) pathway [31], [32]. Visual or mechanical stimulation activates the descending GF neurons (Figure 2A), triggering the stereotypical jump-and-flight response. This behavioral response is associated with a consistent pattern of spiking in both the dorsal longitudinal flight muscle (DLM) and the tergotrochanteral jump muscle (TTM) (Figure 2A). Strong electrical stimulation of the brain can bypass sensory receptors and directly trigger the neuronal circuit at the GF neurons (short-latency response) [32], [33]. Alternatively, with stimulation of the brain at lower intensity, the circuit is activated at GF afferents in the brain (long-latency response) [34]. As shown in Figure 2B (left and middle panels), both the short- and long-latency thresholds (SLT and LLT; the lowest intensities required to trigger short- and long-latency responses in the DLM) were indistinguishable between DopEcR mutants (DopEcRPB1/DopEcRPB1, DopEcRPB1/Df(3L)ED4341 and DopEcRPB1/+) and wild-type flies. This indicates that reducing DopEcR expression does not significantly affect the overall neuronal sensitivity of the GF pathway. In contrast, the refractory period (RP; the minimum time required for the GF system to recover from the 1st stimulus and fire a response to the 2nd stimulus) was significantly reduced in DopEcRPB1/DopEcRPB1 compared to control flies (Figure 2B, right panel). The RP in DopEcRPB1/Df(3L)ED4341 and DopEcRPB1/+ flies also showed a similar tendency, although the differences between these mutants and control flies did not reach statistical significance, possibly due to the weak nature of this DopEcRPB1 phenotype and the small sample numbers. Nonetheless, the shorter RP implies that circuits in DopEcR mutants are less vulnerable or more resistant to activity-dependent modifications than the relevant circuits in controls are.
Diminished neuronal plasticity in DopEcR mutants was unequivocally demonstrated when habituation of the GF pathway was analyzed. Habituation is a simple form of non-associative learning, in which the reaction to a particular stimulus becomes diminished when the stimulus is applied repeatedly. Habituation does not lessen behavioral responses due to sensory adaptation or motor fatigue [35]. When electrical stimulation is repeatedly delivered across the brain, the GF pathway undergoes habituation and the probability of a motor output significantly decreases [36]. Previous studies by us and others revealed that the loci responsible for this neuronal plasticity are localized to the brain, namely neuronal circuits afferent to the GF neurons (aff; Figure 2A) [36]–[40]. Other elements in the GF pathway—including the GF neuron, the peripherally synapsing interneuron (PSI), and the motor neurons that innervate the flight and jump muscles (DLMs and TTMs; Figure 2A)—are robust enough to reliably respond to sustained high-frequency stimuli (up to ∼100-Hz) [32], [33], [36]. In our experiments, control flies became rapidly habituated to 5-Hz stimulation of the brain, as evidenced by a failure of their DLM to respond (Figure 2C, control). The reduced behavioral response was not a consequence of sensory adaptation or motor fatigue because the response was readily recovered by a novel stimulus, such as an air puff (dishabituation; Figure 2C, control). In contrast to controls, DopEcRPB1 homozygotes and DopEcRPB1/Df(3L)ED4341 trans-heterozygotes consistently showed a delay in habituation (Figure 2C), and thus their cumulative response was greater than that of controls (Figure 2D). DopEcRPB1 heterozygotes (DopEcRPB1/+) showed a similar tendency, although the effect was less extreme (Figure 2C and 2D). When habituation was arbitrarily defined as five or more consecutive failures, DopEcRPB1 mutants needed more repetitive stimulations than control flies to reach habituation status (Figure 2E). The average numbers of 5-Hz stimuli required for habituation were 46±31 and 637±236 in control flies and DopEcRPB1 homozygotes, respectively (Figure 2E). DopEcRPB1 heterozygotes also showed a slow habituation phenotype (Figure 2C–E). These results demonstrated that DopEcR is an essential modulatory component of the GF pathway, and that its endogenous role is to positively regulate activity-dependent modification of the relevant CNS neuronal circuits.
In light of the abnormalities in GF habituation, we next tested DopEcR mutants for experience-dependent courtship suppression, an ethologically relevant associative-learning paradigm [41], [42]. In wild-type control males (+/+) and DopEcRPB1 heterozygous males (DopEcRPB1/+), 1 hour of conditioning with a mated female induced “courtship memory”, which was readily detectable 30 minutes after conditioning as a statistically significant, experience-dependent reduction in courtship activity (P = 0.0004 for control and 0.0046 for DopEcRPB1/+; Figure 3A). In contrast, DopEcRPB1 homozygotes and hemizygotes (DopEcRPB1/Df(3L)ED4341) did not display courtship memory (P>0.05; Figure 3A). These results strongly suggested that DopEcR is essential to the processing of courtship memory. The performance indices (PIs; % decrease in courtship index in response to courtship conditioning, see Materials and Methods for details) of these DopEcRPB1 mutants at 30 minutes post conditioning were significantly lower than that of wild-type flies (P<0.05; Figure 3A). Notably, although DopEcRPB1 homozygotes did not display courtship memory at both 15 and 30 minutes after conditioning (P>0.05), they exhibited memory immediately after courtship conditioning (P = 0.00026). The PIs of DopEcRPB1 homozygotes for 0 and 30 minutes after conditioning were significantly different from each other (Krustal-Wallis One-Way ANOVA; P<0.05; Figure 3B). These results indicated that DopEcR mutants retain the ability to acquire courtship memory, but that the memory is labile and severely disrupted within 30 minutes.
To confirm that the memory phenotype in DopEcR mutants is due to the defect in DopEcR function, we examined the effects of DopEcR RNAi on courtship memory. When the DopEcR RNAi was conditionally and globally expressed in adult flies using the RU486-inducible driver tubulin5-GeneSwitch-Gal4 (tub5-GS-Gal4; a gift from Dr. Pletcher, University of Michigan) [43], the level of DopEcR transcripts was significantly reduced in an RU486-dependent manner (Figure S2). When DopEcR expression was conditionally knocked down by RNAi in this context, the courtship memory phenotype of the DopEcR mutants was mimicked (Figure 3C). These results support our conclusion that adult male flies require functional DopEcR for normal courtship memory.
Next we sought to identify the sites within the nervous system in which DopEcR is required for the processing of courtship memory. We found that DopEcRPB1 males displayed courtship memory (P = 9.6×10−6) when the wild-type DopEcR transgene was expressed using DopEcR-Gal4 (Figure 4A). In contrast, control DopEcRPB1 males carrying only DopEcR-Gal4 or UAS-DopEcR were defective for courtship memory (Figure 4A). The PIs of these control males were significantly lower than that of DopEcRPB1 males carrying both the Gal4 and UAS constructs (P<0.001 and P<0.05, respectively; Figure 4A). DopEcR-Gal4 directed gene expression in the adult brain, particularly in the neurons of the MB (Figure 1C). These observations, together with the importance of the MBs in processing courtship memory [25], [44], [45], led us to suspect that the rescue of the DopEcR memory phenotype by DopEcR-Gal4 was a consequence of the expression of wild-type DopEcR in the MB. This possibility was tested by performing rescue experiments for DopEcRPB1 mutants in which UAS-DopEcR expression was driven using three MB-positive Gal4 lines: c772, c739 and 201Y. Courtship memory was restored in DopEcRPB1 males when the wild-type DopEcR cDNA was expressed using either c772 or c739 (P = 2.9×10−5 or 0.0063; Figure 4B). In contrast, the 201 y driver failed to rescue the memory defect of DopEcRPB1 mutants (Figure 4B). c772 and c739 drive gene expression in all three types of MB neurons (α/β, α′/β′ and γ) and primarily in the α/β neurons, respectively, whereas 201 y drives gene expression mainly in the γ neurons [46]. These results suggested that the MBs, in particular the α/β neurons, are the key anatomical site in which DopEcR regulates courtship memory. In support of this idea, expression of the DopEcR RNAi in wild-type MB neurons using c772 or c739 led to a lack of 30-minute courtship memory in males (Figure 4C). The PIs of males carrying both the Gal4 and UAS-RNAi constructs were significantly lower than that of control males (Figure 4C).
Dominant temperature-sensitive 3 (DTS-3) is a dominant mutant allele of molting defective (mld; personal communication, P. Maroy, University of Szeged, Szeged, Hungary), a gene that encodes a putative transcription factor required for ecdysone biosynthesis [47], [48]. We previously reported that, unlike wild-type males, DTS-3/+ males did not exhibit an increase in 20E levels in response to 7-hour courtship conditioning, and that they were defective in long-term courtship memory (courtship LTM) [25]. As shown in Figure 5A, DTS-3/+ males did not show courtship suppression 30 minutes after 1-hour conditioning. Intriguingly, when DTS-3/+ males were fed 20E (0.1 mM) for 10 minutes immediately before courtship conditioning, the courtship-memory defect was rescued and courtship suppression was observed (P = 0.0043) (Figure 5A).
disembodied (dib) is one of the Halloween-family genes encoding the cytochrome P450 enzymes that are essential for ecdysone biosynthesis [49]. When dib expression was conditionally suppressed by treating mature adult males carrying the UAS-dib RNAi (gift from Dr. O'Connor, University of Minnesota) and tub5-GS-Gal4 with RU486, they exhibited a defect in courtship memory (Figure 5B). As with DTS-3/+ males, when dib-knockdown flies were fed 20E (0.1 mM) before courtship conditioning, they displayed experience-dependent courtship suppression (P = 0.006). In the dib-knockdown flies, this rescue effect of 20E was not observed when DopEcR expression was suppressed using the DopEcR RNAi (Figure 5B). Functional DopEcR is thus required for 20E-dependent courtship memory. These findings, together with the phenotypes of the DopEcR-mutant and DTS-3/+ males (Figure 5A), strongly suggest that ecdysone signaling plays a critical role in 30-minute courtship memory, and this signaling is mediated by DopEcR.
In addition to ecdysteroids, dopamine has been shown to be a direct ligand for DopEcR [29]. We fed flies 3-Iodotyrosine (3-IY) to block dopamine synthesis and examined the effect on courtship memory. As reported previously, courtship memory was defective in these flies [50], [51] (Figure 5C). We found that when the flies were additionally fed 20E (0.1 mM) 10 minutes before courtship conditioning, courtship memory was restored in spite of the block in dopamine synthesis (P = 0.00017; Figure 5C) and the PI was significantly increased (P<0.05; Figure 5C). The compensatory effect of 20E was also observed in 3-IY-treated flies of a different genetic background (Figure 5D). In contrast, when DopEcR RNAi was conditionally expressed in dopamine-depleted adults, 20E was not able to rescue courtship memory (Figure 5D). These results show that 20E compensates for the adverse effect of dopamine deficiency on courtship memory through the actions of DopEcR.
We next examined which intracellular signaling events are involved in the regulation of courtship memory by DopEcR. Here we focused our attention on the cAMP signaling pathway, because it plays a central role in learning and memory processes in diverse animal species [52]. We investigated whether 20E and DopEcR exert their effects on courtship memory via this signaling. The functional significance of cAMP for DopEcR-mediated signaling was indicated by a previous study in heterologous cell-culture systems, showing that DopEcR modulates intracellular cAMP levels in response to ligand binding [29]. One Drosophila gene that is crucial for regulating cAMP signaling is rutabaga (rut), which encodes a type I Ca2+/CaM-dependent adenylyl cyclase (AC) [53], [54]. Loss-of-function rut mutations result in lower cAMP-synthesizing activities and affect various forms of neural plasticity, including habituation of the GF pathway [36] and experience-dependent courtship suppression [55]. Habituation in the GF pathway was suppressed in both DopEcR and rut mutants (Figure 2C–E) [36], implying that the encoded proteins may have related functions in regulating neural plasticity.
Consistent with a previous report [42], males carrying a hypomorphic rut mutant allele (rut2 or rut1084) were defective for courtship memory and showed no experience-dependent courtship suppression 30 minutes after 1-hour courtship conditioning (P>0.05; Figure 6A). Remarkably, the memory defect in rut mutants was restored when they were fed 20E (0.1 mM) for 10 minutes immediately before courtship conditioning (P = 0.0043 and 5.5×10−5 for rut2 and rut1084, respectively; Figure 6A). The PIs for rut2 and rut1084 increased significantly following treatment with 20E (P<0.05 and P<0.01 for rut2 and rut1084, respectively; Figure 6A). This pharmacological rescue of the rut memory phenotype was not observed in rut and DopEcRPB1 double mutants (Figure 6B). These results strongly indicated that DopEcR mediates the compensatory effect of 20E on defective memory in rut mutants.
Considering the significance of the MB and rut for DopEcR-mediated memory processing, we examined their relationship. Courtship memory was analyzed in adult rut2 mutants overexpressing DopEcR in the MB neurons. Courtship memory was restored by conditional overexpression of DopEcR using RU486-inducible MB-GS-GAL4 [56] (P = 1.4×10−9), and the PI increased significantly (P<0.001; Figure 6C). Although the memory defect in rut1084-mutant males was not rescued by solely overexpressing DopEcR in the MB (Figure 6D), feeding them a low concentration of 20E (0.01 mM) led to significant courtship suppression (P = 1.2×10−6; Figure 6D). Notably, administering 20E at this concentration was not sufficient to rescue the rut1084 memory phenotype in the absence of DopEcR overexpression (Figure 6D, middle). The different requirements for rescuing courtship memory in rut2 and rut1084 may reflect differences in the severity of the mutations. Indeed, an olfactory-associated memory defect in rut1084 mutants is similar to that in mutants of a presumptive rut null allele (rut1) [53], [57], whereas the rut2 memory defect is milder [58], [59]. Overall, these findings demonstrate that the memory defect in rut mutants can be compensated by strengthening DopEcR-mediated ecdysone signaling in MB neurons.
Another “memory gene” involved in cAMP signaling is dunce (dnc), which encodes a cAMP-specific phosphodiesterase (PDE) that is required for cAMP degradation [60], [61]. Like rut mutants, dnc loss-of-function mutants are defective for various types of neuronal and behavioral plasticity [36], [61], [62]. In contrast to rut mutants, dnc mutants display an increased rate of GF habituation [36], possibly reflecting the fact that rut and dnc mutations have opposite effects on cAMP levels.
As shown previously [42], hypomorphic dnc mutants (dnc1 and dnc2) did not exhibit experience-dependent courtship suppression 30 minutes after 1-hour courtship conditioning, and were therefore defective for courtship memory (Figure 6E). Notably, double mutants carrying both dnc and DopEcR loss-of-function mutations displayed courtship suppression (P = 0.0016 for dnc1/Y; DopEcRPB1 and 0.002 for dnc2/Y; DopEcRPB1; Figure 6E). In addition, dnc1 males displayed courtship memory, which manifests as significant experience-dependent courtship suppression (P = 0.0274) when DopEcR was conditionally down-regulated using tub5-GS-Gal4 in conjunction with the DopEcR RNAi (Figure 6F). dnc2 males showed a similar tendency when DopEcR was down-regulated, although the difference in CIs between naïve and conditioned flies was not statistically significant. Overall, these experimental results with rut and dnc mutants strongly suggest that DopEcR exerts its critical function in courtship memory by regulating cAMP signaling pathway.
The courtship-memory defect in rut mutants can be attributed to their inability to appropriately increase intracellular cAMP levels during courtship conditioning. Because 20E feeding and DopEcR overexpression resulted in restoration of courtship memory in rut mutants, we hypothesized that strengthening DopEcR-mediated ecdysone signaling would increase in cAMP levels in brain regions critical for memory processing, such as the MBs. To investigate this possibility, we examined the effects of 20E feeding on cAMP levels in the MBs of live adult flies using UAS-Epac1-camps, a Förster (fluorescence) resonance energy transfer (FRET)-based cAMP reporter [63]. The reporter was expressed in MBs using c772, one of the MB drivers that were effective in the rescue and phenocopy experiments described above (Figure 4). Immediately after the flies were fed 20E, the cAMP levels in the MBs were assessed by cAMP-induced changes in FRET, as the ratio between YFP and CFP signals. In wild-type males, feeding either 3 mM or 1 mM 20E caused a time- and dose-dependent decrease in the average FRET (Figure 7A). The effects of 20E on FRET are statistically significant (P<0.001; Figure 7B). Because a decrease in FRET corresponds to an increase in cAMP levels, our results indicated that 20E increases cAMP levels in the MB. This effect was eliminated by simultaneously expressing the DopEcR RNAi in the MB (Figure 7C and 7D). These data indicate that the activation of ecdysone signaling rapidly increases cAMP levels in the MBs, and that it does so through DopEcR.
Here we used genetic, pharmacological, and behavioral approaches in Drosophila to demonstrate that the steroid hormone 20E rapidly regulates behavioral plasticity via a non-genomic mechanism that is mediated by the GPCR-family protein DopEcR. This non-canonical steroid signaling pathway was found to have strong functional interactions with the classical “memory genes” rut and dnc, which encode the central components of the cAMP pathway. The identification of 20E as an important modulator of cAMP signaling in the adult Drosophila brain reveals an unprecedented opportunity—that of taking advantage of fly genetics to dissect the molecular and cellular mechanisms responsible for the non-genomic steroid signaling that underlies neuronal and behavioral plasticity.
Our electrophysiological analyses revealed that the GF pathway of DopEcR mutant flies is more resistant to habituation than that of control flies (Figure 2). Direct excitation of GF or its downstream elements would lead to a short-latency response of the DLM, which could follow high-frequency stimuli up to several hundred Hz [32], [33], [36]. In contrast, the afferent input to the GF leads to a long-latency response that is labile and fails to follow repetitive stimulation well below 100 Hz and displays habituation even at 2–5 Hz [36]–[40]. Although there is the possibility that DopEcR-positive thoracic neurons may modulate thoracic motor outputs and contribute to certain parameters of the habituation process not characterized in this study, the more effective modulation would occur in the more labile element afferent to the GF circuit rather than the robust GF-PSI-DLMn downstream pathway, which is responsible for the reliability of the escape reflex. Thus, the mutant phenotype in habituation indicates that DopEcR positively controls activity-dependent suppression of neuronal circuits afferent to the GF neurons in the brain.
Moreover, our finding that DopEcR and rut mutants have a similar GF habituation phenotype raises the possibility that DopEcR positively regulates cAMP levels in the relevant neurons following repetitive brain stimulation. Besides GF habituation, Drosophila displays olfactory habituation, which is mediated by the neural circuit in the antennal lobe [38]. Interestingly, Das et al. found that olfactory habituation is induced by enhancement of inhibitory GABAergic transmission, and that rut function is required for this neuronal modulation [64]. Similar modulation of GABAergic transmission may also be responsible for habituation of the GF pathway. It will be interesting to examine whether and how DopEcR contributes to the regulation of rut and enhanced GABAergic transmission in GF habituation.
Several studies already suggested that 20E has rapid, EcR-independent effects in Drosophila and other invertebrate species. For example, 20E was shown to reduce the amplitude of excitatory junction potentials at the dissected Drosophila larval neuromuscular junction (NMJ), and to do so within minutes of direct application [65]. Whereas treatment with 20E did not change the size and shape of the synaptic currents generated by spontaneous release, it led to a reduction in the number of synaptic vesicles released by the motor nerve terminals following electrical stimulation [65]. A similar effect of 20E was observed in crayfish, and it was suggested that the suppression of synaptic transmission by 20E may account for the quiescent behavior of molting insects and crustaceans [66]. These observations suggested that 20E suppresses synaptic efficacy under certain conditions by modulating presynaptic physiology through a non-genomic mechanism. It is possible that such actions of 20E are mediated by DopEcR. To detail the mechanisms underlying DopEcR-dependent neural plasticity, it will be worthwhile to determine if and how DopEcR contributes to 20E-induced, rapid synaptic suppression at the physiologically accessible larval NMJ, and to determine the extent to which non-genomic mechanisms of steroid actions are shared between the larval NMJ and the adult brain.
One surprising finding made in this study is that ecdysone signaling can modify the phenotypes associated with mutations in the classic “memory genes”, namely rut and dnc, through the actions of DopEcR. rut and dnc encode central components of the cAMP pathway, which is required for memory processing in vertebrates as well as invertebrates. Our demonstration that genetically and/or pharmacologically enhancing DopEcR-mediated ecdysone signaling restores the courtship memory phenotype of loss-of-function rut mutants (Figure 6A–D) suggests that 20E-mediated DopEcR activation triggers an outcome similar to rut activation, i.e., increased cAMP levels. This assumption is supported by our finding that loss-of-function dnc mutants restore courtship memory when DopEcR activity is suppressed (Figure 6E and 6F). A similar restoration of the dnc memory phenotype was previously reported in a dnc and rut double mutant [58], again supporting the idea that DopEcR positively regulates cAMP production.
The results of rescue and phenocopy experiments (Figure 4) indicate that the MB is critical for the DopEcR-dependent processing of courtship memory. Although the endogenous expression pattern of DopEcR is not known, DopEcR is thus likely to modulate cAMP levels in the MB in response to 20E during courtship conditioning. We have recently generated a new Gal4 line, in which a portion of the first coding exon of DopEcR is replaced with a DNA element that contains the Gal4 cDNA whose translation initiation codon is positioned exactly at the DopEcR translation start site (Q. Li and Y. Rao are preparing a paper describing the details of this Gal4 line). When this line was used to drive UAS-GFP, the reporter gene was widely expressed in the adult brain with prominent signals in the MB (unpublished observation). This preliminary result strongly indicates the endogenous expression of DopEcR in the MB. We have also directly shown that cAMP levels in the MB increase rapidly in flies fed 20E (Figure 7A), and that this increase does not occur when DopEcR expression is down-regulated specifically in the MB (Figure 7B). Taken together, these findings suggest that DopEcR expressed in the MB responds to 20E and acts upstream of cAMP signaling in a cell-autonomous manner.
Surprisingly, enhancement of DopEcR-mediated ecdysone signaling restored courtship memory in flies harboring a strong hypomorphic allele of rut (rut1084) (Figure 6A and 6D). A similar result was obtained even in mutants harboring a presumptive rut null allele rut1 (data not shown). These results suggest that, upon stimulation by 20E, DopEcR may be able to signal via another adenylyl cyclase that can compensate for the lack of Rut. This interesting possibility requires further investigation.
In this study, we have focused on the roles and mechanisms of action of DopEcR-mediated, non-genomic ecdysone signaling. As we previously found that 20E levels rise in the head during courtship conditioning [25], the data presented here suggest that DopEcR is activated by 20E during conditioning, triggers a rise in cAMP levels and induces physiological changes that subsequently suppress courtship behavior. This interpretation assumes that 20E directly activates DopEcR to increase cAMP levels. Previous cell-culture studies suggested that DopEcR also responds to dopamine to modulate intracellular signaling [29]. Furthermore, Inagaki et al. have demonstrated that flies respond to starvation by sensitizing gustatory receptor neurons to sugar via dopamine/DopEcR signaling [30]. We thus need to consider whether dopamine is directly involved in the processing of courtship memory through DopEcR. There is a possibility that 20E initially stimulates the production and/or release of dopamine, and that it in turn activates DopEcR and elevates cAMP levels to induce courtship memory. We think that this possibility is unlikely because even when courtship memory is disrupted by pharmacological suppression of dopamine synthesis, 20E feeding can compensate for decreased dopamine and allow restoration of memory (Figure 5C and 5D). Although dopamine plays a significant role in courtship memory [50], our results suggest that DopEcR does not act as the major dopamine receptor in this particular learning paradigm. We thus favor the possibility that dopamine contributes to courtship memory in parallel with, or upstream of, DopEcR-mediated ecdysone signaling. Consistent with this view, Keleman et al. reported that the formation of courtship memory depends on the MB γ neurons, which express DopR1 dopamine receptors, receiving dopaminergic inputs [51]. Notably, our results indicate that the processing of courtship memory requires DopEcR expression in the αβ, but not γ, neurons of the MB (Figure 4), which makes it unlikely that DopEcR is directly influenced by the dopaminergic neurons innervating γ neurons.
Ecdysone signaling through nuclear EcRs is necessary for forming long-term courtship memory that lasts at least 5 days, but appears not to have a significant effect on short-term courtship memory [25]. In contrast, we found that DopEcR-mediated ecdysone signaling is critical for habituation and 30-minute courtship memory. These findings suggest that DopEcR and EcRs control distinct physiological responses to courtship conditioning, and that the former regulates short-term memory, while the latter regulates long-term memory. Although non-genomic actions of steroid hormones have been implicated in vertebrate learning and memory [67], [68], such actions have been attributed mainly to the classical nuclear hormone receptors that function outside of the nucleus and exert roles distinct from those of steroid-activated transcription factors [12]. Although recent evidence has shown that membrane-bound receptors independent of the classical estrogen receptors are involved in estradiol-induced consolidation of hippocampal memory [69], the molecular identities of these proteins have not been established. Our findings here provide a novel framework for dissecting GPCR-mediated steroid signaling at the molecular and cellular levels. Furthermore, future analysis of the functional interplay between genomic and non-genomic steroid signaling pathways is expected to reveal novel mechanisms through which steroid hormones regulate plasticity of the nervous system and other biological phenomena.
Flies were reared at 25°C and 64% humidity, in a 12-hour light/dark cycle and on a conventional glucose-yeast-cornmeal agar medium. The DopEcRPB1 strain used in this study was produced by outcrossing with Cantonized w mutant flies. The DopEcR-Gal4 and UAS-DopEcR strains were generated in this study. For DopEcR-Gal4, the putative promoter region of DopEcR (a 588-bp upstream sequence) was fused to the yeast Gal4 gene. For UAS-DopEcR, the DopEcR coding sequence was inserted downstream of the UAS (upstream activating sequence) in the pUAST vector. Other fly strains used in this study were obtained from the following sources: c772, c739, 201 y, UAS-CD4-tdGFP and UAS-Epac1-camps (55A) (Bloomington Drosophila Stock Center); tub5-GS-Gal4 (Scott D. Pletcher, Baylor College of Medicine, Houston, TX, USA); MB-GS-Gal4 (Ronald L. Davis, The Scripps Research Institute, Jupiter, FL, USA); UAS-DopEcR RNAi (VDRC); UAS-disembodied RNAi (Michael B. O'Connor University of Minnesota, Minneapolis, MN, USA); DTS-3 and Samarkand (Anne F. Simon, Western Ontario University, Ontario, Canada). The Canton-S (2202u) strain was used as the wild-type control.
Adult brains were dissected from 3 to 5-day-old male flies in PBS and fixed for 1 hour with 3.7% formaldehyde at 25°C, in PBS containing 0.05% Triton X-100 (PBST). The brains were blocked with PBST containing 0.1% normal goat serum for 1 hour. Rabbit anti-GFP antibody (1∶1000; A11122, Invitrogen) was used for the primary antibody. The brains were counter-stained with nc82, the mouse anti-Bruchpilot antibody (1∶20; Developmental Studies Hybridoma Bank, University of Iowa). Alexa Fluor 555-conjugated anti-rabbit IgG (1∶300; Invitrogen) and Alexa 647-conjugated anti-mouse IgG (1∶300; Invitrogen) were used as secondary antibodies for detection of anti-GFP and nc82, respectively. Images were acquired as a z-stack, using FV1000 confocal microscope (Olympus). Volume-rendered images were displayed using FluoRender (http://www.fluorender.com).
Total RNA was prepared from 20 fly heads of each genotype using TRIzol solution (Invitrogen), and subjected to a reverse transcription reaction using a poly-dT20 primer and Superscript II enzyme (Invitrogen), according to the manufacturer's instructions. The DopEcR cDNA sequence was amplified by PCR using the following primers: forward, 5′-ATGCAGGAAATGAGCTACCT-3′ and reverse, 5′-CTAGTCATCTGGGTCCAACC-3′. rp49 was used as the internal control (forward, 5′-ATGACCATCCGCCCAGCA-3′ and reverse, 5′-AATCTCCTTGCGCTTCTTGG-3′). The gel images were processed using ImageJ software, to estimate the quantity of PCR products.
The preparation of flies, stimulation, recording, and analysis of muscle responses were performed as described previously [36], with some modifications. Electrical stimuli (0.1 msecond pulse) were delivered across the brain through two uninsulated tungsten electrodes inserted in the eyes (anode normally in the right eye). The action potentials in the left side leg extensor (TTM) and the right side wing depressor (DLM) were recorded as an indicator of GF pathway output [36]. Flies were given tissue paper balls (less than 1 mm in diameter) to inhibit flight, but were free to perform normal jump-and-escape reflexes. All recordings were carried out in an experimental Faraday cage covered with a black plastic sheet to reduce ambient light. To minimize the possible effects of handling and anesthesia, flies mounted for recording were rested for at least 1 hour in a humid chamber before recording. After being assessed for response thresholds (during an inter-stimulus interval, ISI, of 30 seconds), flies were rested for 5 minutes before the habituation test. Three classes of responses, with progressively greater thresholds, were identified: long-latency, intermediate-latency, and short-latency. These responses could easily be distinguished in individual flies, and were used as an “internal gauge” on which to base the stimulation intensity for the habituation test. For each test, the stimulus intensity was set at the mean value of the thresholds for the long-latency and short-latency. To avoid causing artifacts by improper handling of flies, flies that had abnormally high activities or failed to respond more than twice, consecutively, were excluded from data analysis. Dishabituation stimuli (air puffs) were provided by gently squeezing a rubber bulb connected, by tubing, to a pipette nozzle mounted 2 cm to the anterior-left of the fly. All habituation data were recorded using the software pCLAMP 5, and analyzed with clampfit in pCLAMP10. The cumulative curves of the habituation responses were plotted with custom-designed software on the Matlab 7 platform.
The courtship conditioning assay was performed at 25°C and 64% humidity, in an environment room under white light, as described previously [25], with some modifications. All males were 3–5 days old at the time of testing. They were anesthetized with CO2 and stored in isolation for at least 24 hours prior to experiments. Females used as “trainers” in courtship conditioning were 3-days old and were fertilized a day before conditioning. In the conditioning phase, virgin males were placed with unreceptive, non-virgin females (or alone in ‘pseudo-training’ experiments for naïve control males), in single-pair-mating chambers containing food medium (15 mm diameter×5 mm in depth), for 1 hour. After conditioning, males were rested individually for 30 minutes, in a glass tube (12 mm in diameter ×75 mm in depth, VWR International) containing food medium. Memory tests were performed in a courtship chamber (15 mm in diameter ×3 mm in depth) containing a freeze-killed virgin female. The male courtship behaviors were videotaped for a 10-minute test period, using DVD camcorder (Sony DCR-DVD105), and were manually scored for courtship index (CI). The CI was defined as the proportion of time spent for courtship behaviors (orientation, tapping, singing, licking and copulation attempts). We did not exclude from analysis males with a low courtship level. To compare CIs for conditioned and naïve males, we analyzed the data non-parametrically, using the Mann-Whitney U test, because the CI values were often not distributed normally. When CIs for conditioned and naïve males were significantly different (P<0.05), male courtship behavior was considered to be suppressed in an experience-dependent manner (courtship memory). Experimental data are presented in the figures as the performance index (PI), which was calculated using the following formula (after CIs were subjected to arcsine square root transformation to approximate normal distributions): PI = 100×(CIAve naïve−CIconditioned)/CIAve naïve, where CIAve naïve and CIconditioned represent the averaged CI for naïve flies and a CI for each conditioned fly, respectively. Naïve courtship levels of Canton-S, DopEcRPB1 and DopEcR RNAi (UAS-DopEcR RNAi/+; tub5-GS-Gal4/+) flies were shown in Table S1. The CIs were not statistically different between Canton-S and DopEcRPB1 (P = 0.086) and there was no statistical difference between the CIs of DopEcR RNAi males with or without RU486 treatment (P = 0.8459). Mann-Whitney U test.
Flies carrying the RU486-inducible transgene (GeneSwitch strains) were fed food containing 500 µM RU486 (Mifepristone, Sigma) or vehicle (ethanol; final concentration <2%) for 3 days prior to the experiment. 20E was fed for 10 min using Kimwipe paper soaked in 1M sucrose solution containing a particular concentration of 20E (Sigma). The 20E stock solution (25 mM) was prepared in ethanol. 3-Iodotyrosine (3-IY) was mixed into yeast paste with a final concentration of 10 mg/ml. Up to 10 newly eclosed flies were placed in vials containing fly food with 3-IY-yeast paste for 4 days.
The change in cAMP levels was monitored using the genetically encoded cAMP reporter Epac1-camps [63]. This reporter was expressed in MB neurons using the c772-GAL4 driver. Two α-lobe tips and clusters of calix cell bodies were set as a region of interest (ROI), and observed through the head cuticle. Test flies were immobilized on an observation plate by gluing the dorsal portion of the head and neck with nail polish. The observation plate was a large glass coverslip (24×60 mm) attached to a small plastic coverslip (22×22 mm) with a hole (7 mm diameter). The fly thorax was positioned at the edge of the hole so that the fly head was directly attached to the glass coverslip. Confocal images were obtained using a Plan-Neofluar 20× objective on a Zeiss 510 inverted confocal microscope (Zeiss, Oberkochen, Germany). Epac1-camps fluorescence was scanned with a 458 nm Argon ion laser line. YFP-FRET and CFP-donor emissions were separated by means of a NFT545 dichroic mirror and BP475-525 and LP560 emission filters. YFP and CFP signals were scanned simultaneously onto separate photomultiplier tubes, and obtained every 20 seconds. After 3 minutes of baseline FRET (YFP to CFP ratio) measurement, the test fly was fed 20E-sucrose solution or vehicle control for 1 minute using a Kimwipe (10 mm×10 mm) soaked with the solution. The 20E-sucrose solution contained blue food dye (Acid Blue 9, 0.125 mg/ml) as an indicator of ingestion. The effects of 20E on FRET were observed for 30 minutes and analyzed as described by Shafer et al [63]. To compare the FRET time-course among different experiments, the YFP/CFP ratio values were normalized to the value of the first time-point.
|
10.1371/journal.pgen.1002291 | Essential Roles of BCCIP in Mouse Embryonic Development and Structural Stability of Chromosomes | BCCIP is a BRCA2- and CDKN1A(p21)-interacting protein that has been implicated in the maintenance of genomic integrity. To understand the in vivo functions of BCCIP, we generated a conditional BCCIP knockdown transgenic mouse model using Cre-LoxP mediated RNA interference. The BCCIP knockdown embryos displayed impaired cellular proliferation and apoptosis at day E7.5. Consistent with these results, the in vitro proliferation of blastocysts and mouse embryonic fibroblasts (MEFs) of BCCIP knockdown mice were impaired considerably. The BCCIP deficient mouse embryos die before E11.5 day. Deletion of the p53 gene could not rescue the embryonic lethality due to BCCIP deficiency, but partially rescues the growth delay of mouse embryonic fibroblasts in vitro. To further understand the cause of development and proliferation defects in BCCIP-deficient mice, MEFs were subjected to chromosome stability analysis. The BCCIP-deficient MEFs displayed significant spontaneous chromosome structural alterations associated with replication stress, including a 3.5-fold induction of chromatid breaks. Remarkably, the BCCIP-deficient MEFs had a ∼20-fold increase in sister chromatid union (SCU), yet the induction of sister chromatid exchanges (SCE) was modestly at 1.5 fold. SCU is a unique type of chromatid aberration that may give rise to chromatin bridges between daughter nuclei in anaphase. In addition, the BCCIP-deficient MEFs have reduced repair of irradiation-induced DNA damage and reductions of Rad51 protein and nuclear foci. Our data suggest a unique function of BCCIP, not only in repair of DNA damage, but also in resolving stalled replication forks and prevention of replication stress. In addition, BCCIP deficiency causes excessive spontaneous chromatin bridges via the formation of SCU, which can subsequently impair chromosome segregations in mitosis and cell division.
| BCCIP is a BRCA2- and p21-interacting protein. Studies with cell culture systems have suggested an essential role of BCCIP gene in homologous recombination and suppression of replication stress and have suggested that BCCIP defects causes mitotic errors. However, the in vivo function(s) of BCCIP and the mechanistic links between BCCIP's role in suppression of replication stress and mitotic errors are largely unknown. We generated transgenic mouse lines that conditionally express shRNA against the BCCIP, and we found an essential role of BCCIP in embryo development. We demonstrate that BCCIP deficiency causes the formation of a unique type of structural abnormality of chromosomes called sister chromatid union (SCU). It has been noted in the past that impaired homologous recombination and resolution of stalled replication forks can have detrimental consequences in mitosis. However, the physical evidence for this link has not been fully identified. SCU is the product of ligation between sister chromatids, likely formed as a result of unsuccessful attempt(s) to resolve stalled replication forks. Because the SCU will progress into chromatin bridges at anaphase, resulting in mitosis errors, it likely constitutes one of the physical links between S-phase replication stress and mitotic errors.
| Loss of genomic integrity is a hallmark for tumorigenesis. Mammalian cells maintain genomic integrity by ensuring DNA replication fidelity in S-phase, equal chromosome distribution into daughter cells during mitosis, error-free repair of sporadic DNA damage throughout the cell cycle, and a coordinated cell cycle progression [1]. Homologous recombination (HR) plays roles not only in repair of DNA double strand breaks (DSB) but also in replication fidelity [2], [3]. When the replication forks stall during S-phase, one-ended DSBs are produced on one of the sister chromatids at the stalled replication fork. Subsequently, the HR machinery uses the 3′-end of a single-stranded tail of the one-ended DSB to invade the intact double-stranded DNA at the collapsed replication fork, which leads to the resolution of the stalled fork. Failure to do so causes excessive replication stress, which is often defined as the inefficient progression of the replication forks. Replication stress is a status highly susceptible to genomic instability.
The BRCA2 tumor suppressor gene plays critical roles in HR, mainly by mediating RAD51 function [4], [5], including the strand invasion step during the resolution of stalled replication forks. Although mutations of BRCA2 are involved in only a small percentage of human cancers, the germline BRCA2 mutations are of high penetrance in malignant neoplasms. This suggests that the entire molecular network of BRCA2 is critical for cancer prevention, and defects of other proteins related to BRCA2 may contribute to additional tumors [6]. Thus analyses of BRCA2-interacting proteins offers opportunities to identify additional genetic factors involved in tumorigenesis.
BCCIP is a BRCA2- and CDKN1A(p21)- interacting protein [7]–[10]. In human cells, two major isoforms are expressed due to RNA alternative splicing: BCCIPα and BCCIPβ [9]. Although the human BCCIPα isoform was originally identified as a p21 and BRCA2 interacting protein, later studies found that the BCCIPβ isoform also interacts with p21 and BRCA2 [10]–[12]. BCCIP down-regulation has been reported in cancers [9], [13], [14]. Human BCCIP is known to function in HR, G1/S cell cycle checkpoint, and cytokinesis [10]–[12], [15]–[18]. Furthermore, BCCIP deficiency leads to accumulation of spontaneous DNA damage and single-stranded DNA in human cells [16]. The Ustilago maydis homologues of BCCIP and BRCA2 (BCP1 and Brh2) also interact with each other, and BCP1 deficiency causes replication stress [19]. However, the in vivo function of BCCIP has not been determined.
To determine the role of BCCIP in vivo, we established a conditional BCCIP knockdown transgenic mouse model. We show that developmental defects in the BCCIP-deficient embryos occurred before day E6.5, and this was associated with a significant reduction of cell proliferation. In addition to an impaired repair of exogenous DNA damage, BCCIP deficiency significantly induced spontaneous chromatid aberrations that often associate with replication stress. The chromosome abnormalities in BCCIP-deficient mouse cells are characterized by the elevated formation of sister chromatid unions (SCUs) and chromatid breaks, yet a modest increase of sister chromatid exchange (SCE). This suggests an essential role of BCCIP in maintenance of chromatid stability and embryonic development in mice.
Although human cells express two major isoforms (BCCIPα and BCCIPβ) due to alternative RNA splicing [9], mouse tissues appear to express only the BCCIPβ isoform. In the previous studies, human BCCIP has been shown to function in DNA repair, cell cycle regulation, cytokinesis, and maintenance of chromosome stability [7], [10]–[12], [15]–[18]. Reduced or absence of BCCIP expression have been reported in human cancers [9], [13], [14], [20]. To further understand BCCIP's role in development and tumorigenesis, we generated a mouse model with BCCIP deficiency. Similar to the human BCCIP gene structure [9], the mouse uroporphyrinogen III synthase (UROS) is “head-to-head” with the BCCIP gene, and the UROS promoter is located in the intron of the BCCIP gene. The mouse DEAD/H box polypeptide-32 (DDX32) gene is “tail-to-tail” with the BCCIP gene. We adapted the RNAi based conditional knockdown approach developed by Coumoul and colleagues [21]–[23]. Briefly, the U6 promoter that normally drives the expression of short hairpin RNAs (shRNAs) is disrupted by insertion of a LoxPneoLoxP cassette, thus is only functional upon the conditional deletion of the LoxPneoLoxP cassette (Figure 1A). The conditional shRNA expression construct against BCCIP gene was integrated into the mouse genome using standard transgenic mouse techniques. Two founder homozygous transgenic mouse lines with the conditional expression cassette were generated. The two independent homozygous transgenic lines, designated LoxPshBCCIP+/+-4 and LoxPshBCCIP+/+-13, were fertile, grow normally, and have the same lifespan as wild type mice. The LoxPshBCCIP+/+ transgenic mice were crossed with a mouse line expressing Cre recombinase to “pop-out” the LoxPneoLoxP segment. As reported elsewhere [23], the single LoxP site left in the U6 promoter after Cre-recombination does not affect the U6 promoter activity. This reconstitutes the U6 promoter activity, leading to the expression of the anti-BCCIP shRNA (Figure 1A), to achieve a Cre-dependent conditional knockdown of BCCIP.
To verify the BCCIP knockdown, MEFs from the two founder lines were established. As predicted and shown in Figure S1, mouse cells only express one isoform. Expression of Cre in the MEFs derived from both mouse founder lines cells efficiently knocked down BCCIP (Figure S1). These MEF cells, designated MEF4-LoxPshBCCIP and MEF13-LoxPshBCCIP, were used further in vitro studies. It should be pointed out that BCCIP can be knocked down in heterozygous LoxPshBCCIP+/− cells by expression of Cre because one copy of the LoxPshRNA cassette is able to express shRNA against BCCIP.
In an attempt to generate mice with BCCIP knockdown, we bred the FVB/N LoxPshBCCIP+/+-4, and LoxPshBCCIP+/+-13 with the FVB/N EIIaCre+/− mouse [24] that carries a Cre transgene under the control of the adenovirus EIIa promoter. The EIIa promoter drives the expression of Cre recombinase early in embryogenesis [24]. As shown in Table 1, breeding between wild type with EIIaCre+/− mice resulted in approximately 1∶1 ratio of LoxPshBCCIP−/−;EIIaCre+/− and LoxPshBCCIP−/−;EIIaCre−/− mice. However, breeding of LoxPshBCCIP+/+ with EIIaCre+/− mice resulted in a significantly smaller number of LoxPshBCCIP+/−;EIIaCre+/− than LoxPshBCCIP+/−;EIIaCre−/− newborns. In addition, the litter size (5.1 for founder line-4 or 6.6 for founder line-13) from the breeding between LoxPshBCCIP+/+ and EIIaCre+/− was significantly smaller than that of wild type (LoxPshBCCIP−/−) mice (10.3/litter). These data suggest that down-regulation of BCCIP causes embryonic lethality. Although there were significantly less Cre positive mice with this breeding scheme, it was noted that some Cre positive mice were viable. However, further analyses confirmed that some of these viable LoxPshBCCIP+/−;EIIaCre+/− newborn mice had lost the LoxBCCIPshRNA cassette (data not shown). This suggests that the LoxBCCIPshRNA cassette in mice is subject to spontaneous loss.
To confirm that BCCIP knockdown causes embryonic lethality, embryos from crosses between LoxPshBCCIP+/+-4 and EIIaCre+/− were analyzed at day E11.5. As exemplified by Figure 1B–1D, among a total of seven embryos of the same litter, four (labeled as No. 1–4 in Figure 1B) were abnormal and three (labeled as No. 5–7 in Figure 1B) were normal. The abnormal embryos have the EIIaCre-positive genotype, while the normal embryos are EIIaCre-negative (Figure 1C). As expected, the expression of BCCIP in the abnormal embryos was clearly down-regulated, while the healthy embryos expressed normal levels of BCCIP protein (Figure 1D). Analysis of Cre-dependent conditional knockdown embryos derived from another founder line (LoxPshBCCIP+/+-13) is shown in Figure S2. Altogether, these data suggest that down-regulation of BCCIP during embryogenesis causes embryonic lethality prior to day E11.5. Figure 1, Figure S1, and Figure S2 also illustrate that our conditional knockdown strategy indeed achieved the anticipated down-regulation of BCCIP upon expression of Cre-recombinase in the conditional transgenic mice.
Given our observations that BCCIP down-regulation causes developmental arrest at day E11.5, we anticipate the anomaly in embryonic development initiates a few days prior. To define the precise timeframe for the effects of BCCIP knockdown in early embryogenesis, we analyzed embryos at different timepoints, including embryonic days E6.5, E7.5, and E8.5. As shown in Figure 2A–2C, wild type embryos were well developed during this period. By E6.5, wild-type embryos (Figure 2A) displayed normal growth and egg cylinder elongation, extraembryonic and embryonic ectoderm and pro-amniotic cavities. By day E7.5 (Figure 2B), wild-type embryos underwent gastrulation; the amniotic cavity was sealed off and three distinct cavities (amniotic cavity, exocoelom, and ectoplacetal cleft) were well developed. The neural plate, a developed notochord, a confined head and tail folds were visible at day E8.5 in a wild type embryo (Figure 2C). The mid-trunk region remained apparently attached to yolk sac, which is consistent with normal mouse embryo development [25], [26]. However, in the BCCIP knockdown embryos, there was a significantly delayed and abnormally developed embryos as evidenced by the mass size of the embryonic tissues at day E6.5 (Figure 2D). At day E7.5 and E8.5 (Figure 2E and 2F), the BCCIP knockdown embryos were developmentally retarded. There was no apparent formation of amniotic cavity, and no mesoderm differentiation at day E7.5 (Figure 2E). Also, development of the neural plate and notochord was not evident at day E8.5 (Figure 2F). These morphological observations suggest that the developmental defects caused by BCCIP knockdown in the analyzed mouse embryos are likely initiated before day E6.5.
During mouse embryogenesis, mesoderm development occurs around day E6.5. Brachyury can be used as a marker of the primitive streak, nascent mesoderm, the node and notochord [27]–[29]. To confirm that the developmental delay occurs prior to day E6.5, we examined the expression of the Brachyury protein by immunohistochemistry (IHC) at day ∼E6.5. As shown in Figure S3A, the Brachyury expression was readily detectable in the primitive streak and mesoderm in wild-type embryos, which is a sign of mesoderm differentiation (Figure S3A). However, in BCCIP deficient embryos of the same age, little Brachyury expression was detected (Figure S3B). This confirms that the embryonic development retardation in BCCIP deficient mice was likely initiated prior to day E6.5.
As shown in Figure 2 and Figure S3, the BCCIP knockdown embryos display histological development defects around day ∼E6.5. Ki67 expression is commonly regarded as a proliferation marker. To determine whether cellular proliferation is impaired in BCCIP deficient embryos at about the same time, Ki67 expression in embryonic tissues was assessed by IHC (Figure 3A). A proliferative index, defined as the ratio of the number of Ki67-positive nuclei in the embryo tissue preparations over the total nuclei number, was determined (Figure 3B). As shown in Figure 3A and 3B, there was only a slight reduction of Ki67 expression in BCCIP knockdown embryos when compared to wild type embryos at day E6.5. However at day E7.5, the proliferation index was significantly reduced, from ∼80% in wild type to ∼11% in the BCCIP knockdown embryos (Figure 3B).
To confirm the cell proliferation assessment data, incorporation of 5-bromo-2′-deoxyuridine (BrdU) into DNA during the S phase of the cell cycle was measured at days E6.5 and E7.5. As shown in Figure 3C and 3D, there was little difference in labeling index at day E6.5 between wild type and BCCIP knockdown embryos. However at day E7.5, wild type embryos had 52% BrdU-positive nuclear staining compared to 10% in BCCIP knockdown embryos (Figure 3D). These results strongly suggest that the proliferation defects of BCCIP deficient embryos are initiated by day E6.5, consistent with the data from histological analyses (Figure 2 and Figure S3).
To determine if the growth defect of BCCIP deficient embryos is associated with an excessive level of programmed cell death, embryo serial tissue sections at days E6.5 and E7.5 were analyzed by terminal deoxynucleotidyl transferase (TdT)-mediated dUTP nick end labeling assay (TUNEL) and anti-cleaved caspase-3 staining. At day E6.5, there was little apoptotic and caspase-3-positive cells in the wild type and the BCCIP deficient embryos (Figure 4). However, at day E7.5, clear apoptotic signals were detected in BCCIP deficient but not in wild type embryos (Figure 4). This indicates that programmed cell death in BCCIP knockdown embryos is increased as early as day E7.5, which is in strong agreement with the impaired embryo development around this time as shown in Figure 2.
In early mouse embryogenesis, prior to the implantation, the inner cell mass (ICM) inside the blastocysts forms one of the earliest structures of embryos, and eventually give rise to the definitive structures of the embryo. In vitro Blastocyst outgrowth offers an opportunity to observe ICM growth and to assess the early post-implantational development. To assess the role of BCCIP in embryonic development prior to day E6.5, LoxPshBCCIP+/+ mice were bred with EIIaCre+/+ mice (breeding between LoxPshBCCIP+/+ and with EIIaCre−/− as the control). Blastocysts were collected by uterine flushing at day E3.5, and cultured in vitro. The growth of ICM from the blastocysts was monitored daily while in culture. The numbers of blastocysts analyzed are summarized in Table S1. Among 71 BCCIP knockdown blastocysts, 28 (or 39%) successfully attached to the culture dish, which was a slightly lower frequency than the control blastocysts (28/58, or 48%). For the attached blastocysts, there was little morphological difference between control and BCCIP deficient blastocysts after one day in culture (equivalent to day E4.5 in vivo). Figure 5A illustrates the representative growth morphology of the blastocysts in culture at days 2, 4, and 5. Normally, the blastocysts hatch from the zona pellucida around day 1 to 2 in culture. As shown in Figure 5A, there was little apparent morphological difference at day 2 shortly after blastocysts hatching in vitro. After day 2 in culture, growth of the ICM from the BCCIP deficient blastocysts was clearly defective, although the difference in trophoblast giant cell growth between control and BCCIP deficient cells appears to be less significant (Figure 5A). To quantify the growth of the ICM in vitro, the relative areas of ICM were calculated using the ImageJ program. As shown in Figure 5B, the growth of BCCIP deficient ICM was significantly impaired when compared with wild type blastocysts starting at day 3 (equivalent to day 6.5 in vivo) in culture. These results imply that BCCIP defects affect the ICM growth, which is consistent with the in vivo observation of growth retardation in BCCIP knockdown embryos as described in Figure 2.
The in vivo and in vitro data above have shown growth retardation in BCCIP knockdown embryos, suggesting that the mouse BCCIP is essential for cell proliferation and growth. To investigate the underlying mechanism(s), we used the MEF4-LoxPshBCCIP cells. At passage 1, the MEF4-LoxPshBCCIP cells were infected with retroviruses expressing Cre-recombinase to reconstitute the functional U6 promoter in order to achieve BCCIP knockdown. The control groups were infected with retrovirus expressing the YFP. As shown in Figure 6A, the MEF4-LoxPshBCCIP cells infected with Cre-virus grew slower than those infected with YFP expressing virus (Control). This slowed growth of BCCIP knockdown MEF cells is coincident with a reduced level of PCNA (a proliferation marker) and increased level of p21 (Figure 6B). We also observed an increase of Ser-15-phosphorylated p53 in the BCCIP knockdown MEFs (Figure 6B), suggesting a spontaneous activation of DNA damage signaling in the BCCIP deficient cells.
We further assessed the roles of mouse BCCIP in DNA damage sensitivity. Because of poor colony formation by the primary MEF culture, a clonogenic survival assay was technically infeasible. Thus, we performed growth inhibition assay to assess the MEF's response to modest dose of irradiation. As shown in Figure 7A, BCCIP knockdown cells exhibited greater growth inhibition by irradiation compared to control MEFs. Irradiation with 1–4 Gy of γ-rays showed a similar trend of inhibition of cell growth (Figure S4).
Under physiological conditions, without exogenous DNA damage, HR is thought to play a major role in relieving replication stress. We treated the MEF cells with low concentrations of alphidicolin (APH). After washing off the APH, cells were immediately incubated with Bromodeoxyuridine (BrdU) in APH-free medium. At various time points, the fraction of cells with BrdU incorporation was scored after immunofluorescent staining (see Materials and Methods), which reflects the recovery from replication blockage. As show in Figure 7B, when normalized to the un-treated cells, the re-incorporation of BrdU was less efficient among BCCIP knockdown cells than the control MEFs, reflecting a delayed recovery from replication stress. Figure 7C shows representative fields of BrdU-labeled cells. Together, these data suggest that the BCCIP deficient cells are not only more sensitive to DNA damage but also less efficient to recover from replication stress than control cells.
To directly assess the DSB repair capability, we measured the kinetics of γH2AX removal following irradiation. As shown in Figure 8A, 15 min after irradiation, all cells have a similar level of γH2AX. However, at 4 and 8 hours after irradiation, the BCCIP knockdown MEFs have significantly more γH2AX nuclear foci than the control MEF cells. Similarly, the fractions of cells with 5 or more γH2AX foci were higher in the BCCIP knockdown cells than the controls (Figure 8B). Figure S5 shows representative γH2AX foci at different times after irradiation. These observations indicate that the control cells remove DSBs more efficiently than BCCIP knockdown MEFs, and that down regulation of BCCIP impairs DSB repair after irradiation. In addition, an alkaline comet assay revealed more residual DNA damage at 4 hours after irradiation in the BCCIP deficient MEFs when compared to control cells (Figure 8C and 8D). These data strongly suggest an impaired repair capability in the BCCIP deficient cells, consistent with the slower growth of the BCCIP knockdown MEFs after irradiation (Figure 7A).
Because human RAD51 focus formation is associated with BCCIP [10], [30], we further assessed the potential role of BCCIP in mouse Rad51 response to radiation. As shown in Figure 9A, BCCIP deficiency resulted in a significant reduction of Rad51 foci in response to radiation. Furthermore, there was a reduction of Rad51 protein level in BCCIP deficient cells compared to control cells (Figure 9C). This is consistent with a role of BCCIP in HR dependent DSB repair.
The human BCCIP interacts with BRCA2, and BCCIP deficiency reduces endogenous level of Rad51 (Figure 9). Both BRCA2 and Rad51 are key proteins involved in HR. Under physiological condition, a key function of HR is to resolve stalled replication forks [2], [3], and impaired HR would cause spontaneous structural chromosome alterations. Thus, we investigated whether BCCIP deficiency would cause spontaneous chromosome abnormalities. First, Giemsa-stained chromosome metaphase spreads were prepared from control and BCCIP-deficient MEFs. As represented in Figure 10A–10B, we observed two types of spontaneous chromatid aberrations in BCCIP knockdown MEFs: single chromatid breaks with un-paired chromatid fragments, and SCU (sister chromatid union). We also observed some paired sister chromatid fragments (pSCF) that may be companions with SCU (when the SCU is formed by fusion of telomere-less broken chromatid arms). Figure 10G summarizes the spontaneous frequencies of the types of chromosome abnormalities caused by BCCIP-deficiency. BCCIP-deficiency results in a 3.5-fold increase on single chromatid breaks, and 3.4-fold increase in occurrence of paired sister chromatid fragments. The most dramatic increase is in SCU occurrence. While there was little SCU in the control cells, there was ∼20-fold increase of SCUs in BCCIP knockdown cells. Because the BCCIP knockdown MEF population has more polyploid cells than control MEF, the frequencies of chromosome abnormalities were normalized to the number of chromosomes. The frequency of abnormality, normalized to number of metaphase cells, can be found in Figure S6.
The chromatid break is indicative of failed restart of collapsed replication forks, which generates one-ended DSBs. We further measured whether BCCIP deficiency causes increased sister chromatid exchange (SCE), which is seen in Bloom syndrome and several genetic disorder related to replication stress [31]–[33]. Consistent with the results from Giemsa-stained chromosome metaphase spreads, we observed the induction of SCUs alone with chromatid breaks in BCCIP-deficient cells (Figure 10C and 10D). However, the increase of SCE in BCCIP deficient cells is modest (Figure 10H). This may reflect a potential role of BCCIP in supporting Rad51-dependent strand invasion during the restart of replication forks (see Discussion for details), which is consistent with the observation that BCCIP deficiency causes Rad51 down regulation (Figure 9).
The formation of SCUs is a unique phenotype in BCCIP deficient cells. To our best knowledge, the earliest literature that described this form of chromatid alteration was in 1938 with Drosophila by Kaufmann [34], but has been rarely described since then. We reasoned that SCUs may be produced by two mechanisms: telomere fusion between the sister chromatids; or the re-ligation of the broken sister chromatids. Although the second possibility is suggested by the presence of paired chromatid fragments in the BCCIP deficient cells, telomere FISH was performed to distinguish these possibilities. As can be seen in Figure 10E and 10F, the SCUs were associated with loss of telomere signals and were not caused by telomere fusion. We often observed paired telomere signals from the acentromeric chromatid fragments in the same cells with SCUs. These observations suggest SCU as a consequence of ligation of two broken telomere-less sister chromatids, and unlikely a fusion after telomere erosion. With the same telomere FISH experiments, we observed induction of chromatid breaks with single sister chromatid fragment (sSCF) in BCCIP deficient cells (Figure 10F). We also found an increase in percentage of chromatids that have lost telomere FISH signals (Figure 10I). Altogether, these data (Figure 10) strongly suggest that BCCIP deficiency causes spontaneous chromatid aberrations associated with replication.
We further analyzed chromosome abnormalities at 2 and 8 hours after 2 Gy of γ-irradiation. Again, there was a significant increase of spontaneous chromosome abnormalities, including SCU (Figure 11). At 2 hours after the irradiation, SCU frequency is significantly higher in the BCCIP knockdown cells than the control cells, but the other forms of damages are not significantly different between the BCCIP knockdown and control cells. The control cells exhibited significantly less chromosome abnormalities at 8 hours, than at 2 hours after irradiation, indicating repair of DNA damages associated with these forms of abnormalities. However, there remained a significantly higher level of chromosome abnormalities in the BCCIP knockdown cells than the control cells at 8 hours. Noticeably, the SCU level at 8 hours remains as high as at 2 hours, suggesting that the BCCIP deficient cells repaired little damages leading to SCU during the 2–8 hours following irradiation. These data support the notion that BCCIP is not only required to repair different forms of DNA damages but also has a significant role in protecting the cells against SCU.
Because BCCIP deficiency spontaneously activates p53 Ser-15 phosphorylation in the MEFs (Figure 6B), and it has been shown that p53 deficiency can partially delay the embryonic lethality conveyed by BRCA1 and BRCA2 deficiency in mice [35], we measured the in vitro growth rates of the p53 mutant and wild type BCCIP deficient MEFs. As shown in Figure S7, deletion of p53 can only partially rescue the growth retardation of BCCIP deficient cells. Next we asked whether p53 deficiency can completely rescue the embryonic lethality in BCCIP deficient mice. We used three different strategies to breed the constitutive p53 null mice originally generated by Jacks et al [36]. Table 2 shows the distribution of genotypes among viable newborns after breeding: 1) between (p53+/+;LoxPshBCCIP+/+;EIIaCre−/−) and (p53+/+; LoxPshBCCIP−/−;EIIaCre+/−), 2) between (p53+/−; LoxPshBCCIP +/+;EIIaCre−/−) and (p53+/−; LoxPshBCCIP −/−;EIIaCre+/−), and between (p53+/−; LoxPshBCCIP +/−;EIIaCre−/−) and (p53+/−; LoxPshBCCIP −/−;EIIaCre+/−). As shown in Table 2 (see Table S2 for detailed breeding data), p53 deletion retain approximate 1∶1 ratio between EIIaCre(+/−) and EIIaCre(−/−)mice in LoxPshBCCIP(−/−) background. The EIIaCre(+/−) and EIIaCre(−/−) ratio was significantly less than 1 (16∶86) in p53 wild type mice, and this reduced ratio was not increased in p53 deficient or p53 heterozygous background (0∶24 and 11∶72 respectively). These data suggest that p53 deletion failed to completely rescue the embryonic lethality induced by BCCIP knockdown, suggesting that DNA damage activated p53 signaling cannot fully account for the embryonic death completely.
BCCIP is a BRCA2 interacting protein in human and Ustilago maydis [7]–[10], [19]. In this study, we have found that BCCIP deficiency causes chromatid abnormalities especially a dramatic induction of sister chromatid unions (SCUs), and impairs mouse embryo development.
In addition to DSB repair, a major function of the HR machinery is to preserve genomic integrity via resolving replication blockage to reduce replication stress, which is loosely defined as the inefficient progression or stalling of replication forks [3], [37], [38]. During replication, replication forks may be stalled by encountering single-strand breaks or damaged nucleotides that are not by-passed by DNA translesion synthesis. This often produces a one-ended DSB, which can be processed to yield a single stranded 3′-end to initiate a strand invasion and form a single Holliday junction at the stalled replication fork. After branch migration (or replication fork regression) and resolution of the Holliday junction, the stalled replication fork can be re-started. It is believed that many factors of the HR pathway, including BRCA2 and associated proteins, are required in this process.
Replication stress is often manifested by excessive levels of spontaneous single-stranded DNA (ssDNA), or DNA strand breaks. On the cytogenetic level, excessive level of chromatid breaks and SCEs is a signature of replication stress. It has been suggested that endogenous replication stress induced by HR defects may not be detected by the S-phase checkpoint machinery. Thus cells with excessive replication stress can enter mitosis to cause mitotic errors [37]. In a previous report, it was shown that BCCIP deficiency results in accumulation of spontaneous DNA strand breaks and single-stranded DNA in human cells [16]. In this study, we have observed an increase of spontaneous chromatid breaks and SCUs in BCCIP deficient cells (Figure 10), and impaired repair of radiation damages that lead to SCU formation (Figure 11). These results are consistent with a role of BCCIP in suppressing replication stress and repair of DNA damage.
We have observed a significant spontaneous increase in sister chromatid breaks (3.5-fold) yet a modest increase of SCE (∼1.5 fold) in BCCIP deficient cells (Figure 10). These abnormalities have often been used as markers for genomic instability. Although the molecular mechanisms for SCE formation are complex, it is generally believed that the 3′-end of the one-ended DSB of the stalled replication fork initiates the process with strand invasion [31]. Once strand invasion is initiated to form the Holliday junction, branch migration (or fork regression) and resolution of the Holliday junction would produce a SCE (Figure 12A). Therefore, factors that increase the production of one-ended DSBs (e.g. excessive levels of SSBs or inability to carry out translesion synthesis) have the potential to stimulate SCE [31]. Additionally, deficiencies in proteins involved in branch migration of the Holliday junction (e.g. BLM and RecQL5) may favor SCE upon Holliday Junction resolution [31]–[33].
On the other hand, defects in proteins involved in strand invasion may have different consequences on SCE. It has been shown that BRCA2 and RAD51 defects do not significantly increase spontaneous SCE in mammalian cells [31], [39]–[41]. Since strand invasion is a critical step to produce SCE, defective RAD51 and its accessory factors may reduce or only modestly increase SCE due to ineffective strand invasion even in the context of excessive one-ended DSB and replication stress. As a consequence, this would significantly increase chromatid breaks (Figure 12B). In this study, we observed reduced basal level of mouse Rad51 protein and focus formation in the BCCIP deficient cells (Figure 9). This observation is consistent with the increase in chromatid breaks and formation of paired sister chromatids in BCCIP-deficient cells and the reduction of RAD51 focus formation in BCCIP deficient human cells [10], [30]. A possible mechanism for the formation single chromatid breaks is illustrated in Figure 12B.
A question is how BCCIP deficiency may cause reduced Rad51 protein level. It is known that Rad51 preferably expresses in S phase cells. A tempting explanation is that reduction of S-phase cell fraction may reduce the overall Rad51 level in the BCCIP deficient cell population. However, this is unlikely the case because BCCIP deficiency causes replication stress but did not cause overall reduction of S-phase fraction ([11] and data not shown). Although we cannot rule out the possibility that Rad51 protein stability is altered in BCCIP deficient MEFs, we found that the BCCIP deficient MEF cells had reduced Rad51 mRNA level based on RT-PCR analysis (data not shown), suggesting that a down-regulated Rad51 transcription may contribute to the Rad51 protein level.
A characteristic structural chromosome alteration in the BCCIP deficient MEFs is SCU, which is not only induced spontaneously but also remains high 2–8 hours after irradiation (Figure 10 and Figure 11). The telomere FISH experiments have suggested that SCUs are likely the consequence of ligation between two broken sister chromatids. We envision that SCU may be caused by the following scenario (Figure 12C). When one-ended DSB resection and subsequent strand invasion fails, an excessive level of single sister chromatid breaks and further collapse of the replication fork result in three one-ended DSBs (as shown in Figure 12C). Then, SCU may occur upon re-ligation of the sister chromatid DSB ends. The proximate DNA fragment may resume replication due to the presence of multiple replication origin sites. This produces paired chromatid fragments. However, we would like to emphasize that alternative mechanisms to produce SCU are possible. For example, late S-phase cells with failed resolution of HR intermediates and/or replication termination structures may form DSB on sister chromatids, thus SCU. Although erosion of telomeres in telomerase deficient cells may expose the chromtid ends to form SCU, this scenario is unlikely to be the cause of SCU in BCCIP deficient MEF cells, because the BCCIP deficient cells were cultured in vitro for only a few passages. It would be interesting to investigate whether eroded telomere ends can form SCU in Tert deficient cells after long-term culture.
Nevertheless, the SCUs will likely form chromatin bridges between daughter nuclei at anaphase. It is expected that this form of structural abnormality will result in chromosome segregation errors and numerical chromosome instability in daughter cells.
The phenotypes of BCCIP deficient embryos are consistent with BCCIP's orthologs in lower eukaryotes, and its interaction partner BRCA2 [35]. Several BRCA2 knockout mouse models have been developed. Depending on the specific regions deleted in the knockout model, the embryonic phenotype of BRCA2 mutant mice varies [35]. However, most mouse models with large deletions on BRCA2 produce embryonic lethality [35]. In this study, we have established a LoxP-Cre based conditional BCCIP knockdown mouse model. Using this model, we have shown that the mouse BCCIP gene is essential for embryonic development. Although many mechanisms may contribute to embryonic abnormality of BCCIP-deficient mice, the accumulation of spontaneous DNA damage, excessive replication stress, and formation of lethal chromatid aberrations in BCCIP deficient cells are considered the major initiating factors. Down regulation of BCCIP has been shown to cause spontaneous DNA damage in human cells [16]. In this study, we observed spontaneous activation of p53 together with up-regulation of p21 in BCCIP deficient MEFs (Figure 6B), and increased cell death through apoptosis at day E7.5 follows reduction in cell proliferation (Figure 5). These observations are consistent with the scenario that BCCIP deficiency leads to accumulation of spontaneous DNA damage, thus growth inhibition and cell death, which lead to embryonic lethality. However, the accumulation of spontaneous DNA damage along with activation of p53 may not fully account for embryo lethality, as the p53 deletion did not completely rescue the embryonic lethality of BCCIP deficiency despite that it can partially rescue the growth retardation of BCCIP deficient MEFs in vitro (Figure S7). Second, BCCIP deficiency may inhibit proliferation by disrupting cell division in mitosis. We found that BCCIP-deficient cells had significantly increased levels of spontaneous chromatid breaks and SCUs at metaphase (Figure 10). It is anticipated that the chromatid breaks will cause a net loss of chromosomal materials after mitosis, and the SCUs will evolve into chromatid bridges at anaphase and telophase to disrupt chromosome segregation, both scenarios are potentially lethal to the cells and can contribute to the embryo development defects.
The conditional knockdown approach offers advantages over conventional knockout approach. It may avoid interference with the overlapping genes. Second, while the conventional knockout approach would only offer either homozygous or heterozygous gene ablation, the knockdown approach may grant us the ability of mimicking abnormal protein expression that might occur in human diseases. Down regulation of BCCIP has been shown in cancers [7], [9], [13], [14]. Considering the strong genomic instability phenotype in BCCIP deficient cells and the multiple functions of BCCIP [10], [11], [15]–[18]. It is likely that BCCIP deficiency may contribute to tumorigenesis in mice. Because the EIIa-Cre mediated BCCIP knockdown causes embryonic lethality, tissue specific conditional knockdown is in process to address whether BCCIP down-regulation contribute to tumorigenesis.
In summary, our study suggests a critical role of mouse BCCIP gene in maintaining genomic stability and embryonic development. The formation of characteristic sister chromatid union in BCCIP deficient cells may reflect a unique molecular function of BCCIP in resolving stalled replication forks, and may contribute significantly to embryonic development defects.
The animal works presented in this study were approved by Institutional Animal Use and Care Committee of Robert Wood Johnson Medical School-UMDNJ. We follow our institutional guideline regarding to animal welfare issues.
The pBS/U6-pLoxPneo vector [23] was kindly provided by Dr. Chuxia Deng (National Institute of Diabetes and Digestive and Kidney Disease, NIH). A pair of mouse BCCIP specific oligonucleotide (5′-GGATGAAGATGAGATCTTTGGTTCAAGAGACCAAAGATCTCATC TTCATCCTTTTTT-3′ and 5′-AATTAAAAAAGGATGAAGATGAGATCTTTGGTCTCTTGAACCAAAGATCTCATCTTCATCCGGCC-3′) were annealed, and then ligated into the pBS/U6-pLoxPneo vector digested with ApaI and EcoRI. This results in the conditional mouse BCCIP knockdown vector designated pBS/U6-pLoxPneo-shBCCIP. The effectiveness of this vector to knockdown mouse BCCIP was confirmed by stably transfecting vectors into mouse NIH3T3 cells, and then transiently expressing the Cre recombinase in the cells.
The conditional BCCIP knockdown vector (pBS/U6-pLoxPneo-shBCCIP) was digested by KpnI and NotI. The linearized 2.3 kb DNA fragment containing the conditional LoxPshRNA expression cassette was injected into pronuclei of fertilized oocytes isolated from superovulated FVB/N mice. Then the injected oocytes were implanted into pseudopregnant recipient females. Genomic DNA was extracted from tail biopsies of the resulting litters and analyzed by PCR and Southern blot. Among the 27 mice obtained from the injections, 7 were found to be positive for the LoxPshBCCIP transgene cassette. The U6-LoxP-shBCCIP positive mice were crossbred with FVB/N wild type mice to identify the mouse lines capable of germline transmission. Through this procedure, two founder lines with high germline transmission were identified. They were designated as LoxPshBCCIP-4 and LoxPshBCCIP-13, and both were successfully bred into homozygsity (LoxPshBCCIP+/+). By breeding with wild type mice, the homozygous transgenic mice (LoxPshBCCIP+/+) is distinguished from heterozygous mice (LoxPshBCCIP+/−) because the homozygous transgenic mice are able to produce 100% of LoxPshBCCIP positive newborns while the heterozygous mice (LoxPshBCCIP+/−) can produce only 50% of LoxPshBCCIP positive mice. The PCR primer pairs used for genotyping were: 5′-TCTAGAACTAGTGGATCCGAC -3′, and 5′-TCGTATAGCATACATTATACG-3′. The probe used for Southern blot was generated by a PCR amplification of the conditional knockdown vector using the following primers: 5′-ATTGAACAAGATGGATTGCACGCA, and 5′-TCAGAAGAACTCGTCAAG AAGG-3′.
The homozygous FVB/N-Tg (EIIaCre+/+) C5379Lmgd/J mice (Lakso, 1996), were purchased from Jackson Laboratory (stock number: 003724), and crossed with wild type FVB/N mice to obtain EIIaCre+/− heterozygous mice. Then the LoxPshBCCIP+/+ homozygous mice were bred with the EIIa-Cre+/− heterozygous mice. Theoretically, this will generate offspring with two genotypes: [LoxPshBCCIP+/−;EIIaCre+/−] with BCCIP knockdown, and [LoxPshBCCIP+/−; EIIaCre−/−] as a control at a 1∶1 ratio. If the knockdown of BCCIP is lethal during embryogenesis, reduced newborn ratio of [LoxPshBCCIP+/−; EIIaCre+/−] to [LoxPshBCCIP+/−; EIIaCre−/−] is anticipated. The PCR primers to genotype EIIaCre were 5′CCTGTTTTGCACGTTCACCG3′ and 5′ATGCTTCTGTCCGTTTGCCG3′, which results in a PCR product of ∼270 bp. The animal works were approved by Institutional Animal Use and Care Committee of Robert Wood Johnson Medical School.
To generate rabbit anti-mouse BCCIP antibodies, mouse cDNA coding for C-terminal 292aa was cloned into pET28 vector (Novagen, Madison, WI). Recombinant (6×His)-tagged mouse BCCIP protein was expressed and purified with BL21 (DE3) cells, and the GST-mouse BCCIP protein was expressed and purified in BL21 cells using pGEX vector as previously described [42], [43]. The HIS-tagged BCCIP was injected into rabbits to produce polyclonal antibodies, and GST-mouse BCCIP was used for affinity purification of polyclonal anti-BCCIP antibodies. Anti- PCNA (PC-10), p21 (F-5), p53 (FL-393), and c-myc monoclonal antibody were purchased from Santa Cruz Biotechnology (Santa Cruz, CA). Anti- γH2AX, phospho-p53 (Ser15) and anti-cleaved caspase-3 antibodies from Cell Signaling (Danvers, MA); anti-pericentrin antibody from Covance Research Products Inc (Berkeley, CA), anti-γ tubulin antibody from Sigma (St, Louis, MO), and anti-Brachyury and anti-Ki67 from Abcam (Cambridge, MA). Western blots were performed with procedures as described previously [7], [11], [15], [16], [18], [44].
Uteri from female mice were isolated at days E6.5–8.5, the individual decidual swellings were isolated transversely according to the methods of Smith (Smith, 1985), rinsed with cold PBS, fixed overnight in 4% paraformaldehyde at 4°C, then embedded in paraffin. Serials of 5 µm sections were cut and stained with hemotoxylin and eosin. Anti-BCCIP polyclonal antibody (1∶100), anti-Ki67 polyclonal antibody (1∶300), anti-cleaved caspase3 polyclonal antibody (1∶100), and anti-Brachyury (1∶100) antibodies were used for immuno-histochemical staining of the corresponding proteins using previously developed protocols [20].
To measure DNA synthesis in embryo mouse tissues, BrdU (100 µg/g of body weight) was intraperitoneally injected into pregnant female mice. One hour later, the entire uteri were removed, and the individual decidual swellings were isolated, fixed in 4% paraformaldehyde at 4°C overnight, embedded in paraffin, and sectioned (5 µm). To stain incorporated BrdU, the sections were de-paraffinized, treated with 2 N HCl for 30 min at 37°C, incubated with anti-BrdU monoclonal antibody (Becton Dickinson, Franklin Lakes, NJ) at a 1∶500 dilution for 2 hr at 37°C, and then incubated with anti-mouse-HRP secondary antibody for 1 hr. 3,3′-Diaminobenzine tetrahydrochloride hydrate (DAB) color developed. BrdU positive cells are visualized by their brown color with DAB, and BrdU negative cells display blue color by hematoxylin.
Paraffin embedded tissue sections (5 µm) were used to detect apoptotic cells using DeadEnd Fluorometric TUNEL System (Promega, Madison, WI). Briefly, sections were rinsed 3 times with distilled H2O, once in PBS, and permeabilized with 200 µg/ml of Proteinase K in PBS for 15 min. The permeabilized sections were incubated with equilibration buffer for 10 min at room temperature. DNA strand-break labeling and colorization were performed according to the manufacturer recommended procedures, mounted with VECTASHIELD fluorescent mounting media with DAPI, and the results were recorded with fluorescent microscope.
Homozygous BCCIP female FVB/NJ mice (3.5–4 week old) were given 5 IU of pregnant mare's serum gonadotropin by intraperitoneal injection between 3–4 pm. At 46–48 h post-PMSG injection, they were treated with 5.0 IU of human chorionic gonadotrophin (hCG) by intraperitoneal injection, and then mated with wild type and EIIaCre+/+ male mice individually. Next morning, the female mice with positive mating-plugs were separated from male mice. At embryo day 3.5 (E3.5), blastocysts were collected by flushing the uteri of female mice, and individually cultured for 5 days in 24-well plates in ES cell culture media without leukemia inhibitory factor (Liu, 1996, Suzuki, 1997) with 5% CO2 at 37°C. The growth of the cultured blastocysts was monitored daily and photographed.
Primary mouse embryo fibroblast (MEF) cells were generated from day 13.5–14.5 embryos of LoxPshBCCIP+/+ female mice (founder line 4) mated with LoxPshBCCIP+/+ homozygous male mice according to the protocols by Hertzog [45]. The MEF cells were counted and plated into 10 cm dishes at a density of 0.5–1×105 per cm2 in DMEM medium containing 10% FBS, and incubated at 37°C with 5% CO2. After 24 hr, the medium, cellular debris, and any unattached cells were removed. The attached MEF cells were designated as passage 0. After 2–3 days of culture, each 10 cm plate of cells was split into 3–5 of 10 cm plates, and the split cultures were then designated as MEF cell passage 1. All in vitro experiments, except specifically noted, were carried out with the first passage MEF cells.
A retroviral packaging cell line specific for mouse cell lines (φEco), and mouse embryo fibroblast cells (MEF), were cultured in DMEM medium supplement with 10% fetal bovine serum, 100 U/ml of penicillin, 100 µg/ml of streptomycin, and 1% of glutamine. The φEco cells were transfected with pLXSP-YFP and pLXSP-myc-Cre retrovirus vector separately. Forty eight hours after transfection, transfected cells were selected by puromycin (1 µg/ml) for 2 days. Then the cells were grown to 80–90% confluence in regular culture medium. Virus suspensions were collected, filtered with 0.45 sterile syringe filters, and mixed with 8 µg/ml of polybrene (Sigma, St, Louis, MO). The MEF cells were infected 3-times with the virus during a 2 day period, and then selected with 2.5 µg/ml puromycin for 2 days prior to phenotype analyses.
For cell growth analysis, cells were counted using a Coulter counter (Beckman Coulter, Fullerton, CA). Cells were initially seeded onto 6 cm dish at a density of 0.1×106 per dish, then cell number was determined daily for the next 5 days after the initial plating. Triplicates for each group at each time point were used in the measurements.
To assess the ability of MEFs to recover from replication blockage, 0.1×106 MEF cells were grown on 18 mm cover slides in 6-well plate. Cells were treated with 0.4 µM APH in DMEM media for 37°C for 30 hours. After removing APH containing media by rinsing with sterile PBS, the recovery of replication of was measured by measuring Bromodeoxyuridine (BrdU) incorporation. Briefly, BrdU was added to each well to a concentration of 10 µM and slides were fixed at 0, 2, 3, 4, and 5 hours after adding BrdU using 4% paraformaldehyde for 10 minutes at room temperature. The fixative was removed by washing the cover slides three times with 1× PBS. The slides in the wells were treated with 1 M HCl for 10 min in ice, 10 min at room temperature, and 40 min at 37°C to denature DNA. Acid was removed and neutralized by washing the cover slides three times with borate buffer (pH 8.5). Cover slides were then washed three times in PBS+ 0.05% Tween 20 [PBS/T20], blocked with 1 ml of PBS/T20/2% normal goat serum at 37°C for 30 minutes. Cells were immuno- stained with mouse anti-BrdU antibody (1∶200 in 0.1 ml of PBS/T20/2% normal goat serum) and incubating at room temperature for 1 hour or at 4°C overnight. The cells were washed three times with PBS+ 0.05% Tween-20 and stained with donkey anti-mouse Rhodamine conjugate diluted to 1∶500 in 0.1 ml PBS+ 0.05% Tween-20 with 3% BSA and incubated at room temperature for 1 hour. Cover-slides were washed three times with PBS/T20, and mounted on glass slides using Vecta shield+DAPI mounting media. Slides were evaluated using immunofluorescent microscopy and the percentage of BrdU positive cells were counted by counting BrdU positive and DAPI stained cells on each slide.
To prepare metaphase chromosome spreads (Brown, 2000; Ko, 2008), cells at 80–90% confluence were subcultured into fresh medium, and incubated at 37°C for 24 hours. Colcemid (Sigma, St. Louis, MO) was added at final concentration of 0.2 µg/ml and incubated at 37°C for 4 hours. Cells were trypsinized, and suspended in 75 mM KCl hypotonic solution at 37°C for 15 minutes, and then fixed in fresh 3∶1 methanol/acetic acid. After 3–4 times of additional fixation, suspended cells were dropped onto cold wet slides, allowed to dry at room temperature, and stained with 1% Giemsa. At least 50 metaphase cells were analyzed under 1000× magnification with microscope for each group. Gross chromosome aberrations were scored. Statistical analyses for frequency of aberrations were performed using t-test, and a P value of <0.05 was considered significant.
The methods developed by Williams et al were used [46], [47]. In brief, after the MEF cells were subcultured into fresh medium and cultured for 24 hours, 0.2 µg/ml Colcemid was added for 6 hours to accumulate mitotic cells. Cells were trypsinized with Trypsin-EDTA (Gibco, Carlsbad, CA) and suspended in 75 mM KCl hypotonic solution at 37°C for 15 minutes before fixation. After four times of repeated fixation in fresh 3∶1 methanol/acetic acid, cells were dropped onto cold slides and allowed to dry slowly in a humid slide box. A probe to telomeric DNA was prepared by synthesizing an oligomer having the sequence (CCCTAA)7 and was labeled by terminal deoxynucleotidal transferase tailing (Roche, Florence, SC) with SpectrumRed-dUTP (Vysis, Des Plaines, IL) according to the manufacturer's instructions. A hybridization mixture containing 0.4 µg/ml probe DNA in 30% formamide and 2×SSC (1×SSC is 0.15 M NaCl, 0.015 M sodium citrate) was applied to slides that had been denatured in 70% formamide, and 2× SSC at 70°C for five minutes. Following an overnight hybridization at 37°C in a moist chamber, the slides were washed in 2× SSC at 42°C (5 times, 15 min each) twice, and then placed in PN Buffer (100 mM Na2HPO4, 50 mM NaH2PO4, 0.1%Triton X-100) at room temperature for 5 minutes and mounted in fluorescence mounting medium with DAPI. Metaphase cells examined with a Zeiss fluorescence microscope and images were captured with HAL100 camera. At least 20 metaphase cells were analyzed for each group. Chromosome aberrations (breaks, fragments, and sister chromatid union) were scored. Statistical analyses for frequency of aberrations were performed using t-test.
The method developed by Wang et al. was modified [48]. The MEF cells were cultured with medium containing 10 µM bromodeoxyuridine (BrdU) for 24 hour and then cultured in growth medium for another 24 hour, and treated with 0.2 µg/ml of colcemid 6 h before collection. The harvested cells were treated with 75 mM KCl hypotonic solution at 37°C for 15 minutes, and fixed with fresh 3∶1 methanol/acetic acid. The cell suspension were dropped onto slides and air-dried. The slides were incubated with 10 µg/ml Hoechst 33258 in ddH2O for 20 min, and rinsed with MacIlvaine solution (164 mM Na2HPO4, 16 mM citric acid pH 7.0) for three times. The slides were mounted in MacIlvaine solution, and exposed to UV light for 45 min. After washing with PBS for 3 times, the slides were incubated in 2× SSC (0.3 M NaCl, 0.03 M sodium citrate) solution at 62°C for 1 hour, and then stained with 1% Giemsa solution at pH 6.8 for 20 min. Metaphase cells were examined with Olympus microscope and images were captured with PictureFrame. At least 20 metaphase cells were analyzed for each group. Statistic analyses for frequency of aberrations were performed using t-test.
The heterozygous p53 knockout mice [36] were crossed with LoxPshBCCIP+/+-4 mouse and EIIaCre+/+ mouse respectively to generate p53+/−;LoxPshBCCIP+/−, and p53+/−;EIIaCre+/− mice. The PCR primers used to genotype p53 are: p53ex6F: 5′-GTATCCCGAGTATCTGGAAGACAG-3′, p53neoF: 5′-GCCTTCTATCGCCTTCTTGACG-3′, p53ex7RN: 5′-AAGGATAGGTCGGCGGTTCATGC-3′. The same PCR primer pairs as described earlier in this report were used for BCCIPshRNA and EIIaCre genotyping. The p53+/−;LoxPshBCCIP+/+ mice were obtained by crossing p53+/−;LoxPshBCCIP+/− females with p53+/−;LoxPshBCCIP+/− males. The p53+/−;LoxPshBCCIP+/+ or p53+/−;LoxPshBCCIP+/− mice were crossed with p53+/−;EIIaCre+/− mice respectively.
|
10.1371/journal.pntd.0003306 | Approaches to Refining Estimates of Global Burden and Economics of Dengue | Dengue presents a formidable and growing global economic and disease burden, with around half the world's population estimated to be at risk of infection. There is wide variation and substantial uncertainty in current estimates of dengue disease burden and, consequently, on economic burden estimates. Dengue disease varies across time, geography and persons affected. Variations in the transmission of four different viruses and interactions among vector density and host's immune status, age, pre-existing medical conditions, all contribute to the disease's complexity. This systematic review aims to identify and examine estimates of dengue disease burden and costs, discuss major sources of uncertainty, and suggest next steps to improve estimates. Economic analysis of dengue is mainly concerned with costs of illness, particularly in estimating total episodes of symptomatic dengue. However, national dengue disease reporting systems show a great diversity in design and implementation, hindering accurate global estimates of dengue episodes and country comparisons. A combination of immediate, short-, and long-term strategies could substantially improve estimates of disease and, consequently, of economic burden of dengue. Suggestions for immediate implementation include refining analysis of currently available data to adjust reported episodes and expanding data collection in empirical studies, such as documenting the number of ambulatory visits before and after hospitalization and including breakdowns by age. Short-term recommendations include merging multiple data sources, such as cohort and surveillance data to evaluate the accuracy of reporting rates (by health sector, treatment, severity, etc.), and using covariates to extrapolate dengue incidence to locations with no or limited reporting. Long-term efforts aim at strengthening capacity to document dengue transmission using serological methods to systematically analyze and relate to epidemiologic data. As promising tools for diagnosis, vaccination, vector control, and treatment are being developed, these recommended steps should improve objective, systematic measures of dengue burden to strengthen health policy decisions.
| Dengue is the most common mosquito-transmitted viral disease. It represents a formidable public health problem that is expanding in both infection rates and geographical range. Fortunately, vaccines, improved diagnostics, innovative vector control approaches, and other disease control methods are under development. Despite the importance of dengue, there is substantial uncertainty about the magnitude of the disease burden and economic cost of dengue, particularly in the number of symptomatic dengue infections. There is substantial variation in national reporting systems for dengue, which hinders accurate estimates of total cases and, therefore, of economic burden. Here we suggest a combination of immediate, short-term, and longer term strategies to address this knowledge gap. Immediate strategies include, for example, documenting the number of ambulatory visits before and after hospitalization. Short-term recommendations include merging multiple data sources, such as cohort and surveillance data to improve estimates of dengue incidence. Long-term efforts include increasing the collection and analysis of seroprevalence and economic data, use of enhanced surveillance (e.g., use of incentives to improve reporting, include private sector sentinel sites). Implementing these steps would give policy makers more reliable, systematic data for strengthening and refining policies about the application and financing of new technologies to control dengue.
| Dengue presents a formidable global economic and disease burden with around half the world's population estimated to be at risk of infection [1], [2]. Dengue transmission has intensified in the past decades, with outbreaks increasing in frequency, magnitude, and countries involved [3], [4]. Dengue disease varies across time and age of persons affected. This complexity results from the transmission of four different viruses affected by vector density, the host's immune status, age, pre-existing medical conditions and other factors [5], [6]. The impact of dengue has been measured in terms of both monetary value and public health metrics, such as disability-adjusted life-years (DALYs) [7], [8]. Here we use the term “burden of dengue illness” to refer to the amount of clinically apparent disease and mortality imposed by dengue in a population. Economic burden has three main components: (i) costs of illness, estimated from the total symptomatic episodes multiplied by the average costs per episode [9], [10], (ii) costs of dengue prevention, surveillance, and control strategies [11], [12], and (iii) other impacts of dengue, usually harder to estimate, such as effects of dengue outbreaks on tourism [13], co-morbidities and complications associated with dengue virus (DENV) infection [14]–[16], or the effects of the seasonal clustering of dengue on health systems [17]. Accurate estimates of the economic and disease burden of dengue are critical to track health progress, assess program impact and results, and inform decisions about health policy, research, and health service priorities [7], [18]–[20]. However, estimates of dengue burden have substantial variability due to limitations in the availability, quality, and use of data.
As promising technologies for vaccination, vector control, and disease management are being developed, more reliable measures of dengue illness burden are needed to produce better data on the economic cost of dengue. This systematic review aims to identify and examine estimates of dengue burden and their main sources of uncertainty and to develop an agenda for immediate, short-term, and long-term strategies to improve these estimates. Our main focus in this article is on the costs of illness, particularly from the challenges to estimate the total episodes of symptomatic DENV infections.
Available data on the economic and disease burden of dengue are limited. We conducted a systematic literature review of articles published or indexed in the Web of Science, MEDLINE, or in WHO's Dengue Bulletin, combining the keyword “dengue” with the following list of keywords: surveillance, incidence, reporting, sensitivity, capture-recapture, cohort, economics, costs, burden, Aedes aegypti, and control. In addition, we added findings from previous literature reviews on dengue disease and economic burden [9], [10], [21]. For relevance to current dengue surveillance and management, we included articles published from 1995 through 2013 in English, Spanish, French, or Portuguese. The inclusion criteria for articles at each step of the review process (i.e. identification, screening, eligibility, and inclusion) are shown in the PRISMA flow diagram [22] (Figure 1). The review process left us with 88 articles. Our goal was not to obtain numerical findings from the individual studies, but rather to summarize the main strategies and data used to estimate the economic and disease burden of dengue and the sources of variability in the burden estimates.
Estimates of the disease and economic burden of dengue were derived by combining surveillance, clinical, and cost data. Since dengue is a reportable disease in many endemic countries, the incidence of dengue in a population can be estimated initially from cases reported to the surveillance system. But because surveillance systems are not designed to capture all episodes of symptomatic dengue, relatively low reporting rates lead to conservative incidence estimates [23]–[26]. Further, national dengue reporting systems show great diversity in design and implementation, and some developing countries have important resource limitations that hamper their ability to produce any systematic dengue-related data. Recent efforts to improve estimates of dengue burden include merging multiple data sources (e.g., health and surveillance data, private laboratories, experts' opinion) [27], [28], analyzing the relationship between cohort studies and routine reporting [25], [26], [29], and estimating incidence and/or reporting rates using covariates (e.g., healthcare access and quality, geographic and climate variables) [2], [8], [23].
To illustrate, there were about 2.2 million reported episodes of dengue illnesses to WHO in 2010, but estimates of total symptomatic dengue incidence vary widely. Bhatt and others [2] estimated 96.0 million dengue episodes globally. Their study combined a range of evidence of dengue transmission [1] with various sources of occurrence data (outbreak reports, cohort studies, online reporting, etc.), adjusting for the probability of occurrence of dengue based on socioeconomic, urban, and environmental covariates. Much of the dengue reporting occurs in areas of high transmission or during disease outbreaks, creating an upward trend in reports of dengue occurrence. Resulting models may have overstated total DENV infections. Also, the ratio of inapparent to apparent DENV infections varies substantially, depending on the age of patients, herd immunity and the circulating virus strain. The 2004 WHO global burden of disease (GBD) [30], estimated 9 million dengue episodes globally based on country-level datasets and information, and a systematic review of population-based incidence and mortality studies. By adjusting surveillance data with the rate of reporting of symptomatic DENV infections to health authorities, Shepard and others estimated about 30 million annual episodes treated in the medical system globally [31]. Last, Murray and others' GBD 2010 study [8] estimated global incidence of 0.2 million dengue episodes in 2010. Noting that their approach underestimated disease burden for dengue and other neglected tropical diseases; improved updates for 2013 are underway [32].
A review of studies on the economic burden of dengue in 2011 highlighted the relatively sparse literature and conflicting results of existing cost studies [33]. A recent report proposed procedures for costing dengue illness [19]. Extensions to that document that may help in refining dengue economic evaluations include estimating unit costs that are sensitive to productivity loss for workers that are not part of the formal economy (e.g., estimating the local marginal productivity of labor based on local wages averaged over the dengue season), examining local health-seeking behavior involving pharmacies or traditional healers, or using macro-costing techniques, which allow one to allocate overall operating costs among the outputs of a health facility [34] when person-level costs are unavailable. Estimates of health system congestion costs are also important; when health facilities are close to their capacity, the costs of an outbreak should also include costs that additional episodes impose on the system as a whole, like degradation of treatment quality of non-dengue patients [17].
Most important, improving current estimates of total dengue episodes is critical to quantify the disease and economic burden of dengue. Understanding the main sources of variability in the availability, quality, and use of reported data will allow for more comprehensive burden estimates. Consequently, through our literature review we identified the major sources of uncertainty, as a preliminary step in this direction.
The number of limitations in reporting symptomatic dengue infections makes it difficult to estimate the true burden of dengue illness, which is probably underestimated in most studies. In this section, we suggest immediate, short-, and long-term refinements in data collection and analysis to improve the accuracy of estimations of the total dengue episodes and other components of disease burden. Table 1 lists the main sources of variability in dengue burden estimates and possible ways to improve data collection, including a few examples for some suggested improvements [6], [70], [91]–[94]. In the remainder of this section, we discuss possible analysis refinements with currently available data, or at least, data that could be gathered in the short run with marginal additional efforts.
Multiple factors contribute to the variability in estimates of dengue burden, making it challenging to obtain accurate estimates. We recommend a series of strategies for improving dengue-burden estimates; however, some of them may be costly and therefore harder to achieve, and strategies themselves may need to be evaluated for their cost-effectiveness. Possibly the most important limitation has to do with limited availability, quality, and use of dengue surveillance data in many countries. New prospective studies to ascertain dengue burden better are needed, particularly in areas where reporting is least complete (or nonexistent), such as Africa or South Asia. However, several improvements in economic and disease burden estimates may be achieved with available data. Reported surveillance data should include a narrative about the system's main characteristics, including whether it includes the private sector, ambulatory episodes, cases of all ages, and type of lab confirmation, if any, of DENV infections reported. Most importantly, reporting to national surveillance systems should record each dengue episode as either hospitalized or ambulatory (i.e., never hospitalized). The use of covariates to estimate the burden of dengue can adjust for underreporting and/or to extrapolate to areas where there is no reporting at all [2], [8], [23]. It would be important to characterize the context for epidemiological dengue studies to describe why these studies were conducted at the specific time and place, and how those settings compare to others in the country or region. Understanding how specific variables affect the burden of dengue will help researchers improve burden estimates. The greatest source of uncertainty in existing burden of dengue studies comes from underreporting of symptomatic DENV infections, followed by the type of treatment of episodes. Probabilistic sensitivity analyses and tornado diagrams are helpful to understand the proportion of a confidence interval that arises from various sources of uncertainty [10], [40]. The biggest payoff for burden of dengue estimates would come from studies that can link and analyze existing data. For example, data from cohort studies and clinical trials could be re-analyzed and compared with officially reported dengue episodes to estimate EFs [104] and population-based economic burden. Understanding the health-seeking behavior of people with symptomatic DENV infections would, for example, allow researchers to estimate the probability that a dengue episode is reported as a function of setting (inpatient or outpatient), sector (public or private), case severity, age, type of facility, access to healthcare, and other variables in the surveillance system. We also expect that neglected impacts of dengue, such as decreases in tourism or health system congestion, would represent substantial costs during outbreaks.
We hope that future studies will obtain more accurate and comparable measures of economic and disease burden of dengue, for example, by documenting surveillance reporting criteria and adjustments used to estimate total symptomatic DENV infections (including adjustments for dengue episodes treated in the private sector or alternative health providers); using consistent case definitions; stratifying by treatment setting (hospitalized and non-hospitalized), severity, and age; using probabilistic sensitivity analysis to estimate uncertainty; and including comprehensive analysis of prevention and control costs. These improved estimates will be crucial for public health advisors and policy makers to identify optimal and cost-effective dengue control technologies and financing. Compared to other diseases with higher mortality rates or more frequent chronic symptoms, the DALY burden of dengue is relatively low; nevertheless, dengue poses a substantial burden on a large share of the world population. Estimates of dengue burden are sparse and there is significant room for refinement. Understanding the factors that shape the uncertainty around dengue burden and reporting will enable improvement of current estimates. Improving the methods to quantify dengue endemicity, for example, by using a measure of DENV incidence rather than disease, would also be a major improvement towards the goal of controlling dengue as it may allow more direct cross-country comparisons [93]. In the long run, we aim to identify the most cost-effective ways to control dengue, by combining various data sources and improving analytical tools. Costing studies can help us examine existing preventive and treatment approaches. Economic and epidemiological models can project costs and effectiveness of existing and alternative approaches in a range of settings.
Most likely the future paradigm of dengue prevention and control will require an integration of vaccine, vector control, and anti-viral strategies, and systematic, comparable measures of dengue burden will be increasingly important. Several organizations have called for the improvement of health data [18]. We, too, believe this is an essential global public good that will help prioritize and improve public health decisions locally and globally.
|
10.1371/journal.pgen.1000366 | A Gene-Based Linkage Map for Bicyclus anynana Butterflies Allows for a Comprehensive Analysis of Synteny with the Lepidopteran Reference Genome | Lepidopterans (butterflies and moths) are a rich and diverse order of insects, which, despite their economic impact and unusual biological properties, are relatively underrepresented in terms of genomic resources. The genome of the silkworm Bombyx mori has been fully sequenced, but comparative lepidopteran genomics has been hampered by the scarcity of information for other species. This is especially striking for butterflies, even though they have diverse and derived phenotypes (such as color vision and wing color patterns) and are considered prime models for the evolutionary and developmental analysis of ecologically relevant, complex traits. We focus on Bicyclus anynana butterflies, a laboratory system for studying the diversification of novelties and serially repeated traits. With a panel of 12 small families and a biphasic mapping approach, we first assigned 508 expressed genes to segregation groups and then ordered 297 of them within individual linkage groups. We also coarsely mapped seven color pattern loci. This is the richest gene-based map available for any butterfly species and allowed for a broad-coverage analysis of synteny with the lepidopteran reference genome. Based on 462 pairs of mapped orthologous markers in Bi. anynana and Bo. mori, we observed strong conservation of gene assignment to chromosomes, but also evidence for numerous large- and small-scale chromosomal rearrangements. With gene collections growing for a variety of target organisms, the ability to place those genes in their proper genomic context is paramount. Methods to map expressed genes and to compare maps with relevant model systems are crucial to extend genomic-level analysis outside classical model species. Maps with gene-based markers are useful for comparative genomics and to resolve mapped genomic regions to a tractable number of candidate genes, especially if there is synteny with related model species. This is discussed in relation to the identification of the loci contributing to color pattern evolution in butterflies.
| Butterflies and moths (called the Lepidoptera) are a large and diverse group of insects that has long captured the attention of biologists and laymen. The colorful patterns on the wings of butterflies, in particular, offer an ideal system to investigate which genes and developmental mechanisms contribute to evolutionary diversification. Genetic analyses that try to find the position of genes along chromosomes are invaluable for such efforts, also because they allow researchers to compare chromosome content between species. Here, we report on a study which built a gene-based map for the chromosomes of a butterfly “lab rat” and identified chromosomes carrying color pattern genes. We compare our map to that of the reference lepidopteran species, the silkworm. Despite these species having diverged some 100 million years ago, there is much conservation in terms of which genes are found together in chromosomes and even how genes are ordered within chromosomes. However, because we were able to compare positioning of many more genes than had ever been reported before for this group, we also found evidence of several large- and small-scale chromosomal rearrangements. We discuss the advantages of gene-based maps in understanding the genetic basis of color pattern evolution.
| With the need for a wider sampling of biological diversity [1]–[3], the availability of tools for large-scale genetic and genomic analysis is rapidly being extended beyond a handful of classical model systems. Gene collections are growing for various species and with them, the need for methods to assign genes to genetic maps and to assess synteny with relevant sequenced genomes. Gene-based linkage maps are invaluable in the search for the loci that contribute to phenotypic evolution; they are more easily transferable and comparable between species than anonymous markers, and facilitate resolution of mapped genomic regions to candidate genes, also via comparisons of maps or gene functions between species.
The Lepidoptera (butterflies and moths) are a diverse order of insects with an abundance of species, including many agricultural pests, and one of two species of domesticated insects. Lepidopterans have some unusual genetic properties, such as holocentric chromosomes, heterogametic females, and male-restricted meiotic recombination, whose underlying mechanisms and consequences for genome evolution remain to be fully explored. However, lepidopteran species are relatively under-represented in terms of genomic resources with little available outside the model silkworm Bombyx mori [4]. Comparative genomics among lepidopterans and a detailed comparative analysis of the B. mori genome have been hampered by the relative scarcity of relevant genomic information. Dipterans, the closest insect lineage with available sequenced genomes, diverged from lepidopterans more than 200 MYA, and there is relatively little genomic information within the Lepidoptera. This is especially striking for butterfly species (derived from moths some 100 MYA), despite much interest in their diverse, derived, and ecologically-relevant wing patterns.
Color patterns on butterfly wings include some compelling examples of adaptation and have regained interest in evolutionary developmental biology's quest to understand the mechanistic basis of phenotypic variation [5]–[7]. A number of candidate genes well described in relation to wing development in Drosophila melanogaster have been implicated in formation (reviewed in [5],[8]) and variation [9]–[11] of wing patterns in butterflies. Despite the success of the Drosophila-based candidate gene approach, it is clear that a more unbiased approach will be necessary. For example, for those cases where there are no obvious candidate genes [12], and because it is conceivable, if not likely, that genes other than those described for a derived model system will be relevant for traits that are restricted to a lineage diverged more than 200 MYA. For this reason, there have been a number of recent efforts to push forward butterfly genomics [13],[14], including construction of large EST collections [15]–[17], and genetic linkage maps [18]–[21] for a few target species. The latter are, however, largely or exclusively composed of anonymous markers, limiting broad-coverage comparative analysis of gene co-segregation and order across species. Recent studies based on a limited number of pairs of mapped orthologous markers have proposed conservation of syntenic blocks and gene order between B. mori and Manduca sexta moths and/or Heliconius melpomene butterflies [22]–[25]. Extending this type of analysis to many more pairs of mapped orthologs will be crucial to exploring the use of B. mori as a pan-lepidopteran genomics reference, and to allow integration of genomics information now accumulating for different species of butterflies [14].
Bicyclus anynana is probably the closest to a butterfly equivalent of a “lab rat”. This species was introduced to captivity some two decades ago and it has since been the focus of studies on the evolution and development of wing patterns and other phenotypes [26]. Two key processes in morphological evolution are captured on the wings of these butterflies; diversification of evolutionary novelties (as are the scale-based color patterns of butterflies [27]), and of serially-repeated structures (as are the eyespots of many Nymphalids [28]). Laboratory populations of B. anynana have been used to examine the genetic, developmental, and physiological basis of phenotypic variation [29], and have provided the material for identification of anonymous [30] and expressed gene-based [15],[31] markers. Here, we describe a study that genetically maps SNPs in a large number of ESTs to B. anynana chromosomes. We use a mapping panel composed of a number of small families to maximize the number of mapped markers and produce the densest gene-based map available to date for any butterfly species. This map includes a number of color pattern loci defined by spontaneous Mendelian mutations and enabled a large-scale analysis of synteny with the lepidopteran reference species. The usefulness of gene-based linkage maps and comparative analysis of chromosomal composition is illustrated in relation to the identification of color pattern loci.
We used a biphasic linkage mapping method [32] to map 508 markers on expressed genes and seven color pattern loci in B. anynana linkage groups (LGs). With the largest collection of anchor loci mapped to date for any butterfly species, we were able to do a broad-coverage comparison of gene co-segregation and gene order between B. anynana butterflies and the lepidopteran reference species, Bombyx mori. Our analysis confirmed previous reports of conserved synteny in the Lepidoptera and also detected several small- and large-scale chromosomal rearrangements separating B. anynana butterflies and B. mori moths.
We selected 768 SNPs in expressed genes to genotype in a mapping panel composed of 288 individuals from 12 F2 families (Table S1). These markers correspond to 745 SNPs in 744 UniGene contigs (marker name starting with BaC) and 23 SNPs in 14 selected candidate genes (marker name starting with BaG). The contigs were identified from the assembly of over 100,000 EST reads, and the candidate genes were selected based on their developmental roles (see Methods). We selected a single SNP for most genes, with the exception of 5 genes whose potential role in development warranted extra effort. Seventy percent (533 of 768) of the target SNPs converted into good assays, defined as those with 90% of the panel individuals being genotyped and with a minor allele frequency greater than 5%. For these SNPs (Table S2), each of the 12 families had an average of 60 SNPs that were informative in females only (ranging from a maximum of 84 to a minimum of 43), 63 SNPs that were informative in males only (ranging from 81 to 44), and 141 SNPs that were doubly-informative (ranging from 155 to 119). On average, each of those SNPs was male-informative only in 1.4 families, female-informative-only in 1.3 families and both male and female informative in 3.2 families. Upon visual inspection of the genotypes for the 533 markers (Table S2), we identified 513 markers with autosomal segregation patterns, 9 with segregation patterns consistent with sex linkage, and 11 with several Mendelian inconsistencies which were excluded from further analysis.
Absence of recombination in lepidopteran females can be exploited to construct genetic linkage maps via “biphasic mapping” [32]; marker pairs that are female fully-informative are used to initially assign markers to segregation groups, and male informative markers are then used to order markers within those groups. We used this strategy and CRIMAP software for pedigree analysis [33] and were able to assign 508 SNPs to 28 B. anynana linkage groups (Table 1; Figures 1–4), possibly corresponding to the 27 autosomes and Z sex chromosome of this species [20]. We were able to map 10 of our 14 candidate genes (BaG markers): cubitus interruptus (ci), Ecdysone receptor (EcR), engrailed (en), APC-like (Apc), naked cuticle (nkd), cinnamon (cin), Henna (Hn), echinus (ec), Catalase (Cat), and Heat-shock protein 70 (Hsp70). Failure to map the other four candidate genes was due to a failed assay (split ends, spen), Mendelian inconsistencies (scabrous, sca), or the SNP being fixed in the mapping panel (wingless, wg and groucho, gro).
We also attempted to map nine Mendelizing mutations affecting body or larval coloration which were segregating in some of the 12 full-sib mapping families (Table 2). Two of the visible markers could not be mapped (LOD score not significant), and the other seven were assigned to six LGs. Mapped markers typically had poorly resolved map positions, often corresponding to the entire length of the chromosome (Figures 1–4, Table 2). Among the mapped visible mutants, two are particularly worrisome: 1) the Spotty mutation for which a 2 LOD support interval included positions at either end, but excluded the middle region of LG10 , and 2) the Goldeneye mutation which mapped to LG28 whose validity we are uncertain of (see below). Poor mapping resolution for the visible markers likely reflects the fact that: 1) any given mutation was typically only segregating in 1–4 families (Table 2), 2) in the case of non co-dominant mutations a fraction of the segregants needed to be scored as “missing” which resulted in further loss of resolution, and 3) the mutations may not be 100% penetrant. Nonetheless, the mapping of these mutations to chromosomes is a very valuable first step towards efforts to clone the corresponding loci. Fine mapping efforts need now only employ markers in the same linkage groups.
With 508 markers in expressed genes and seven visible mutants, this is the densest non-anonymous marker map ever reported for a butterfly species. Up until now, the most anchor loci mapped in this group was 101 for Heliconius melpomene (cf. [23]), another Nymphalid.
For all the SNPs assigned to a given segregation group we used male informative markers to build a map for that group (Figures 1–4). For 297 of the 508 gene-based markers, we were able to assign a position in the corresponding LG (hereafter, “ordered markers”; Table S3). The remaining 211 markers were not assigned to a unique position, but their position was typically narrowed to two or three intervals (hereafter, “unordered markers”; Table S4). LGs had on average 10.6 ordered and 7.5 unordered markers with standard deviations of 5.5 and 4.6 respectively (Table 1). Three linkage groups (LG24, LG27, and LG28) consisted only of single ordered markers at the tips and zero to two extra, unordered markers. In addition, despite a total of 24 markers assigned to LG13, this LG only had markers placed at the tips. The reasons for poor marker resolution in LG13 are unknown.
Our total estimated map length, based on LG “male-based” distance between terminal markers, was 1642.2 cM, with individual LG length varying between ∼14 cM (LG14) and ∼122 cM (LG19) (Table 1). This map length is well within that estimated for different butterfly species (1430 cM–2542 cM, [18],[19],[21]) and close to that estimated for B. anynana based largely on AFLP markers (1354 cM or 1873 cM depending on the mapping software used cf. [20]). However, because of the probable non-zero distance between terminal markers and chromosome ends, the “male distance” between terminal markers can be an underestimate of actual LG lengths.
Our dataset allows for two types of quality control of map assignments. First, estimates of map distance in females (which should be zero since they do not have recombination) is a measure of potential map expansions due to errors. About 26% (70 of 269) of the “female distances” between neighbor ordered markers were greater than 0 cM (Table 1). The average distance for the non-zero distances was 3.0 cM, and included 11 distances greater than 5 cM, and 4 greater than 10 cM (Table S3). The total female map is 212.4 cM implying a map expansion due to genotyping errors and/or the mapping algorithm of ∼12.9%. The extent of this expansion varies greatly between LGs (Table 1); while for most, the expansion is lower than 10%, for LG12 it reaches 55% (due mainly to a single terminal marker; see Table S3). Secondly, male recombinational distance between multiple SNPs at the same gene measures error in map position assignments. We have two genes where multiple markers have been ordered, Apc (LG6) and EcR (LG10). For Apc two of the three ordered markers overlap and the third maps at a distance of 5.4 cM from them, while for EcR all three markers map to positions within 2.6 cM from each other (Figures 1–4). The average maximum distance between ordered non-overlapping markers at the same locus is 4 cM, and the average distance of the four possible distances between consecutive markers (0, 5.4, 1.9, 0.7) is 2 cM. Some of this error is certainly associated with genotyping errors, but it may also result from our mapping approach which attempts to integrate marker information over several families (see below). In any case, this analysis suggests that distances smaller than ∼5 cM might not be well resolved in our map.
Our method was designed to maximize the number of gene-based markers assigned to linkage groups with minimum de novo SNP identification. This approach involved: 1) focusing on SNPs identified in EST collections (thus, in expressed genes) and for which the minor allele was seen at least twice (thus making it less likely that SNPs are cDNA-related errors; [31]), 2) using a mapping panel composed of a number of small families rather than one large one (maximizing the number of mapped markers at the expense of their mapping resolution; see below) and CRI-MAP software for pedigree analysis, 3) using Illumina GoldenGate genotyping technology (without any per SNP assay optimization), and 4) following a biphasic linkage mapping method [32] which takes advantage of the fact that there is no recombination in lepidopteran females.
We chose to use a mapping panel made up of a number of small families rather than the more typical single large family. With this strategy we maximize the chance of assigning any given SNP to a LG (as this only requires having one female informative family), and, once assigned to a linkage group, we maximize the chance of identifying at least a second family in which that SNP is (also) male informative. Of the 508 mapped markers, 11 corresponding to nine Z-linked loci and to two autosomal markers (BaC645 on LG2 and BaC4454 on LG11; cf. Methods) were not female-informative (i.e. heterozygous) in any family. Similarly, only three SNPs were not male informative in any family (Table S2). A panel derived from several independent parental pairs, additionally allows for estimates of population SNP frequencies, which will be useful in future mapping experiments. The disadvantages of this strategy are noticeable in terms of mapping resolution when a marker is male informative only in a single (or few) families and because of the need to integrate marker information across families. The majority of the SNPs (264 of 508) were male-informative (including both SNPs informative only in males and those doubly-informative) in at least five families (corresponding to at least 120 individuals in the mapping panel), and three SNPs were informative for a maximum of ten families (240 individuals).
In downstream uses of this map (e.g., for mapping QTLs or visible mutants), we will be able to choose from mapped, intermediate-frequency, informative SNPs, to design assays for larger mapping panels derived from a smaller number of founders. For this, having a large number of gene-based markers (even if mapped with limited resolution) and knowledge of SNP frequency is more useful than having a very accurate map of sparse markers (which may not be informative in another context).
A recent very high density SNP map for Bombyx mori [34] combined with a new assembly (unpublished) of the whole-genome sequence of this species [35],36 with larger average scaffold sizes, may be used as a pan-Lepidoptera reference. Using blastn, we assigned 1711 of the 1755 mapped SNPs in the silkworm B. mori [34] to the recent genome-sequencing scaffolds [37] (Table S5). The mapped markers aligned to 185 of the 645 different scaffolds, consistent with the highly skewed distribution of scaffold lengths. Of the 185 scaffolds to which markers mapped 29%, 10%, 6%, and 6% had one through four mapped markers, respectively. On the other hand, ∼90% of the mapped markers were assigned to only 91 scaffolds having more than four markers each, implying that the bulk of the current B. mori genome assembly is contained in 91 large scaffolds. As a check on the quality of the current assembly, we looked for scaffolds with more than four mapped markers in which at least one marker mapped to a different linkage group than the remainder. We observed seven scaffolds (7.7%) with material coming from two chromosomes and two additional scaffolds (2.2%) with material derived from three chromosomes, suggesting that there are errors with the current assembly. A visual inspection suggests that those assembly errors tend to be associated with the very ends of scaffolds. So, although the fraction of large scaffold with such errors is significant, very little of the assembly is affected. We next fitted local regressions for each scaffold that allowed for predictions of genetic positions (cM) given a physical position (bp) on the scaffold (see Methods). The B. mori map thus generated was the basis for the comparative analysis with our B. anynana gene-based map.
The current B. mori map consists of over 1650 SNPs covering 1413 cM [34]. With 28 chromosome pairs, B. mori has the largest chromosome number of all insect genomes sequenced to date. Previous studies of deep synteny across insects showed that divergent gene order correlates with divergent protein sequence [38], and that there is more conservation of syntenic groups between B. mori and the coleopteran Tribolium castaneum than between B. mori and the hymenopteran Apis mellifera [34]. Both these orders have presumably split from a common ancestor with lepidopterans earlier than dipterans did. However, analysis of synteny blocks with sequenced representatives of the Diptera is hampered by the large difference in chromosome number; typically around 30 pairs in most lepidopteran species [39] and between three and six pairs for the various sequenced dipteran (mosquito and Drosophila) species. Among lepidopterans, and despite the available phylogenetic framework for comparative analysis (e.g. [40]–[44]), relatively sparse genomic resources have resulted in very few attempts to examine synteny. Previous studies compared synteny blocks between moths and Heliconius butterflies based on a modest number of mapped orthologous pairs (maximum 72 with many “unordered” [23]). Here, in a comparison between B. mori and B. anynana genetic maps, we increased this number by more than seven times, with a large fraction of our markers being “ordered”. This type of analysis, hopefully extending also to representatives of the microlepidoptera (all lepidopterans examined to date are macrolepidopterans), will be crucial to put the gene collections and genetic maps, growing for a variety of butterfly species, into phylogenomic context.
We used blast to identify orthologs of the gene-based markers in B. anynana (Neolepidoptera; Papilionoidea; Nymphalidae; Satyrinae) mapped in other lepidopteran species: the butterfly Heliconius melpomene (Neolepidoptera; Papilionoidea; Nymphalidae; Heliconiinae), and the silworm Bombyx mori (Neolepidoptera; Bombycoidea; Bombycidae; Bombycinae). Of the 508 B. anynana markers, 29 (18 ordered and 11 unordered) had an ortholog among the 101 anchor loci mapped in H. melpomene [23], and 462 (269 ordered and 193 unordered) could be assigned to a mapped B. mori scaffold. Of the remaining 46 mapped B. anynana markers (blue in Figures 1–4), 20 had orthologs in B. mori scaffolds which we could not assign to a B. mori LG and 26 did not have significant sequence similarity with any B. mori scaffold.
Despite the ca. 100 MY that separate butterflies and moths [23],[42], there is much conservation of the grouping of genes in linkage groups (Figures 1–5). Our numbering of B. anynana LGs reflects homology with B. mori with the exception of B. anynana LG28, which has only two markers and none with orthologs mapping to B. mori LG28 (Figures 1–4). Of the 462 pairs of mapped orthologous markers in the two species, 425 (∼92%) are found in orthologous LGs (Figure 5). The 37 orthologous genes found on non-orthologous LGs, include 17 that are associated with three large chromosomal rearrangements (involving LG2 and LG24, LG16 and LG23, and LG20 and LG28), and 20 which are potential single gene transpositions. The latter may also include blast false positive (even though only five cases had e-values higher than 1.0e-20; Figure 5), blasts to pseudo- or duplicate genes, or mapping errors (e.g., markers mapping to non-syntenic linkage groups that are isolated at the tips of chromosomes are especially suspicious). Both B. mori and B. anynana have 28 pairs of chromosomes, while basal lepidopterans have 31 pairs [45] and different species of butterflies and moths have very variable numbers [39], [45]–[47]. The instances where individual B. anynana LGs are made up of syntenic blocks from different B. mori LGs suggest that the two lineages have undergone independent karyotype reductions, via non-homologous chromosomal fusions.
A previous study compared linkage group assignment for 72 orthologous pairs of markers available for another Nymphalid butterfly (Heliconius melpomene) and the reference lepidopteran (Bombyx mori) and concluded that extensive synteny existed [22],[23]. Some striking differences, however, are apparent between the genome-wide analysis of macro-synteny for B. mori and B. anynana (this paper) and that for B. mori and H. melpomene [23]. First, the comparison between H. melpomene and B. mori did not detect any of the chromosomal rearrangements we document (Figures 1–4). This may be because these rearrangements are not present in Heliconius, or because they could not be detected given the relatively small number of mapped orthologs pairs in H. melpomene and B. mori. Thus, it remains unclear to what extent the rearrangements we see in B. anynana are characteristic of Nymphalid butterflies or more lineage-restricted. Second, we see no evidence of the six chromosomal fusions proposed to distinguish the H. melpomene and B. mori genetic maps [23]. This probably reflects the fact that Heliconius butterflies have a lower chromosome number (21 pairs instead of the 28 pairs in both B. anynana and B. mori), and must thus have undergone further, or independent, chromosomal fusions relative to B. anynana. It is, however, noteworthy that the proposed fusions separating H. melpomene and B. mori are based on few pairs of mapped orthologous markers (mostly 1–3 pairs [23]) and our analysis shows that single marker transpositions do occur (Figures 1–5).
Figures 1–4 illustrate synteny between B. anynana and B. mori orthologous markers: both in terms of the grouping of markers in LGs (see also Figure 5), and in terms of conservation of gene order along individual LGs. For most LGs with multiple ordered markers, we have evidence for some reordering of genes which suggests multiple inversions separating B. anynana and B. mori.
From the 23 B. anynana LGs with greater than three ordered and non-overlapping markers (i.e. excluding multiple markers mapping to the same genetic position), with a mapped B. mori ortholog on a syntenic block, only LG10 and LG21 have fully conserved marker order (Figures 1–4). For the remaining LGs, we see evidence of order rearrangements ranging from one (e.g. LG9, LG18) to multiple (e.g. LG17, LG19) markers whose relative position in B. anynana differs from that in B. mori. Where the marker order inferred for B. anynana differed from that of the orthologous markers in B. mori, we compared the log10 likelihoods of the two (Table S6). Of the 24 comparisons made (complete LGs or LG fragments with homology to different B. mori LGs), inferred marker order in B. anynana was at least twice as likely than B. mori order in 20 cases (and at least 630 times more likely for 18 of the comparisons). For the four situations where the B. mori order was better supported than the one originally inferred for B. anynana (LG2, LG6, LG10, and LG17), and for LG13 (which had many but very poorly resolved markers and where the original inferred order was only ∼2 times better than that of B. mori), we used the B. mori order as a starting point in CRI-MAP and further improved it (see Methods). In all cases except LG10, and the LG2 segment homologous to B. mori LG2, the final order was different from that in B. mori. The difference between the log10 likelihoods for the final inferred marker order in B. anynana and that in B. mori ranges between 1.1 for LG13 (i.e. the inferred B. anynana order is ∼13 times more likely than that in B. mori) and 34.8 for LG11 (i.e. inferred order ∼10∧34 times better) (Table S6). Because de novo map construction using CRI-MAP uses a “hill climbing” algorithm to maximize marker order likelihood, the map order arrived at is dependent on the particular subset of markers used to initiate a build. This explains why the build corresponding to some B. anynana LGs reached a local maximum that could be improved upon by using the B. mori gene order as seed. Note that marker mapping was further improved by re-adding to the map markers with no mapped B. mori ortholog and by re-assessing unordered markers in those LGs. The mapping information in all Tables and Figures corresponds to the final CRI-MAP builds.
Our data suggest that previous conclusions of highly conserved gene order between H. melpomene and B. mori [23]–[25] may have been over-stated, perhaps as a result of the limited number of markers examined (maximum of 10 ordered orthologous pairs in one syntenic block [25]). Future work adding extra markers and improving marker mapping resolution in B. anynana, and extending comparative analysis to additional species will be crucial to rigorously quantify the extent of inversions separating different lepidopteran lineages. Unfortunately, the number of shared ordered markers in the H. melpomene and B. anynana maps prevents evaluating the consistency of gene order within Nymphalid butterflies. We have a single B. anynana LG (LG15) with greater than two ordered markers with mapped orthologs in H. melpomene. However, of those four markers, only one has a resolved genetic map position in H. melpomene [23] making impossible the assessment of conservation of gene order.
Previous studies that analyzed order of more than three ordered H. melpomene - B. mori marker pairs were much more localized than the study presented here. They either focused on one individual chromosome (and reported on four perfectly aligned markers [24]), or on a BAC-level scale (and reported on nine of ten aligned markers [25]). However, conservation of gene order for small collections of orthologous markers can occur by chance alone (e.g., four perfectly aligned markers occur by chance ∼10% of the time), and comparisons of marker order at the level of single BACs can only infer conservation at sub-centimorgan scales. Here, we extended the analysis of gene order to many more markers in many more linkage groups and alert for the fact that, even though we have syntenic blocks and broad conservation of gene order (see, for example, LG10), we also have clear evidence of multiple rearrangements (see, for example, LG19). These intra-chromosomal rearrangements do not mean that B. mori cannot serve as a pan-macrolepidopteran reference, but they do argue that marker order is likely conserved over tens of centimorgans as opposed to entire linkage groups. Our observations are remarkably similar to the emerging consensus view in the Drosophila clade (including species diverged some 40 MYA), that the assignment of genes to Mullerian elements is highly conserved but gene order within those elements is variable [48],[49]. It will be interesting to look both more widely (across species from different families) and also more narrowly (across multiple species within some selected genera) in the Lepidoptera. It is still unclear how the relatively numerous and relatively small (in insect terms) chromosomes in this diverse group have evolved and what the role of the holocentric chromosome structure and male-restricted recombination has been in the process.
Aside from enabling analysis of macro- and micro-synteny, gene-based maps are of great value in studies attempting to map genes that contribute to phenotypic variation because they greatly facilitate the resolution of mapped genomic regions into a tractable number of candidate genes. This is not only because the mapping analysis itself can exclude candidate genes (namely, those in non-implicated LGs), and identify candidate genes among available markers, but also because conservation of gene grouping and gene order in related species with dense linkage maps might allow identification of extra candidate genes within the implicated genomic regions. For example, the B. mori ortholog to the pigmentation gene black localizes to a B. mori scaffold (nscaf2986 in [37]) mapping to the Chocolate-containing region of B. anynana LG7. This makes black an interesting candidate gene for the Chocolate larval phenotype (Figure 6I).
With the exception of Bigeye and Chocolate (and the more dubiously mapped Spotty; see above), at present we have only mapped our collection of B anynana visible mutants to entire linkage groups (Table 2). While this renders identification of individual candidate genes premature, our analysis enables us to clearly identify “anti-candidates”. For example, from a developmental point of view, the gene engrailed would be a good candidate for several of our Mendelian mutations. The expression of engrailed is regulated in relation to different stages of eyespot development [50], and to changes in eyespot size [29], color-composition [50] and number [51],[52]. The involvement of engrailed in eyespot formation and also in embryonic development [27] makes it a potential candidate gene for mutations such as Goldeneye and Bigeye which affect both embryonic viability and eyespot morphology (Figure 6). However, none of the mapped visible markers maps to the engrailed-containing LG2, and hence none can be alleles at this locus. This, of course, does not mean that the engrailed locus cannot contribute to complex naturally occurring segregating variation or other laboratory mutations affecting wing patterns. Future studies trying to refine the location of each of our mapped color pattern loci will need only to concentrate on markers throughout single LGs, greatly reducing the genotyping effort.
Another exciting aspect of having color pattern loci in gene-based maps of different lepidopteran species is the possibility to investigate to what extent color pattern diversification in different lineages has a similar genetic basis. Recent studies have shown that color pattern loci contributing to race variation map to homologous genomic regions in different Heliconius species [12],[53]. Whether these loci play a role in color pattern variation outside Heliconius and to what extent color pattern diversification has repeatedly recruited the same loci in different lineages are interesting questions in evolutionary (developmental) biology. We looked for H. melpomene and B. mori color pattern loci mapping to orthologous LGs to those where we mapped visible markers in B. anynana (Table 3). Particularly interesting is the case of the B. anynana Bigeye and 067 spontaneous mutations, both affecting eyespot size (Figure 6, Table 2). We mapped these to LG17, which, based on comparisons to B. mori, we know is orthologous to H. melpomene LG15 (Table 3). This is the linkage group carrying the color pattern loci above-mentioned which have been implicated in the race-divergence in three different Heliconius species [6],[12],[23]. Also, the Band mutant with lighter background coloration on the distal section of the wings maps to LG4 whose Heliconius ortholog carries a white/yellow color switch locus [10]. In the future, emerging comparative maps in Heliconius and Bicyclus can be exploited to accelerate the dissection of the genetic basis of wing pattern variation in butterflies; potentially aided by patterns of conserved microsynteny detected for “developmental genes” in insect genomes [54].
With gene collections growing for a variety of species, so is the need for methods that enable the mapping of markers in those genes and comparisons with genetic maps of relevant reference species. These maps will aid in the genetic dissection of phenotypic variation in non-model systems, enable analysis of synteny and genome evolution, and facilitate future sequence-assembly efforts. Here, we report on the mapping of 508 markers in expressed genes and seven color pattern loci in an emerging butterfly model system. We used our map to compare gene grouping and gene order with the lepidopteran reference genome. Based on 462 pairs of orthologous markers mapped in Bicyclus anynana and Bombyx mori, we show that there is extensive conservation of syntenic blocks and gene order but not as much as had been previously suggested. We illustrate how gene-based maps and synteny with relevant species in relation to dissecting the genetic basis of wing pattern variation.
We used different Bicyclus anynana laboratory populations to establish a mapping panel of 288 individuals from 12 families. These were all F2 families composed of a F1 mother and father, and 22 offspring (typically 11 females and 11 males). The F2 families were obtained by using single-pairs of P grand-parents that were either from “outbred”, or 1–3 generation inbred (i.e., single brother-sister mating pairs) populations. DNA from thorax or head of freshly frozen butterflies (killed in liquid nitrogen and stored at −80°C until processed) was extracted using the QIAGEN tissue kit following manufacturer's recommendations. Genomic DNA was checked for quality and yield on agarose gel and NanoDrop spectrophotometer. From each of the 288 individuals in the mapping panel, 1.7 µg of genomic DNA in 100 µl of QIAGEN kit elution buffer was dried down (SpeedVac), re-suspended in 20 µl water, and sent to Southern California Genotyping Consortium - Illumina Genotyping Core Laboratory at UCLA [55].
We selected 768 SNPs to genotype using the Illumina Golden Gate platform [56]. The 768 target SNPs were identified in 759 expressed B. anynana genes (Table S1). These correspond to 744 contigs resulting from the assembly of an on-going, large-scale EST project (sequencing of the new ∼91,000 ESTs (GenBank GE654128–GE745563), assembly of those together with the previously published collection of ∼10,000 ESTs [15] and 13 genes available on GenBank nr database, and discovery and characterization of SNPs will be described elsewhere), and 14 candidate genes identified in previous sequencing efforts (including [15],[29],[57] ). The contig-derived markers (name starting with BaC) correspond to SNPs with a minor allele count of two or greater identified in CAP3 alignments of at least 4 EST reads. We identified ∼1,200 contigs with at least one such “double-hit” SNP and chose the 745 target SNPs based on criteria listed below. The candidate genes (marker designation starting with BaG) were selected based on their potential role in wing color patterns or other phenotypes of interest. The genes ci, EcR, en, and wg, as well as others from the Wingless signaling pathway, Apc, gro, nkd, and spen, are presumably involved in butterfly wing pattern formation [8],[58]. The genes cin and Hn are involved in pigmentation. Other candidate genes represent various key biological processes, such as wing disc development (sca), programmed cell death (ec), lifespan (Cat), and stress response (Hsp70).
We attempted to choose only one high quality SNP for each gene but, in the case of a minority of putative B. anynana homologs of developmental candidate genes, we designed two or more assays. These were: two SNPs in the pigmentation gene yellow (BaC4163), in ci (BaG15), en (BaG21), and nkd (BaG24 and BaG25); and four SNPs in Apc (BaG14 and BaG16), and EcR (BaG19 and BaG20). Many criteria went into choosing the target SNPs, including: the estimated frequency of the SNP (preference given to SNPs with high frequency of the minor allele), absence of secondary polymorphisms in the ∼100 bp up- and down-stream of it, the contig annotation (preference given to markers in genes with sequence similarity to genes in public databases), and score for Illumina “type-ability”. Sequences associated with the 768 markers we attempted to genotype are available in Genbank's sequence or EST archive; accession numbers in Table S1.
A large fraction of the SNPs assayed converted into working assays and ∼75% had a call rate of greater than 95%. The poorest 15% of SNP assays had a call rate of 0%, whereas the best 80% had a minimum call rate of 89%. The individuals in our genotyping panel consistently generated good data; a 95% Confidence Interval on the number of called SNPs over individuals was 626 to 644 with the poorest and second poorest individuals yielding 527 and 600 called SNPs, respectively. Consistent with this narrow confidence interval, we did not consider removing any individuals from the study because of poor quality DNA. The vast majority of SNPs were ascertained from an EST project so the observation that 15–25% of the attempted SNPs did not convert to a useful assay was not unexpected. Reasons for failure to convert include factors such as: SNPs having a low allele frequency in the mapping panel, some SNPs being falsely identified due to over assembly problems, introns resulting in non-functioning Golden-Gate assays, and errors in flanking regions that the Golden-Gate oligonucleotides anneal to [31]. We chose to focus solely on SNPs for which greater than 90% of the genotyped individuals were “called” and whose minor allele frequency over all called individuals was greater than 5%. These criteria resulted in a set of 533 “converting” SNPs (∼70% of assays attempted). Most SNPs not meeting our inclusion criteria were very clearly failed assays, so either relaxing or increasing the stringency for a SNP's inclusion did not greatly change the number of SNPs in further analyses.
We visually examined the dataset for clear genotyping errors that resulted in a SNP showing a pattern of inheritance inconsistent with Mendelian expectations (Table S2). SNPs fell into three categories: 1) inheritance that was sex linked (nine SNPs), 2) several Mendelian inconsistencies (eleven SNPs), or 3) no or a handful of Mendelian inconsistencies (513 SNPs). Sex-linked SNPs were duly noted as they were treated differently in subsequent steps, and SNPs showing several Mendelian inconsistencies (possibly genotyping mistakes, duplicated genes, gene families) were excluded from further consideration. For the SNPs with no or a small number of Mendelian inconsistencies, we manually changed the genotypes of those inconsistent individuals to missing. In the majority of cases this meant discarding the genotype of 1–2 of the 22 full-sib offspring in a family, but in a minority of cases the most parsimonious change involved discarding a parental genotype. This set of 513 hand-annotated putative autosomal SNPs plus the nine sex-linked SNPs were used in all subsequent mapping analysis.
We used marker-pairs that were female fully-informative (e.g., dad = aabb & mom = AaBb) to initially assign markers to segregation groups. For all possible pairs of SNPs in the 513 putative autosomal SNPs, we calculated a LOD score summarizing the evidence for complete linkage (LOG10[L(data;r = 0)/L(data;r = 0.5)]). For any given pair of SNPs that LOD score could be missing (if that pair of SNPs was never female fully-informative across the 12 families) or summarize linkage information from 1 to 12 female fully-informative families. We then grouped SNPs connected by LOD scores of greater than eleven. As a result, a SNP could be assigned to a segregation group despite not having a LOD score of greater than 11 with all the SNPs in that group. Unpublished simulations suggested that this algorithm rarely results in “over-clustering”. At our CRIMAP inclusion LOD score of 11 we identified 27 segregation groups, with the smallest number of markers in any given group being three, the largest 28 and the mean 13.6 SNPs. Increasing the LOD score for inclusion to values as high as 16 resulted in fewer SNPs assigned to segregation groups, and never split a segregation group identified at an inclusion value of 11 into two. Whereas decreasing the LOD score resulted in the merging of segregation groups (and fewer than 27 clusters).
For all the SNPs assigned to a given segregation group we used CRIMAP [33] to build a map for that group. As a result of our having 12 full-sib families, in many instances in which there existed a female informative SNP in one family, at least one other family displayed a: 1) male fully-informative SNP-pair (e.g., dad = AaBb & mom = aabb), 2) a male semi-informative SNP-pair (e.g., dad = AaBb & mom = Aabb), or 3) a doubly-informative SNP-pair (e.g., dad = AaBb & mom = AaBb). CRIMAP was designed for integrating such information in complex human pedigree data [33]. We wrote scripts to take the genotyping data for all the SNPs within a segregation group, irrespective of inheritance patterns, and create input files for CRIMAP. We then used the “build” option of CRIMAP to make a consensus map for each segregation group using default parameters, with the exception of lowering the PUK_LIKE_TOL from 3.0 to 1.0. We built a sex-chromosome map using CRIMAP by simply encoding the “second-allele” in each female as a “9” (i.e., an allele not present).
We manually inspected the resulting maps. In cases where the two “ordered-loci” used to initialize the Expectation Maximization algorithm underlying CRIMAP were loosely linked we re-ran the build option with a different set of random starting loci. In other cases where we observed SNPs that were completely linked in males we re-ran the “build” using the “hap_sys” option for those SNPs. We then used the “flips4” option on the ordered loci to confirm that our maps had the highest possible likelihood, creating a new order when necessary, and re-running the “build” and “flips4” analysis until the order stabilized.
We then used the “two-point” option in CRIMAP in an attempt to assign to segregation groups the remaining 146 putative autosomal markers, not initially assigned. This resulted in our being able to: 1) assign 134 markers to pre-existing segregation groups, 2) merge two pairs of pre-existing linkage groups in single groups, and 3) identify two novel small linkage groups (one having three and the other four SNPs). Typically, added markers displayed a high LOD score for linkage with several members of a pre-existing linkage group and below background LOD scores with members of any other group. The few SNPs that could not be assigned to any segregation group were typically only informative in a single family and/or showed segregation patterns that were unlikely under Mendelian inheritance. Based on the newly defined segregation groups and starting with the markers ordered in the previous round, we carried out another round of “builds”, followed by another round of “flips4”, and iterating until we achieved an ordering for which the “flips4” command could no longer identify orders with higher likelihoods. Details about the mapping of the ordered and unordered markers are found in Tables S3 and S4, respectively.
A total of nine Mendelizing visible mutants affecting adult or larval coloration were segregating in six of the twelve full-sibs families used for mapping (Table 2). All offspring of these families were phenotyped and 22 were selected so that each phenotypic class was represented in approximately similar numbers in the mapping panel. Consequently, segregation patterns of the visible mutants in the mapping families do not follow Mendelian ratios. To assign each visible marker to a linkage group, we used the “two-point” option of CRIMAP. This allowed us to assign seven of the nine mutant genes to linkage groups. Despite attempts with lower LOD threshold scores and/or examining only a subset of families we were unable to assign the other two mutants, comet and Missing, to linkage groups. For the seven mutants mapping to linkage groups we used the “all” option of CRIMAP separately for each mutant and its respective linkage group in an attempt to localize that mutation within a linkage group.
For all genes in the B. anynana map, we used blastn and tblastx analysis (e-score cut-off value of 1.0e-05) against the scaffolds from the B. mori genome assembly (May 1, 2008; only the “nscaf” fasta entries from [37]), and against the mapped anchor loci in Heliconius melpomene [23]. Two genes in our collection (ci and en) did not have a significant direct blast hit to any of the target H. melpomene markers (“na” notation in marker name in Figures 1–4). However, we were able to identify orthologous pairs based on annotation available for both species via blast analysis to collections from other species. For the contigs with a B. mori ortholog, we used custom prediction algorithms (see below) to estimate its position in the B. mori map. Details of the blast analysis with B. mori and H. melpomene can be found in Tables S3 and S4 for the B. anynana ordered and unordered markers, respectively.
We downloaded the new collection of B. mori scaffolds and used blastn to query all the mapped B. mori SNPs from (“DE” accessions from [34]) against the collection [37] (Table S5). We then wished to develop a prediction equation for every scaffold, that when given a base position on that scaffold would return the map position associated with that base position. Such a prediction equation would allow us to estimate a B. mori map position for any B. mori gene. For B. mori scaffolds having greater than four mapped markers this predictor is simply the slope and intercept obtained from a linear regression of map position on base position (of the midpoint of the highest scoring blast hit). For B. mori scaffolds with one to four markers this predictor is simply the average position of the markers mapping to that scaffold. For scaffolds with no mapped markers the predictor is undefined. This heuristic seemed reasonable, as a large fraction of the genome is contained in scaffolds with more than four mapped markers, and scaffolds harboring four or fewer markers are typically small enough that returning a single map position for the midpoint of that scaffold is acceptable. During this annotation effort we discovered a small number of B. mori scaffolds with termini mapping to different chromosomes, we assumed these are mis-assembly errors and removed these sections of scaffold from further consideration.
We wished to ask if within linkage group, inferred marker orders in B. anynana were different from those in B. mori. To do this we used the “fixed” option of CRIMAP to compare the inferred (non-haplotype system) order in B. anynana to that in B. mori for the subset of markers having orthologs. This analysis allowed us to obtain log10 likelihoods for both orders, and identify instances where the B. mori order was more highly supported (Table S6; see Text S1 for an explanation of the contents of all supplementary tables). In those cases we used the B. mori order as a new starting point, incorporated any observed haplotype systems, and reran the “flips” analysis to look for iterative improvements over the B. mori order. In cases where the “flips” analysis improved upon the B. mori order we obtained a log10 likelihood indicating the support for this new order over the B. mori order. We then used the order resulting from the “flips” analysis as a seed for an additional “build” run (to possibly place additional unordered markers and/or previously ordered markers without a B. mori ortholog). This final build went though additional “flips” rounds and then “fixed” was run on the final order to obtain the map displayed in Figures 1–4.
We used the MapChart software [59] to build a graphical representation of the B. anynana genetic map, and of synteny between B. anynana and B. mori chromosomes (Figures 1–4) . The map produced by MapChart was further processed to include unordered markers and visible mutations. B. mori markers were named with the corresponding B. anynana marker name, B. mori scaffold number, and blast e-score (see legend to Figures 1–4). For the graphical display of the synteny analysis, we multiplied estimated map positions of B. mori markers by a factor of two so as to facilitate visualization of homologies with the otherwise relatively condensed B. mori LGs. For markers with an estimated position of less than 0 cM, that marker's position was set as 0 cM and the positions of other markers on same linkage group were adjusted accordingly.
|
10.1371/journal.ppat.1002873 | Listeria monocytogenes Cytoplasmic Entry Induces Fetal Wastage by Disrupting Maternal Foxp3+ Regulatory T Cell-Sustained Fetal Tolerance | Although the intracellular bacterium Listeria monocytogenes has an established predilection for disseminated infection during pregnancy that often results in spontaneous abortion or stillbirth, the specific host-pathogen interaction that dictates these disastrous complications remain incompletely defined. Herein, we demonstrate systemic maternal Listeria infection during pregnancy fractures fetal tolerance and triggers fetal wastage in a dose-dependent fashion. Listeria was recovered from the majority of concepti after high-dose infection illustrating the potential for in utero invasion. Interestingly with reduced inocula, fetal wastage occurred without direct placental or fetal invasion, and instead paralleled reductions in maternal Foxp3+ regulatory T cell suppressive potency with reciprocal expansion and activation of maternal fetal-specific effector T cells. Using mutants lacking virulence determinants required for in utero invasion, we establish Listeria cytoplasmic entry is essential for disrupting fetal tolerance that triggers maternal T cell-mediated fetal resorption. Thus, infection-induced reductions in maternal Foxp3+ regulatory T cell suppression with ensuing disruptions in fetal tolerance play critical roles in pathogenesis of immune-mediated fetal wastage.
| Pregnant women are uniquely susceptible to the bacterium Listeria monocytogenes that preferentially lives inside infected cells. Since infection during pregnancy often triggers prematurity, abortion, or stillbirth, we propose that understanding how these complications occur represent important prerequisites for improving the health of mothers and the developing fetus. Here we investigate the specific interaction between Listeria and a subset of immune-suppressive cells in the mother that expand to create and maintain a hospitable environment for the fetus. We find although Listeria, especially with highdosage infection, can invade the fetus, damage can also occur by infection-induced changes in maternal immune cells that make it markedly less hospitable for the fetus. Under these circumstances, maternal immune-mediated rejection causes fetal injury without direct pathogen invasion. We show this occurs using both reduced-dosages of virulent or weakened Listeria for infection. By comparing pregnancy outcomes and infection-induced changes in the mother that make it more or less hospitable for the fetus, we further demonstrate Listeria gaining entry inside infected cells is the critical factor for immune-mediated fetal injury. These results illustrating infection-induced changes in the mother that lead to fetal injury have important implications for designing new ways to improve the outcomes of pregnancy.
| Listeria monocytogenes (Lm) is a ubiquitous human pathogen with a unique predisposition for invasive disseminated infection during pregnancy that represents a significant etiology of spontaneous abortion, stillbirth, and neonatal infection [1], [2]. Although many Lm-specific proteins required for cell entry and maintaining residence within infected cells have been identified, and some play important roles in placental cell invasion [3], [4], the interplay between Lm and maternal immune cells that sustain fetal tolerance in the pathogenesis of infection-induced fetal injury has not been well-characterized. Recently, the physiological accumulation of immune suppressive maternal Foxp3+ regulatory CD4 T cells (Tregs) during gestation was shown to compromise host defense against Lm and other pathogens that cause prenatal infection [5]. Nevertheless, despite increased infection susceptibility, the sustained expansion of maternal Tregs was more essential because even transient partial ablation to baseline levels was sufficient to disrupt fetal tolerance and trigger fetal resorption [5], [6], [7]. These findings in mice recapitulate the blunted expansion of maternal Tregs with spontaneous abortion and other complications associated with fractured fetal tolerance in human pregnancy [8], [9], [10], [11], [12]. Thus, healthy pregnancy requires the sustained accumulation of immune-suppressive maternal Tregs that maintains tolerance to the developing fetus.
In addition to these quantitative changes, fluctuations in Treg suppressive potency also occur. These shifts fine-tune the delicate fluid balance between immune stimulation and suppression. In particular within the first few days after infection in non-pregnant mice with Lm or other pathogens that primarily cause acute infection, progressive reductions in either Treg number or their suppressive potency have each been described [13], [14]. Similarly, immune activation that coincides with pathogen clearance during more persistent infections occurs with either experimental Treg ablation or naturally occurring reductions in Treg suppressive potency [15], [16], [17], [18]. Accordingly, infection-induced reductions in Treg suppression are likely essential prerequisites for unleashing the activation of immune effectors that efficiently eradicate infection [19]. Importantly however, how pregnancy-expanded Tregs impact infection-induced shifts in suppression, and reciprocally how infection-induced shifts in maternal Treg suppression impact fetal tolerance are each undefined. Given the substantial overlap in pregnancy complications (e.g. spontaneous abortion, stillbirth, prematurity) associated with prenatal Lm infection and disruptions in fetal tolerance induced by experimental or naturally-occurring defects in maternal Treg accumulation, we sought to investigate if infection-induced shifts in maternal Treg suppression might disrupt fetal tolerance and cause fetal wastage. Immune-mediated fetal injury triggered by maternal infection that may occur without in utero pathogen invasion could explain the only modest fraction of concepti with recoverable Lm born to mothers with invasive infection [20], and spur new approaches for improving pregnancy outcomes.
To address these questions, we investigated maternal Tregs and the maintenance of fetal tolerance using escalating dosages of virulent Lm for infection during pregnancy. Foxp3GFP reporter mice were used so that maternal Tregs could be purified based on their lineage-defining marker [21], and a mating strategy using ovalbumin (OVA)-expressing males allowed the maternal response to this surrogate fetal antigen to be precisely characterized [5], [22], [23]. To more specifically evaluate the contribution of immune-mediated fetal wastage in isolation, pregnancy outcomes, maternal Tregs, and fetal tolerance were also compared using Lm mutants lacking defined virulence determinants required for in utero invasion. Together, these studies demonstrate Lm entry into the cell cytoplasm disrupts fetal tolerance sustained by maternal Foxp3+ Tregs that triggers immune-mediated fetal wastage.
To investigate the pathogenesis of prenatal listeriosis, pregnancy outcomes after infection with escalating dosages of virulent Lm beginning midgestation were evaluated. In this regard, although Lm has been described to stimulate fetal resorption and in utero invasion during syngeneic pregnancy [24], [25], [26], [27], this mating scheme does not recapitulate the natural heterogeneity between maternal and paternal antigens, and more pronounced accumulation of maternal Tregs [5], [6], [7]. To bypass these limitations, pregnancy outcomes were enumerated using MHC mismatched strains of inbred mice (Balb/c H-2d males with C57Bl/6 H-2b females) for mating. We found Lm infection midgestation (E10.5) caused dose-dependent reductions in the number of live pups born at term ∼10 days thereafter (Figure 1A). Compared with uninfected pregnancies [8.0±0.6 live pups (mean ± standard deviation)], the number of live pups was reduced by 88% and 54% [1.0±0.5 (104 CFUs); 3.7±0.9 (103 CFUs)] for mice infected with each respective Lm dosage (Figure 1A). Furthermore, no live pups were born for the majority of pregnant mice (9 of 13) infected with 104 Lm. Thus, maternal Lm infection during allogeneic pregnancy in mice triggers fetal wastage.
To more comprehensively evaluate infection-induced fetal injury, the frequency of in utero fetal resorption at an earlier time point after infection (day 5 post-infection, E15.5) was also enumerated. Consistent with dose-dependent reductions in the number of live pups born at term, progressively increased rates of fetal wastage were found with escalating dosages of Lm used for infection (Figure 1B). Interestingly, by culturing each individual resorbed placental-fetal unit, dose-dependent increased rates of fetal Lm invasion that were uniformly reduced compared with the overall resorption frequencies were also identified. In particular, while the majority of resorbed fetuses (87%) contained recoverable Lm after infection with the highest inocula (104 CFUs), the recovery of bacteria among resorbed fetuses declined sharply with reduced inocula (14% for 103 CFUs, and 0% for 102 CFUs) (Figure 1C). Given the previously described decline in recoverable bacteria in the placenta and fetus at earlier time points (from 24 through 72 hours) after maternal infection during syngeneic pregnancy [28], we also enumerated Lm at these time points following infection with the intermediate 103 Lm dosage to investigate the possibility that the absence of recoverable bacteria in resorbed placental-fetal units could reflect Lm invasion that had been cleared by day 5. We found the placenta and fetus each contained no recoverable bacteria (among 26 and 16 individual placental-fetal units 24 and 72 hours, respectively, following maternal Lm infection midgestation) at these earlier time points (data not shown). Furthermore, the absence of direct Lm placental-fetal invasion for the majority of resorbed concepti using this intermediate 103 Lm dosage was reinforced by the absence non-viable bacteria using PCR-based detection with primers specific for Lm hly and lmo0056 among culture negative concepti (data not shown). Together, these results demonstrate although Lm has the potential for in utero invasion, especially after high-dose infection, infection-induced fetal injury can also occur without direct pathogen invasion of the placenta or fetal tissue.
Given the importance of expanded maternal Foxp3+ Tregs in maintaining pregnancy [5], [6], [7], and infection-induced reductions in Treg suppression that unleash immune activation in non-pregnant mice [13], [14], [17], [18], [19], we investigated if similar reductions in maternal Treg suppression occur with infection during pregnancy. We found Lm infection at midgestation did not significantly impact either the number or percent Foxp3+ among CD4 cells, suggesting quantitative reductions in maternal Tregs do not occur in this context (Figure 2). Next, to investigate the potential for infection-induced qualitative shifts in maternal Treg suppressive potency, Foxp3GFP reporter mice on the C57Bl/6 background were substituted for mating with Balb/c males so that maternal Foxp3+ Tregs could be purified as GFP+ CD4 cells by FACS directly ex vivo. Consistent with prior studies in non-pregnant mice [14], [21], maternal Tregs were isolated from pregnant mice with equally high purity by sorting for GFP+ CD4 cells (Figure 3A). By enumerating the efficiency whereby these GFP+ CD4 cells suppress the proliferation of responder T cells in co-culture, no significant difference in suppressive potency were found for Tregs recovered from pregnant compared with non-pregnant control mice (Figure S1 in Text S1).
By contrast with escalating Lm dosages used for infection beginning midgestation, progressive reductions in maternal Treg suppressive potency were identified because the proliferation of responder T cells isolated from naïve mice was more pronounced after co-culture with maternal GFP+ Tregs from infected compared with uninfected control mice (Figure 3B). To evaluate the magnitude of these infection-induced reductions in Treg suppressive potency, we titrated the ratio of GFP+ Tregs to responder T cells in co-culture and found two-fold increased ratios of Tregs from pregnant mice infected with 103 Lm were required to achieve the same level of suppression as Tregs cells recovered from uninfected mice (Figure 3B). Comparatively, GFP+ Tregs recovered from mice infected with 104 Lm suppressed responder cell proliferation even less efficiently; requiring two- to four-fold more Tregs to achieve the same level of suppression compared with GFP+ Tregs recovered from uninfected controls, while Tregs recovered from mice infected with 102 Lm suppressed responder cell proliferation to an intermediate degree compared with Tregs isolated from uninfected mice and those infected with higher Lm inocula (Figure 3B). Thus, Lm infection during pregnancy stimulates dose-dependent reductions in maternal Treg suppressive potency comparable to reductions in Foxp3+ CD4 cell suppressive potency with acute systemic Lm infection in non-pregnant mice [14].
Since the sustained expansion of maternal Foxp3+ Tregs is essential for maintaining tolerance to paternal antigens expressed by the developing fetus [5], [6], [7], we further investigated how these infection-induced reductions in maternal Treg suppressive potency might impact fetal tolerance. To identify maternal cells with fetal specificity, transgenic male mice engineered to express the model antigen, OVA, in all cells behind the β-actin promoter were substituted for mating with C57Bl/6 females so that maternal T cells responsive to peptides within the surrogate fetal-OVA antigen can be tracked using established immunological tools [5], [22], [23]. Following Lm infection at midgestation, we found maternal CD8 T cells with fetal-OVA specificity accumulated and became activated in a dose-dependent fashion (Figure 4). Specifically, compared with the few OVA-specific cells recovered from uninfected pregnant mice that produced only background levels of IFN-γ, fetal-OVA-specific T cells expanded over 50-fold and efficiently produced IFN-γ in pregnant mice infected with 104 Lm (Figure 4). Although the degree of expansion and cytokine production each progressively diminished with reduced Lm inocula, both remained significantly elevated compared with background levels found in uninfected pregnant mice. Together, these results demonstrate Lm infection during pregnancy blunts maternal Treg suppression and disrupts fetal tolerance. Furthermore, since even transient partial reductions in maternal Treg numbers cause resorption and fractures fetal tolerance [5], fetal wastage that occurs without direct invasion of the placental-fetal unit with lower Lm inocula are likely triggered by infection-induced reductions in maternal Treg suppressive potency.
To more specifically investigate the pathogenesis of immune-mediated fetal injury that occurs with infection-induced disruption in maternal-fetal tolerance, we compared pregnancy outcomes after infection with attenuated Lm containing defects in defined virulence determinants required for productive infection that do not cause fetal invasion [3], [4], [26]. These include LmΔLLOΔPLC that cannot escape from the endocytic vacuole and enter into the cell cytoplasm; and LmΔactA that enters the cell cytoplasm, but cannot recruit actin required for intra- and inter-cellular spread [29]. In particular, these mutants were chosen because their ability to stimulate protective T cells in vivo that requires overriding Treg suppression is drastically discordant; LmΔactA readily primes the expansion of protective T cells, whereas LmΔLLOΔPLC does not [30], [31], [32]. Consistent with robust immune activation that occurs with Lm entry into the cell cytoplasm [33], LmΔactA infection midgestation caused sharp reductions in the number of live pups with reciprocal increased rates of fetal resorption compared with uninfected controls (Figure 5A and B). By contrast, the number of live pups and frequency of fetal resorption did not differ significantly between pregnant mice infected with LmΔLLOΔPLC and non-infected controls (Figure 5A and B). Importantly, these differences in pregnancy outcomes could not be attributed to potential differences in relative attenuation between these two Lm strains because 10-fold more LmΔLLOΔPLC compared with LmΔactA was used for infection, and at these dosages no significant difference in bacterial CFUs were found one day post-infection in the liver representing another tissue susceptible to Lm invasion (Figure S2 in Text S1). Furthermore, consistent with the highly attenuated nature of these mutants, LmΔactA and LmΔLLOΔPLC each became eradicated from the liver, and could not be recovered from any resorbed placental-fetal units by day 5 post-infection (data not shown). Together, these results demonstrate Lm cytoplasmic entry is essential for infection-induced fetal wastage.
Given reductions in Treg suppressive potency associated with fetal resorption and fractured fetal tolerance after virulent Lm infection (Figures 1, 3, and 4), we investigated if differences in fetal wastage induced by LmΔactA and LmΔLLOΔPLC also paralleled discordance efficiencies in dampening maternal Treg suppression and disrupting fetal tolerance. GFP+ Tregs could be purified efficiently from pregnant Foxp3GFP reporter mice after LmΔactA and LmΔLLOΔPLC infection similar to mice after WT Lm infection or uninfected controls (Figure 6A). Using purified GFP+ Tregs, we found LmΔactA infection midgestation triggered ∼2-fold reductions in suppressive potency for maternal Tregs compared with GFP+ CD4 cells recovered from uninfected pregnant controls (Figure 6B). Interestingly, the magnitude of these reductions in Treg suppressive potency were almost identical to cells recovered from mice infected with low or intermediate WT Lm dosages that induce fetal resorption with minimal to undetectable in utero invasion. By contrast, the suppressive potency for maternal GFP+ Tregs after LmΔLLOΔPLC infection did not differ significantly compared with cells from uninfected controls (Figure 6B). Thus, reductions in maternal Treg suppressive potency directly parallel fetal wastage induced by LmΔactA and LmΔLLOΔPLC.
To further establish how these reductions in maternal Treg suppression induced by attenuated Lm impact fetal tolerance, the expansion and activation of maternal T cells with specificity to the surrogate fetal-OVA antigen were also enumerated after infection in pregnant C57Bl/6 mice mated with OVA-expressing Balb/c males. Similar to WT Lm, LmΔactA inoculated midgestation primed the robust expansion and IFN-γ production by fetal-OVA specific T cells (Figure 7). Comparatively, LmΔLLOΔPLC failed to stimulate expansion and IFN-γ production above background levels found in uninfected control mice. These findings directly parallel the relative efficiency whereby each attenuated Lm dampens maternal Treg suppressive potency and induces fetal wastage (Figures 5 and 6). Lastly, to more definitively establish infection-induced immune-mediated fetal injury, the impacts of maternal CD4 and CD8 T cell depletion prior to Lm infection on pregnancy outcomes were evaluated. These experiments exploit the highly attenuated nature of LmΔactA that is eliminated even in immune-compromised mice to investigate how depletion of effector and regulatory T cells together impact infection-induced fetal wastage. Remarkably, the rate of LmΔactA-induced fetal wastage became sharply reduced in T cell-depleted compared with T cell-sufficient pregnancies [11.2±3.5% resorption in T cell depleted mice (n = 9); 62.7±9.5% resorption in T cell-sufficient mice (n = 15), P = 0.0005]. Reciprocally, the number of live pups born with LmΔactA infection initiated midgestation was significantly increased in T cell-ablated compared with T cell sufficient pregnancies [5.80±0.49 live pups in T cell depleted mice (n = 10) compared with 3.68±0.60 live pups in T cell-sufficient pregnancies (n = 19), P = 0.027]. Taken together, these results demonstrate Lm cytoplasmic entry is essential for disrupting fetal tolerance that triggers maternal T cell-mediated fetal wastage.
The intracellular bacterium Lm represents a significant infectious cause of pregnancy loss and stillbirth [1], [2]. Herein, we investigate the host pathogen interaction that triggers these unfortunate outcomes, with particular focus on how prenatal infection impacts fetal tolerance sustained by expanded maternal Foxp3+ Tregs. Our results demonstrate infection-induced dampening of Treg suppression that unleashes immune activation required for optimal host defense [19], in the context of prenatal infection when sustained tolerance to fetal antigen is essential, plays a pivotal role in the “immune-pathogenesis” of fetal wastage. The importance of immune-mediated fetal injury is shown by the increased frequency of fetal resorption with reciprocal reduction in the number of live pups following infection with low or intermediate doses of virulent Lm where bacteria are not found in the majority of resorbed concepti. Similarly for attenuated Lm that do not cause fetal invasion [26], fetal wastage and disrupted fetal tolerance occurs only for strains that retain the ability to prime protective T cells through entry into the cell cytoplasm and dampen maternal Treg suppressive potency. Furthermore, depletion of effector T cells along with Tregs prior to cytoplasmic Lm infection sharply reduces the frequency of infection-induced fetal wastage. Together with fetal injury induced by systemic treatment with various TLR ligands [e.g. LPS, poly(I∶C)] shown to modulate Treg suppression in vitro or after in vivo stimulation in non-pregnant mice [34], [35], [36], [37], [38], [39], these results establish the importance of infection- or inflammation-induced disruption of maternal Treg suppression in the pathogenesis of fetal wastage. Although we used intravenous inoculation to recapitulate disseminated infection that occurs with Lm during pregnancy, crucial next steps based on these results are to evaluate if systemic disruption of fetal tolerance is essential, or if local disruption induced by pathogens that primarily reside in the vaginal or cervical mucosa (e.g. bacterial vaginosis, Ureaplasma and Chlamydia sp.) are also sufficient to stimulate fetal injury.
On the other hand, since Lm and other prenatal pathogens can and do cause in utero fetal invasion, immune-mediated fetal injury alone does not fully address the pathogenesis of these infections. In this regard, an important clue from our studies is that although virulent Lm-induced fetal resorption and invasion are each dose-dependent, these processes can be readily dissociated based on the inocula of Lm used for infection. Accordingly, we propose a model whereby low-dose maternal infection dampens Treg suppression enough to stimulate the activation of immune effectors that rapidly eliminate the pathogen so that fetal injury occurs almost exclusively via immune-mediated pathways (Figure 8). Comparatively, with higher-dose infection, blunted maternal Treg suppression that promotes immune activation does not eradicate infection as efficiently. In turn with ongoing disruptions in fetal tolerance, remaining pathogen drawn to inflammation at the uterine-placental interface promotes invasion into the placental-fetal unit (Figure 8). Although clearly an over-simplification, this model suggests overriding maternal Treg suppression is the pivotal feature that dictates whether fetal wastage occurs regardless of in utero pathogen invasion.
Together with prior studies using pregnant guinea pigs where the placental anatomy more closely resembles human tissue that illustrate the placenta is relatively resistant to early infection [40], [41], our findings suggest Lm infection-induced disruption of fetal tolerance with ensuing inflammation at the maternal-placental interface likely represents an important initial step in targeting bacteria for placental invasion. Later after infection when the placenta becomes a possible nidus for ongoing bacterial dissemination, expulsion or resorption of the placental-fetal unit by maternal immune effectors unleashed for activation by reduced Treg suppression provides expanded protection to the mother from infection at the expense of fetal injury [40], [41]. Based on these results, important areas for further investigation are to identify the cellular receptors and secreted cytokines that respond to Lm, and in particular bacteria in the cell cytoplasm, that stimulate reductions in maternal Treg suppression. Given the cellular immunology and transgenic tools that are currently available only in mice, our ongoing studies continue to address these questions experimentally using murine pregnancy models. However, additional investment in developing more refined immunological tools in other complementary animal models will be important for extending this work to more fully unravel the mechanistic steps in fetal wastage triggered by prenatal infection.
Finally, given the increasingly established heterogeneity and functional specialization among Foxp3+ CD4 cells that utilize distinct cell-associated and secreted molecules to mediate context specific immune suppression [42], [43], [44], establishing the Treg-associated molecule(s) that sustain fetal tolerance and distinguishing them from those required for host defense against infection have salient implications for developing therapies for dissociating the beneficial and detrimental impacts of expanded maternal Tregs. In this regard, while IL-10 is non-essential for sustaining fetal tolerance under non-inflammatory conditions, it likely plays more important roles in maintaining pregnancy under inflammatory conditions known to trigger fetal wastage [5], [35], [45], [46]. Accordingly, establishing the importance of other Treg-associated molecules in maintaining fetal tolerance and sustaining pregnancy represent other important areas for future investigation. Given the sharply increased rates of fetal wastage and pregnancy loss that occurs with prenatal Lm infection, the specific Treg-associated molecules essential for maintaining pregnancy may overlap with those required for optimal protection against prenatal infection. Nevertheless, given the importance of maternal Tregs in both sustaining fetal tolerance and compromising host defense against prenatal infection, we propose establishing how these cells work in each context represent critical next steps towards new therapeutic approaches for improving pregnancy outcomes.
This study was carried out in accordance with recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. These specific protocols were approved by the University of Minnesota Institutional Animal Care and Use Committee (Animal Welfare Assurance Number A3456-01).
C57Bl/6 (H-2b), Balb/c (H-2d), and B6.PL-Thy1 (CD90.1) mice were purchased from The National Cancer Institute or The Jackson Laboratory. Foxp3GFP mice backcrossed to C57Bl/6 mice, OVA-expressing backcrossed to Balb/c mice, and OT-I TCR transgenic mice maintained on a CD90.1 background have been described [21], [22], [23], [47]. The timing of pregnancy was determined by visualization of a copulation plug (embryonic day 0.5) after introducing virgin female with male mice.
Lm strains 10403 s (WT), 1942 (ΔactA), and 2319 (ΔLLOΔPLC) were each grown to early log phase (OD600 0.1) in brain heart infusion media at 37°C, washed and diluted with saline to 200 µl, and injected intravenously via the lateral tail vein [29], [31], [32]. The inoculum was verified for each infection by plating serial dilutions onto agar plates. For enumerating recoverable Lm CFUs, each placental-fetal unit or the liver was individually dissected, homogenized in saline containing 0.05% Triton X and cultured onto agar plates as described [5]. For enumerating non-viable Lm by PCR, DNA was extracted from each placental-fetal unit after homogenization, phenol chloroform extraction, and ethanol precipitation, and used as template DNA with primers specific for Lm hly and lmo0056 [hly, 5′-TGATTCACTGTAAGCCATTTC-3′ and 5′-AGCACCACCAGCATCTCCGC-3′; lmo0056, 5′-CCAAGCGAACTACGTGATCG-3′ and 5′-TGCTCTTCTACTGCGTTTGC-3′] [48].
Fluorophore-conjugated antibodies and other reagents for cell surface, intracellular cytokine, and intranuclear Foxp3 staining were purchased from eBioscience or BD Biosciences. The expansion of fetal-OVA-specific CD8 cells among splenocytes, and cytokine production after stimulating splenocytes with OVA257–264 peptide in media containing GolgiPlug (BD Biosciences) was enumerated using previously described procedures [5]. Specifically, for cell transfers, purified CD8+ cells (105) isolated from OT-I TCR transgenic (CD90.1) mice were injected intravenously into recipient (CD90.2) mice one day prior to Lm infection at midgestation. For T cell depletion, purified anti-mouse CD4 (GK1.5) and anti-mouse CD8 (2.43) antibodies (BioXcell) were administered intraperitoneally (500 µg each antibody per mouse) one day prior to LmΔactA infection at midgestation.
For enumerating Treg suppressive potency, CD4 cells were first enriched by negative selection (Miltenyi Biotec) from Foxp3GFP reporter mice, followed by sorting for the GFP+ CD4 subset [21]. In each experiment, GFP+ Tregs were verified to be >98% pure by staining for Foxp3 expression. Responder CD8+ T cells isolated from naïve CD90.1 mice were labeled with CFSE (5 µM for 10 minutes at room temperature), and co-cultured in triplicate in 96-well round bottom plates (1×104 responder cells per well) with purified GFP+ Tregs at the indicated ratios. The relative suppressive potency of Tregs in each experiment was calculated by comparing responder cell proliferation (CFSE dilution) after co-culture with GFP+ Tregs from uninfected control mice as described [14], [18].
The number of live pups, resorbed concepti, cell numbers, and percent cytokine producing cells were first analyzed and found to be normally distributed. Thereafter, differences between groups were analyzed using an unpaired Student's t test (Prism, Graph Pad) with P<0.05 taken as statistical significance.
|
10.1371/journal.pgen.1005896 | Inferring Phylogenetic Networks with Maximum Pseudolikelihood under Incomplete Lineage Sorting | Phylogenetic networks are necessary to represent the tree of life expanded by edges to represent events such as horizontal gene transfers, hybridizations or gene flow. Not all species follow the paradigm of vertical inheritance of their genetic material. While a great deal of research has flourished into the inference of phylogenetic trees, statistical methods to infer phylogenetic networks are still limited and under development. The main disadvantage of existing methods is a lack of scalability. Here, we present a statistical method to infer phylogenetic networks from multi-locus genetic data in a pseudolikelihood framework. Our model accounts for incomplete lineage sorting through the coalescent model, and for horizontal inheritance of genes through reticulation nodes in the network. Computation of the pseudolikelihood is fast and simple, and it avoids the burdensome calculation of the full likelihood which can be intractable with many species. Moreover, estimation at the quartet-level has the added computational benefit that it is easily parallelizable. Simulation studies comparing our method to a full likelihood approach show that our pseudolikelihood approach is much faster without compromising accuracy. We applied our method to reconstruct the evolutionary relationships among swordtails and platyfishes (Xiphophorus: Poeciliidae), which is characterized by widespread hybridizations.
| Phylogenetic networks display the evolutionary history of groups of individuals (species or populations) including reticulation events such as hybridization, horizontal gene transfer or migration. Here, we present a likelihood method to learn networks from molecular sequences at multiple genes. Our model accounts for several biological processes: mutations, incomplete lineage sorting of alleles in ancestral populations, and reticulations in the network. The likelihood is decomposed into 4-taxon subsets to make the analyses scale to many species and many genes. Our work makes it possible to learn large phylogenetic networks from large data sets, with a statistical approach and a biologically relevant model.
| Evolutionary relationships are typically visualized in a tree, which implicitly assumes vertical transfer of genetic material from ancestors to descendants. However, not all species follow this paradigm. If genes can be horizontally transferred between some organisms, a tree is not a good representation of their history. Such reticulate events include hybridization, horizontal gene transfer or migration with gene flow, and require methods to infer phylogenetic networks. While a great deal of research has flourished for the inference of phylogenetic trees from different types of data, methods to infer phylogenetic networks are still limited and under development.
There are mainly two kinds of phylogenetic networks: implicit and explicit. Implicit networks–also called split networks–describe the discrepancy in gene trees, or other sources of data, and methods are well developed to reconstruct these networks [1–4]. These methods tend to be fast. However, implicit networks lack biological interpretation as the internal nodes do not represent ancestral species. Explicit networks, on the other hand, represent explicit reticulation events and each node represents an ancestral species. Combinatorial methods to infer explicit networks (which we call phylogenetic networks here) are fast but ignore gene tree error and incomplete lineage sorting (ILS) as a possible source of gene tree discordance (e.g. [5]). Model-based methods are most accurate but can be computationally challenging. They calculate the likelihood of an observed gene tree given a species network taking into account both reticulation and ILS [6–8]. Their scope was expanded in [9] to search for the most likely phylogenetic network based on multi-locus data (see also [10] for a different likelihood framework, where sites instead of genes are treated as independent and ILS is ignored). The likelihood-based method in [9], implemented in PhyloNet, provides a solid theoretical framework to estimate the maximum likelihood phylogenetic network from a set of gene trees. It has several advantages: it incorporates uncertainty on the gene trees estimated from sequence data, accounts for a background level of gene tree discordance due to ILS, and controls the complexity of the network with a cross validation step. However, its likelihood computation is heavy and becomes intractable when increasing the number of taxa or the number of hybridizations, making this method practical for small scenarios of up to about 10 species and 4 hybridizations in the network.
Here, we provide a fast statistical method to estimate phylogenetic networks from multi-locus data. We first present the theory for the pseudolikelihood of a network. We do so by deriving the proportion of the genome that has each 4-taxon tree (quartet concordance factors) as expected under the coalescent model extended by hybridization events, and we prove the generic identifiability of the model. We then use the observed quartet concordance factors as inferred from the multi-locus data to estimate the species network. Our method SNaQ (Species Networks applying Quartets) is implemented in our open-source software package PhyloNetworks in Julia and publicly available at https://github.com/crsl4.
Like PhyloNet, our method can incorporate uncertainty in estimated gene trees and gene tree discordance due to ILS. Our pseudolikelihood has computational advantages. It is simpler and more scalable to many species, compared to the full likelihood. It also scales to a large number of loci because estimation of gene trees can be highly parallelized, then summarized by only 3 tree frequencies on each 4-taxon subsets used as input in the pseudolikelihood. In simulations, our method showed good performance and scaled to scenarios for which PhyloNet could not run. We also used SNaQ to infer the evolutionary relationships between Xiphophorus fishes, from 1,183 loci across 24 taxa. Our results were congruent with [11] and refined the placement of some hybridizations found in that study. The analyses here presented show that SNaQ can enable scientists to incorporate organisms to the “tree of life” in parts that are more net-like than tree-like, and thus, complete a broader picture of evolution.
Intuitively, a phylogenetic network is a phylogenetic tree with added hybrid edges, causing some nodes to have two parents (but see [12]). Phylogenetic networks can describe various biological processes causing gene flow from one population to another such as hybridization, introgression, or horizontal gene transfer. Hybridization occurs when individuals from 2 genetically distinct populations interbreed, resulting in a new separate population. Introgression, or introgressive hybridization, is the integration of alleles from one population into another existing population, through hybridization and backcrossing. Genes are horizontally transferred when acquired by a population through a process other than reproduction, from a possibly distantly related population. Although these three processes are biologically different, we do not make the distinction when modeling them with a network. In other words, our model takes into account all three biological scenarios, but those scenarios are not distinguishable in the estimated phylogenetic network unless more biological information is provided.
Just like phylogenetic trees, networks can be rooted or unrooted. A rooted phylogenetic network on taxon set X is a connected directed acyclic graph with vertices V = {r} ∪ VL ∪ VH ∪ VT, edges E = EH ∪ ET and a bijective leaf-labeling function f: VL → X with the following characteristics. The root r has indegree 0 and outdegree 2. Any leaf v ∈ VL has indegree 1 and outdegree 0. Any tree node v ∈ VT has indegree 1 and outdegree 2. Any hybrid node v ∈ VH has indegree 2 and outdegree 1. A tree edge e ∈ ET is an edge whose child is a tree node. A hybrid edge e ∈ EH is an edge whose child is a hybrid node. Unrooted phylogenetic networks are typically obtained by suppressing the root node and the direction of all edges. We also consider semi-directed unrooted networks, where the root node is suppressed and we ignore the direction of all tree edges, but we maintain the direction of hybrid edges, thus keeping information on which nodes are hybrids. The placement of the root is then constrained, because the direction of the two hybrid edges to a given hybrid node inform the direction of time at this node: the third edge must be a tree edge directed away from the hybrid node and leading to all the hybrid’s descendants. Therefore the root cannot be placed on any descendant of any hybrid node, although it might be placed on some hybrid edges.
We further assume that the true network is of level-1[1], i.e. any given edge can be part of at most one cycle. This means that there is no overlap between any two cycles (but see the Discussion). Refer to [1] for other types of evolutionary networks. Throughout this work, we denote by
For example, in Fig 1 (center) n = 7, h = 2, k1 = 3 and k2 = 4. The main parameter of interest is the topology N of the semi-directed network. Like phylogenetic trees, this network can be rooted by a known outgroup. The other parameters of interest are t, the vector of branch lengths in coalescent units (see below), and a vector of inheritance probabilities γ, describing the proportion of genes inherited by a hybrid node from one of its hybrid parent (see Fig 1). Only identifiable branch lengths are considered in t. For example, with only one sequenced individual per taxon, the lengths of external edges are not identifiable and are not estimated.
The input for our method is a table of quartet CFs observed from multi-locus data (the X values in Eq (1)), across many or all 4-taxon subsets from the n taxa of interest.
We carried out simulations to compare the speed and accuracy of SNaQ and PhyloNet. Given that PhyloNet uses the rooted and full gene trees, SNaQ can only be expected to perform as accurately as PhyloNet at best. Our simulations show that a pseudolikelihood approach does not compromise too much accuracy, but greatly improves speed.
We simulated g gene trees with ms [30] under four different networks: (n, h) = (6, 1), (6, 2), (10, 1) and (15, 3), with γ values set to 0.2 or 0.3 on each minor hybrid edge (see S1 Text) These network topologies were chosen at random by simulating a tree with n taxa under the coalescent, then choosing two edges at random for the origin and target of each hybridization and rejecting networks of level >1. On 6 taxa all reticulations were hard to reconstruct with k = 4, including a bad diamond I in the case h = 2. On 10 and 15 taxa, both networks also had a diamond, of the bad type II for n = 10. We varied the number of genes between 10 and 3000. All analyses were run on 2.7–3.5 GHz processors.
We first used the true simulated gene trees for inference. The rooted gene trees served as input for PhyloNet and the unrooted quartet CFs as observed in the g gene trees served as input for SNaQ. The semi-directed network returned by SNaQ was rooted by the outgroup species, when compatible with the estimated hybrid edges. Next, we used Seq-Gen [31] to simulate sequences of length 500 under HKY, κ = 2, A, C, G and T frequencies of 0.300414, 0.191363, 0.196748, 0.311475 and population mutation rate θ = 0.036, as in [9]. Gene trees were estimated with MrBayes [28] using 106 generations sampled every 200, 25% burnin and an HKY model. The consensus trees (one per gene) served as input for PhyloNet. The posterior tree samples were then used in BUCKy [26, 27] for each 4-taxon set, to estimate quartet CFs and use them as input for SNaQ. For this pipeline, we used the tools implemened by [32] and available at https://github.com/nstenz/TICR. This procedure was replicated 30 times. The accuracy of each method was measured as the proportion of times that the estimated network matched the true network. To compare rooted networks we used the distance in [33], which is a metric on reduced networks (including level-1 networks) and is implemented in PhyloNet. We used it to detect equality between rooted networks, but not to measure how “close” networks were, because this distance is very sensitive to small differences such as a change in the direction of a hybrid edge.
Fig 6 summarizes the accuracy and speed of SNaQ and PhyloNet. On 10 or 15 taxa PhyloNet was too slow to run (a single replicate with 10 taxa and 300 loci required over 400 hours), so we cannot provide a comparison of accuracy on these 2 larger networks.
For networks with h = 2 or more, the accuracy of SNaQ decreased. So, for each semi-directed network estimated by SNaQ, we determined if its unrooted topology matched that of the true network. Fig 7 shows that in the vast majority of cases when the directed network was incorrectly estimated, its unrooted topology was still correctly inferred from true gene trees and for n = 6 with estimated gene trees. For n ≥ 10, the inferred direction of hybrid edges degraded when gene trees were estimated. In most replicates on 10 taxa, this was because the bad diamond II near the root in the true network had a wrong estimated placement of the hybrid node.
To detemine which features in the network were correctly estimated, we extracted the major tree from each network, that is, the tree obtained by keeping the major hybrid edge and suppressing the minor hybrid edge at each hybrid node. We then compared the true major tree (from the true network) to the estimated major tree using the Robinson-Foulds distance (see Fig 8). The major tree was correctly estimated from 300 or more genes in all scenarios, except when n = 6, h = 2 and 300 genes (1 replicate out of 30) and 1000 genes (1 replicate out of 30). In both cases, the true major tree was displayed in the estimated network but the major hybrid edge was estimated as a minor edge with γ < 0.5. Therefore, the network’s “backbone”, i.e. the major vertical inheritance pattern, can still be estimated accurately even when the full network and hybrid edges are not (Fig 7).
Among cases when the major tree was correctly estimated, we determined the detection accuracy of each true hybridization event. To do so, we compared each estimated hybridization with the true hybridization of interest. In each network (true and estimated), we removed the other hybridizations by suppressing their minor hybrid edges and used the known outgroup to root both networks. We then calculated the hardwired cluster distance between the two resulting networks to determine if the estimated hybridization event matched the true hybridization of interest: connecting the same donor edge to the same recipient edge in the major tree (Fig 9). For n = 6, the hybridizations forming a good diamond were recovered with high accuracy from 100 genes, but the hybridization forming a bad diamond I (case h = 2) was very hard to recover, needing more than 1000 genes for an accurate inference of the hybrid edges’ direction. Still, the unrooted cycle was correctly estimated from 100 genes or more. For n = 10 and n = 15 taxa, the hybridization creating a cycle of k = 4 nodes was also very hard to detect with its correct direction, although its undirected cycle was accurately recovered from a few hundred genes. Hybridizations were recovered more accurately as their cycles spanned more nodes, with a high recovery rate for the hybridizations with k = 6 and k = 7 from 100 genes or more.
We re-analyzed transcriptome data from [11] to reconstruct the evolutionary history of 24 swordtails and platyfishes (Xiphophorus: Poeciliidae). Based on high CFs of splits in conflict with their species tree followed by a series of ABBA-BABA tests [35], [11] concluded that hybridization or gene flow was widespread in the history of these tropical fishes. We re-analyzed their first set of 1183 transcripts. BUCKy was performed on each of the 10,626 4-taxon sets. The resulting quartet CFs were used in SNaQ, using hm = 0 to 5 and 10 runs each. The network with h = 0 and the major tree in the network with h = 1 were identical to the total evidence tree in [11], with X. xiphidium placed within the grade of southern platyfishes (SP), making the northern platyfishes (NP) paraphyletic (see S1 Text). With h ≥ 2 the major tree was almost identical but with NP monophyletic (Fig 10) because X. xiphidium was found sister to the rest of the NP species, but involved in a reticulation (see below). With h ≥ 3, a reticulation within the southern swordtails (SS) was found consistently (γ = 0.43), but with a direction in conflict with SS being an outgroup clade. Its cycle had only k = 5 nodes, 4 of them leading to a single taxon (see S1 Text) so we suspect an error in the inferred hybrid node and gene flow direction. The extra 2 reticulations found with h = 4 and 5 had low γ values (in [0.006–0.16]).
The network scores (negative log-pseudolikelihood) decreased sharply from h = 0 to h = 2 then slightly and somewhat linearly (see S1 Text), suggesting that h = 2 best fits the fish data using a slope heuristic [45, 46]. The network estimated with h = 2 (Fig 10) found X. xiphidium involved in an ancient reticulation, contributing a proportion γ = 0.17 of genes to the lineage ancestral to northern swordtails (NS). This reticulation might explain the placement of X. xiphidium closer to the root in [11], from tree-based methods that do not account for potential gene flow. The second hybridization (γ = 0.20) was found from the population ancestral to X. multilineatus and X. nigrensis into X. nezahuacoyotl, and relates to a high CF found by [11] for a clade uniting X. nezahuacoyotl and the nigrensis group.
Bootstrap data sets were simulated by sampling each quartet CF from a uniform distribution on its 95% credibility interval (conservatively) then normalizing the sampled CFs across the 3 quartets on each 4-taxon set. For each bootstrap data set we estimated a network using 3 runs, and h = 3 (instead of 2) because the third inferred reticulation had a high γ (see S1 Text) and to assess the ability of the bootstrap procedure to identify the best h value. If the bootstrap was consistent with the slope heuristic, we expected high bootstrap support for the placement of the first 2 reticulations and lower support for the third. As expected, this third reticulation and network topology within the SS clade was variable among bootstrap networks (see S1 Text), suggesting uncertainty in the major tree within this clade (Fig 10). The rest of the tree was highly supported, as was the placement of the reticulation involving X. xiphidium. The reticulation involving X. nezahuacoyotl had split support for its donor lineage, with 75% support for a more ancestral lineage (Fig 10).
Many methods are being developped to understand organisms whose evolution behaves more net-like rather than tree-like. There is evidence of reticulation at all levels in the tree of life: deep among early prokaryotic and eukaryotic groups, to shallow among recently diverged species (e.g. [36–38]) or even among populations of the same species. Our new and fast statistical method to infer phylogenetic networks from multi-locus data could be used at these various levels in the tree of life.
Network inference is theoretically and computationally challenging. Split networks can be estimated rapidly, yet lack an evolutionary model and biological interpretability. [39] proposed a very fast distance-based approach to reconstruct topological ancestral recombination graphs (tARGs) from a long alignment, but the biological interpretability of tARGs is still limited. The evolution model in [8] uses an explicit network and satisfyingly accounts for various processes: reticulation events, deep coalescences, and substitutions. Yet a full likelihood estimation of large network (as in [9]) seems beyond computational reach. Our pseudolikelihood method offers an alternative, allowing the estimation of bigger and more complex networks while maintaining biological interpretability and a flexible evolutionary model.
We assumed a level-1 network throughout, where each hybrid node is part of a single cycle. This assumption is quite restrictive, but [40] showed that sequence data and gene trees on present-day species do not contain enough information to reconstruct complex networks, even from many loci. Therefore, some assumption has to be made to limit the network complexity. Extending our method to networks with intersecting cycles will need further work to restrict the search to candidate networks that are distinguishable from each other. Indeed, [40] show that different level-2 networks can have the exact same likelihood, and hence pseudolikelihood. So no method based on gene trees can ever decide which of these level-2 networks is true. Under a model without ILS, using full gene trees and branch length in substitutions per sites comparable across genes, [40] showed that level-1 networks are distinguishable but level-2 networks are not necessarily. Extending our approach to higher level networks, with or without ILS, will require extensive theory to work around this lack of identifiability.
Our approach allows for multiple individuals per species. All alleles from the same species simply need to be treated as a known and fixed polytomy in the network. Future work could include this and other topology constraints on the network, to reduce the computational burden when there are known phylogenetic relationships.
We allow hybrid edge lengths to be 0, but we do not constrain them to be 0 (unlike in [6, 8]) even though each gene flow event has to occur between contemporary populations. If one parental population went extinct or has no sampled descendants, the hybrid edge from this parent has a positive length in the observable network. A second reason is that a long branch can fit a population bottleneck, as might be expected in the formation of a new hybrid species. Not constraining hybrid branch lengths to 0 has a computational burden, however. Future implementations might enforce this constraint, when taxon sampling is thorough and extinction of parental populations can be ruled out.
By considering quartet topologies only, we ignored branch lengths in gene trees. This choice frees us from various assumptions. Using gene tree branch lengths, which are in substitutions per site, would require some assumption on gene rates to make branch lengths comparable across trees, and a molecular clock on gene trees. Other assumptions would also be needed on population sizes, shared or not across lineages. The recent approach in [41] should scale well to many taxa, but makes these strong assumptions because it requires accurate distances obtained from branch lengths in gene trees. On the contrary, our approach should be robust to rate variation across genes and across lineages, and does not require any assumption on population sizes.
Yu et al. [8] already noted a lack of identifiability from rooted gene trees for reticulations with k = 3 from only 4 taxa (including the outgroup). We found a similar lack of identifiability from unrooted quartets if n < 5. In practice, some reticulations are hard to detect even with 5 or more taxa, if some branches are long with no ILS (close to violating A1). However, in these cases the unrooted topology of the network can still be recovered, even if the direction of gene flow and the placement of the hybrid node is not. Therefore, heuristic strategies that keep the unrooted network unchanged, or that just slightly modify it, may improve the search for the best network.
More tools are needed to study unrooted and semi-directed phylogenetic networks. For instance, no distance measure has been developed for such networks, that we know of. Distances between rooted networks would also be needed, that would be less sensitive to small changes in the unrooted or semi-directed topologies than the distance proposed in [33]. New notions of edge equivalence would also be needed on unrooted and semi-directed networks. It would help summarize a bootstrap sample of networks for instance, with no need for an outgroup.
We propose here a tree-based but informative summary by extracting the major tree from each network, obtained by dropping any minor hybrid edge (with inheritance γ < 0.5). Because this tree summarizes the major vertical inheritance pattern at each node, it can be considered an estimate of the species tree. We found that recovering the underlying species tree can be much easier (requiring fewer genes) than recovering the horizontal signal. Even if the species tree is the main purpose of a study, [34] showed that species-tree methods can be inconsistent in recovering the vertical signal if there is gene flow, so using a network can be beneficial to avoid the possible inconsistency of tree-based coalescent methods.
All data analyzed here had full taxon sampling from each gene, and we were able to use all 4-taxon sets. Future work could assess the impact of missing data (gene sequences, or 4-taxon sets) on the method’s accuracy. Missing 4-taxon sets will be necessary for large networks, because the number of 4-taxon sets grows very rapidly with the number of taxa (∼n4/24). With many taxa, one may randomly select a collection of 4-taxon sets and/or choose them specifically. SNaQ calculates the number of quartets involving each taxon and provides information about under-represented taxa, if any. With many individuals per species, one may greatly reduce the collection of 4-taxon sets to be analyzed by randomly sampling from those containing at most one individual per species. If the assignment of individuals to species is correct, any 4-taxon set containing 2 individuals from the same species would be non-informative about the species-level relationships. This strategy is used in [42] to infer species trees under ILS.
Model selection is necessary to estimate the number of hybridizations h, because the pseudolikelihood is bound to improve as h increases, like the likelihood or parsimony score in [43]. We used here the log pseudolikelihood profile with h. A sharp improvement is expected until h reaches the best value and a slower, linear improvement thereafter. Such data-driven slope heuristics can indeed be used with contrast functions (like pseudolikelihoods) for model selection in regression frameworks [45, 46].
Information criteria have already been used to select h (e.g. [44]), but these criteria are inappropriate if the full likelihood is replaced by a pseudolikelihood. Theory is missing to compare the pseudolikelihoods of different networks, because of the possible correlation between quartets from different 4-taxon sets. It can be shown, however, that quartets from two 4-taxon sets s1 and s2 are independent if s1 and s2 overlap by at most one taxon and if the true 4-taxon subnetworks share no internal edges. Future work could exploit this partial independence to construct hypothesis tests.
Cross validation has been proposed by [9], and was shown to have good performance. In our framework, the cross-valication error could be measured from the difference between the quartet CFs observed in the validation subset and the quartet CFs expected from the network estimated on the training set. Because K-fold cross-validation requires partitioning the loci into K subsets and re-estimating a network K times at each h value, this approach can be computationally heavy.
Finally, [32] proposed a goodness-of-fit test, also based on quartet CFs, to determine if a tree with ILS fits the observed data or if a network is needed instead. This test could be extended to networks, to decide if a given h provides an adequate fit. One advantage to this approach is that testing the adequacy of a given h does not require to estimate a larger network with h + 1 hybridizations, whereas other approaches above would require estimation of both networks in order to decide that the simpler network is sufficient.
After submission, we learned about similar work using subnetworks and a pseudolikelihood approach [47], which scales to many taxa. In [47], the pseudolikelihood is based on rooted triples whereas we use unrooted quartets. There are fewer triples, so the method in [47] is potentially faster. However, fewer triples means less information. For example, the networks Ψ1 and Ψ2 shown in Fig. 2 of [47], which are not distinguishable from triplets, are in fact distinguishable from quartets (see S1 Text). Our thorough study of the network identifiability allowed us to implement a search that avoids jumping between networks that are not distinguishable, which facilitates convergence. The downside of our approach is the assumption of a level-1 network. Instead, [47] do not assume any restriction on the network. Finally, our method does not require rooted gene trees as input, which we view as a major advantage because rooting errors are avoided.
|
10.1371/journal.pntd.0005338 | Unusual pattern of chikungunya virus epidemic in the Americas, the Panamanian experience | Chikungunya virus (CHIKV) typically causes explosive epidemics of fever, rash and polyarthralgia after its introduction into naïve populations. Since its introduction in Panama in May of 2014, few autochthonous cases have been reported; most of them were found within limited outbreaks in Panama City in 2014 and Puerto Obaldia town, near the Caribbean border with Colombia in 2015. In order to confirm that Panama had few CHIKV cases compared with neighboring countries, we perform an epidemiological analysis of chikungunya cases reported from May 2014 to July 2015. Moreover, to understand this paucity of confirmed CHIKV cases, a vectorial analysis in the counties where these cases were reported was performed.
Chikungunya cases were identified at medical centers and notified to health authorities. Sera samples were analyzed at Gorgas Memorial Institute for viral RNA and CHIKV-specific antibody detection.
A total of 413 suspected cases of CHIKV infections were reported, with incidence rates of 0.5 and 0.7 per 100,000 inhabitants in 2014 and 2015, respectively. During this period, 38.6% of CHIKV cases were autochthonous with rash and polyarthralgia as predominant symptoms. CHIKV and DENV incidence ratios were 1:306 and 1:34, respectively. A phylogenetic analysis of E1/E2 genomic segment indicates that the outbreak strains belong to the Asian genotype and cluster together with CHIKV isolates from other American countries during the same period. Statistical analysis of the National Vector Control program at the district level shows low and medium vector infestation level for most of the counties with CHIKV cases. This index was lower than for neighboring countries.
Previous training of clinical, laboratory and vector workers allowed a good caption and detection of the chikungunya cases and fast intervention. It is possible that low/medium vector infestation level could explain in part the paucity of chikungunya infections in Panama.
| Chikungunya virus (CHIKV) is a mosquito borne pathogen that causes fever with rash and arthralgia, which are often confused with Dengue virus (DENV) infections. It has been reported that when CHIKV colonizes regions without previous circulation, it generally results in explosive human epidemics. In Panama, the first CHIKV infections were detected in May 2014. However, unlike many countries in the Americas, Panama presented with few autochthonous cases during the outbreak. In this study, we investigated the likely reason for the paucity of cases. Low vector infestation level, along with the surveillance programs, preparedness and early outbreak response possibly influenced the low number of cases observed during the Panamanian CHIKV outbreak.
| The chikungunya virus (CHIKV, Alphavirus, Togaviridae) is an RNA single-stranded arthropod-borne pathogen that was first recognized in 1952–1953 in southeast Tanzania and northern Mozambique [1]. In Africa, CHIKV was associated with small outbreaks in rural areas, however, in Asia between 1960–1970, the virus was associated with explosive urban epidemics [2]. Based on the geographical distribution and genetic profiles, CHIKV is divided in three major lineages: the East, Central and South African (ECSA), the West African and the Asian lineages [3]. After a 2004 epidemic in coastal Kenya, CHIKV spread to the Indian Ocean island of La Reunion in 2005–2006, and caused approximately 266,000 cases [4]. In India, a subsequent epidemic reported about 1.4 million cases; several cases were imported to Europe, North and South America [5–8]. A single mutation A226V in E1 gene of the ECSA strains, which increases transmission by Aedes albopictus, allowed the emergence of CHIKV Indian Ocean lineage (IOL) [9,10]. The E1-A226V mutation allowed the rapid diversification of CHIKV IOL via a second wave of mutations in the E2; providing evidence that this mutation enhances the likelihood of IOL transmission, and consequently increases the risk of worldwide expansion [11]. Therefore, it was believed that ECSA genotype would be established in the Americas, where Aedes albopictus populations are present. However, in 2013 autochthonous CHIKV cases due to the Asian genotype were detected in the French Caribbean Island of Saint Martin. This CHIKV genotype later spread to other Caribbean Islands and the Americas causing epidemics in several countries [12,13].
The introduction of CHIKV into naive populations is followed by an explosive epidemic that affects a large number of people [9,10]. The interaction of some variables most likely favors this phenomenon: 1) susceptible human populations; 2) the presence of both mosquito vectors Aedes aegypti and Aedes albopictus; 3) and mutations in the virus that increase its infectivity [9,10]. From December 2013 to July 2015, a total of 1,118,763 suspected and 25,463 confirmed autochthonous CHIKV infections were reported in the Americas [14]. In 2015 an atypical presentation of the disease, that included distal extremity necrosis, was reported in Venezuela [15,16].
In Panama, the first imported case of CHIKV was reported in May 2014, and the first autochthonous case in August 2014 [13]. The majority of imported cases were associated with the Colombian, Dominican Republic and Venezuelan epidemics. Two main, but limited outbreaks, were detected in Panama: 1) in Rio Abajo (a county in Panama City) in August 2014 and 2) Puerto Obaldia (a town within an indigenous region near to Colombian border) in January 2015. The incidence rates for 2014 were low (0.5 per 100,000 inhabitants), and no fatalities or severe cases were detected, despite intensive surveillance. This contrasts with the high incidence in 2014 observed in Dominican Republic (5,182.5 per 100,000 inhabitants), Colombia (189 per 100,000 inhabitants) and Venezuela (131 per 100,000 inhabitants) (www.paho.org). This country also reported severe clinical presentations [15,16]. In order to understand these differences and the paucity of CHIKV outbreaks in Panama, we analyzed the CHIKV outbreak response and control in addition to the surveillance data of dengue, chikungunya and Aedes vectors between May 2014 and July 2015.
All information was obtained during the outbreak response and through the National Dengue and Chikungunya Surveillance programs, thus IRB approval was not necessary to submit (0277/CBI/ICGES/15), however all personal information was removed to perform the analysis and patient identification was codified to respect confidentiality.
The suspected chikungunya case definition was set up at the beginning of chikungunya outbreak in the Americas to alert all-medical personnel of this new introduction in dengue endemic countries. Overall, 114/413 (27.6%) cases of CHIKV infections were confirmed through CHIKV surveillance. A total of 60.7% were women and 88.6% were more than 15 years old; the mean age was 37.8 years old (SD ± 17.9) (Table 1). All (100%) of patients presented to the health center with fever; the majority presented with polyarthralgia (98%), myalgia (88%), headache (79.3%), chills (77.2%) and rash (76%) (Table 1). A total of 29/413 (7.0%) of the suspected CHIKV cases were confirmed as DENV infections. A high number of suspected cases (65.1%; 269/413) were negative for both viruses.
From May (19th epidemiological week) 2014 to July (26th epidemiological week) 2015, a total of 413 patients met the suspected case definition for CHIKV infection in the country. The majority of the suspected CHIKV cases (70.2%, 290/413) were located in the Districts of Panama City and San Miguelito, which are the most densely inhabited districts in the country. Similarly, around 65.8% of confirmed CHIKV infections (75/114) were detected in Panama City and San Miguelito (Fig 1, Table 1).
From all suspected cases, 114/413 (27.6%) were confirmed CHIKV infections, of these 114 cases, 26.3% were detected in 2014 (71/270) and 30.2% in 2015 (43/142) respectively (Table 1). Of these confirmed CHIKV infections, a total of 70/114 (61.4%) were imported infections (Table 1, Fig 2A), 53 of them detected in 2014 and 17 in 2015, respectively.
Two mains CHIKV outbreaks were detected: one in Panama City during 2014 (epi-weeks 31 to 40), with a total of 28 cases, 13 of which were autochthonous and one in 2015 (epi-weeks 22 to 26), which had 19 cases detected in Puerto Obaldia (Figs 1 and 2A). Of the confirmed CHIKV cases, 51 were detected by RT-PCR, showing a concordance of 92.2% with the clinical and molecular diagnosis (≤8 days of symptoms), while 65 were detected by IgM ELISA) with concordance of 92.3% within the clinical and serology diagnosis (≥9 days of symptoms) (Table 1). A total of 29/413 (7.0%), were confirmed cases of DENV infection, and 269/413 (65.1%) were negative for both viruses (Table 1) and for the three mains groups: alphavirus, flavivirus and phlebovirus.
The CHIKV incidence rates were 0.5 and 0.7 per 100,000 inhabitants in 2014 and 2015 (until epidemiological week 26), respectively (Table 2). Data from National Dengue Surveillance system were used to compare CHIKV and DENV infections rates for both years, the ratio indicate that CHIKV and DENV infections were 1:306 and 1:34, respectively. The incidence rates varied between provinces, and Panama province (divided in four health regions: West Panama, Panama, San Miguelito and East Panama) reported the majority of cases (Tables 1 and 2).
In order to reduce selection bias influence and to increase the possibility of detection of CHIKV cases that were not caught through National Chikungunya Surveillance, samples from Dengue surveillance were also analyzed. All the samples received during the time of the study met the quality control requirement and none was rejected for analysis. From a total of 1489 samples (564 DENV positive and 925 DENV negative) received at ICGES through the National Dengue surveillance system representative from all provinces, we randomly selected and tested a total of 879 dengue negative samples for CHIKV detection (612 from 2014 and 267 until July 2015) (S1 Table). From these, only two samples were CHIKV positive (1/612 in 2014 and 1/267 in 2015).
The national epidemic curve for all dengue positive cases (6459) detected between 2014 to July 2015 in Panama shows an increase in the number of cases at the beginning of 2014, before CHIKV detection in May that same year (Fig 2B). This epidemic curve shows a similar pattern to the epidemic curve of dengue cases detected by ICGES during the same period of time, suggesting that dengue laboratory surveillance at ICGES (8.9% [577/6459] of total cases) represented the behavior through time of all reported dengue cases in Panama. CHIKV infections were detected starting in May 2014 when dengue cases were also lower than the previous months (Fig 2B). Acute DENV and CHIKV co-infection was not detected, however seroconversion for both viruses was detected in 5 patients, and one presented with IgM-dengue and CHIKV detected by RT-PCR[13].
To describe vector infestation level of Ae. aegypti and Ae. albopictus and their possible relation with the low CHIKV secondary transmission, the vector infestation rates in Panama and San Miguelito were analyzed. Both districts reported most of the Chikungunya confirmed cases (imported and autochthonous), with the majority being detected in 2014 (Fig 3A).
Vector infestation levels observed during the 2014 outbreak fluctuate between 0.4–4.2% (Fig 3B). Low infestation levels were documented mainly during the dry season (January-April) and moderate infestation levels during the rainy season (May-December), where the rate of infection was 4.2%. This corresponds the increase detection of imported CHIKV and secondary transmission (Fig 3A). During 2015, we observed low levels of infestation during the dry season and moderate levels in subsequent months (Fig 3B), when only imported cases were detected (Fig 3A). High infestation levels (8.2 to 14.3%) were observed in the rainy season of 2015. The student t-test applied for 2014–2015 showed no significant difference (P >0.05). When the values of infestation were compared between the months of 2014, it was determined that there was significant difference (P <0.05) between infestation rates only during the months of the rainy season, whereas in the dry season there was no difference. By 2015, only significant difference was observed in the months of April and May due to high incidence and outliers.
The comparison of the standard deviations of the time series with, and without interpolations, corroborated their low variability in the infestation levels through epidemiological weeks (Fig 3B). The minimum quadratic showed that through the eighty-three epidemiological weeks analyzed, the levels of Aedes sp. vector infestation increased by a rate of 0.053 per epidemiological week, the graphic show the growing trend behavior observed in the level of infestation (Fig 3B). The abundance analysis shows that Ae. aegypti is two times more abundant than Ae. albopictus (P<0.05) (Fig 3C).
In 2014, imported cases were reported mainly in seven counties in Panama and San Miguelito districts, all of which had low or medium infestation levels, other than the Las Cumbres County, which had a high infestation rate (Fig 4A). Autochthonous cases were detected in only 3 counties: Rio Abajo and Pueblo Nuevo from the District of Panama and Amelia Denis de Icaza from the District of San Miguelito. Rio Abajo County presented a high proportion of imported cases, but low infestation like in Amelia Denis de Icaza County, while Pueblo Nuevo County had medium infestation level. In 2015 (January-July), seven counties had imported cases, all with low or medium infestation levels, other than Las Cumbres. However only four counties reported one autochthonous case for each (Fig 4B), one of them was Rio Abajo again. The infestation rates in Rio Abajo remained between low and medium during 2014 and 2015 respectively (Fig 4B).
To determine if variations over time in vector infestation levels are associated with CHIKV and DENV outbreaks, the CHIKV epidemic curve with imported and autochthonous cases, as well as dengue epidemic curve were superimposed along with the monthly mean infestation rate for Rio Abajo (S4 Fig). The increase of vector infestation levels just before the detection of imported cases was associated with an appearance of autochthonous cases. However, as soon as CHIKV cases were reported, there was a decrease in infestation rates that could be associated with the vector control intervention and educational campaign. Juan Diaz County also maintained a low infestation rate during the period of study. Although imported cases of CHIKV were reported, no autochthonous cases were detected in this County (S3 Fig). Even if vector infestation rates remained medium and low, DENV cases were detected in both counties for most months included in the analysis (S4 Fig).
The imported index case was detected in May 2014 and autochthonous index case was reported in August 15th [13]. In July that year, 3 imported cases and 2 autochthonous cases in the convalescent phase were detected in Rio Abajo County (S4 Fig). From August 9th to 12th, three chikungunya cases during the acute phase were detected, all from the same street in that county. Three convalescent cases were detected also that month. MINSA was notified of laboratory confirmation on September 8th and the active febrile surveillance and mosquito control measures were started on September 11th. During this intervention, a total of 15 febrile cases were detected (onset of symptoms ranged from 0 to 8 days), all negative for dengue and two of them were confirmed as CHIKV infection. In total from August 15th to October 6th 2014, 37 acute cases from Rio Abajo were tested for CHIKV infection, of which 7 were confirmed by RT-PCR.
The optimal Maximum-likelihood (ML) tree, based on a 975 nt segment of the E1/E2 genes sequences of 53 CHIKV strains (Fig 5), shows that the 13 Panamanian imported and two autochthonous isolates (Genbank accession number KX255061-63 [13], KX355507-16) cluster together within the American clade of the Asian lineage. The Panamanian strains (256821, 257245 and 257263) were obtained from febrile patients that came from Venezuela, while the strain 256899 was obtained from a patient traveling from El Salvador. The majority of imported strains included in our study were obtained from patients with febrile disease that came from Dominican Republic. Two strains were autochthonous to Panama (256619, 256629). Sequences from all Panamanians strains (imported and autochthonous) included in our study are nearly identical to each other, and cluster together with strains isolated in Saint Martin in 2013 [12] and the British Virgin Islands during 2014 [29]. No known mutations of vector adaptation were found in the analyzed E1/E2 sequence.
While many tropical Caribbean and Latin American countries that reported autochthonous infections after 2013 have experienced explosive CHIKV epidemics [30], in Panama, from May 2014 to July 2015, we confirmed only 46 autochthonous CHIKV cases (44 through CHIKV surveillance and 2 through dengue surveillance), that corresponds to the 38.6% of total CHIKV confirmed cases (Table 1 and S1 Table). Clinical and laboratory findings of CHIKV infections in the Panamanian outbreak were similar to those reported in previous epidemics [31–33]. The most predominant signs and symptoms observed in CHIKV Panamanian cases were rash and polyarthralgia, similar to previous reports [31–33]. In previous reports polyarthralgia was more prevalent in CHIKV when compared to DENV infections [34]. In the Panamanian cases, no fatalities or atypical clinical presentations, such as extremity necrosis occurred [15,16]. Therefore, suspected chikungunya cases could be misdiagnosed here as suspected dengue cases. Even though the viremia is higher and lasts longer than dengue, our data show that even through both surveillance programs, few cases of CHIKV were detected [35].
High CHIKV incidence rates and severe cases were reported in several countries of Latin America from July 2014 to July 2015: Dominican Republic: 5,182.5 and 0.6 per 100,000 inhabitants; El Salvador: 2,135.4 and 375.8 per 100,000; Guatemala: 178.1 and 50.8 per 100,000, Honduras: 66 and 607.1 per 100,000, Nicaragua: 70.6 and 331.9 per 100,000, Colombia: 189 and 612 per 100,000, and Venezuela: 131 and 42 per 100,000, respectively. The overall mortality rates in Latin America was 194 in 2014 and 41 in July 2015 [14]. In comparison, in Panama, the incidence rates were 0.5 and 0.7 per 100,000 inhabitants in 2014 and 2015, respectively.
In both years 2014 and 2015, the CHIKV incidence rate was lower compared to DENV. The marked difference observed between CHIKV and DENV in 2014 is probably due to the high number of dengue cases detected during January-April 2014 that corresponded with a dengue outbreak that began just before Chikungunya detection (Table 2). Dengue is an endemic disease in Panama, and about 23 years have passed since reintroduction. This fact may influence physicians’ criteria for differential diagnoses. The sensitivity of dengue diagnosis through the decentralized surveillance should be greater than the recent implemented centralized CHIKV surveillance. In addition, the majority of DENV infections are asymptomatic and sustained DENV transmission by people who are infected without developing detectable clinical symptoms have been proposed [36]. This fact may reduce the effectiveness of traditional control methods that in contrast could be effective in the control of CHIKV infections as the majority of cases are symptomatic [2,31]. Future studies need to be conducted to determine if there are ecological or virus-specific constraints on DENV or CHIKV, due to usage of the same vectors and reservoir.
The majority of CHIKV cases were found in the Panama City metropolitan area (Districts of Panama and San Miguelito). This is likely related to the fact that the most densely populated areas are Panama City and San Miguelito, and both have a high proportion of immigrants inhabiting or working there. The majority of imported cases in 2014 were from Dominican Republic and in 2015 from Colombia and Venezuela. The latter two countries had reported CHIKV outbreaks during that year [14]. Imported CHIKV infections were not reported in Puerto Obaldía during 2015. Human migration of native communities through the Panamanian-Colombian border is very usual. This migration is most likely not reported; therefore, it is difficult to determine if these cases were autochthonous. As Colombia had an important CHIKV epidemic during 2015, it is possible that migration over this border was the most likely way the virus was introduced in the east side of the country.
In this study, we sought to provide information about viral genetic differences that could explain the paucity of CHIKV cases observed in Panama. However, our phylogenetic analysis indicates that these CHIKV strains from the Asian genotype are nearly identical to those strains that are circulating in the rest of the Americas [13,29]. These results suggest that a single introduction of CHIKV from the Asian genotype in the Caribbean islands in 2013 spread to the rest of the Americas, and that Panama is not an exception.
No mutations of vector adaptation in the E1/E2 genes of CHIKV strains included in our studies were found. Further studies of 3’UTR of Panamanian CHIKV strains, should be addressed because it has been proposed that this genomic region may reduce the fitness of the Asian genotype for efficient transmission by mosquitoes [37]. Together, this information suggests that the naive population, along with the presence of Aedes aegypti [38] were the major forces that facilitate the dissemination of the Asian genotype through the Americas and not a specific mosquito adaptation like in the La Reunion outbreak [10].
As no specific adaptive mutations were found, other possible explanations for the paucity of Panamanian CHIKV cases are: a) heterologous alphavirus antibodies cross-protect against CHIKV infection and/or disease; b) low level of vector infestation before and after CHIKV introduction; c) and the early case detection and implementation of control measures.
Experimental infections in mice with Mayaro virus (MAYV) and the Alphavirus encephalitis viruses (VEEV and EEEV) have shown cross-protection that seems to last around 2 months [39]. Cross-reactivity between MAYV and CHIKV has been reported [40]. However, it is not clear whether previous infections with VEEV, MAYV or Una virus (UNV) have an impact in CHIKV infection or disease. Moreover the distribution of the main vectors of VEEV, EEEV and probably MAYV and UNV is sylvatic and rural [41,42]. This contrast with the current distribution of Ae. albopictus and Ae. aegypti in Panama that are mainly urban and peri-urban [41] and with the distribution of the CHIKV confirmed cases. Further studies should explore the effect of Alphavirus cross-protective immunity in the pattern of CHIKV emergence observed in Panama.
Experimental studies have shown that Ae. aegypti and Ae. albopictus efficiently transmit CHIKV of the Asian genotype with a major fitness in Ae. aegypti [43]. Our analysis of vector infestation levels shows a predominant low to moderate risk for CHIKV epidemics. The detection of only 3 secondary cases during the main CHIKV outbreak in Panama City (Rio Abajo County) suggests that the observed infestation levels along with interventions after a suspected case detection may play a role in the observed pattern of CHIKV outbreak in Panama. Nevertheless, we were not able to test the hypothesis that rapid vector control measures impacted the dynamics of CHIKV transmission in Panama; little information of vector infestation levels in specific locations before and after the control response was available. However, biochemical studies have shown that strains of Ae. aegypti in Panama City are sensitive to the insecticides used currently for control campaign [44]. The sensitivity of the more abundant Panamanian vector to insecticides and the low number of CHIKV cases in the community following vector control measures suggest that these local control measures applied by MINSA in areas close to confirmed cases were efficacious to limit the number of human cases and the expansion of CHIKV throughout the country. Additionally, because behavior and lifestyle have been proposed to increase or restrict the transmission of DENV rather than the effects of climate in some places [45], the role of lifestyle in the distribution pattern of CHIKV infections observed in Panama, should be addressed in future studies.
The population of Ae. aegypti appears to be two times more abundant than Ae. albopictus in Panama City where the majority of cases occurred. Ae. albopictus has expanded across Panama and predictions indicate that it will colonize the entire Pan-American highway, potentially increasing the area of CHIKV transmission [41]. However the experimental and field studies of vector competence and transmission of Asian CHIKV strains [38,43] suggest that Ae. aegypti may be currently the principal vector of CHIKV transmission in Panama. Variations in the temperature have shown to reduce Ae. aegypti vector competence for DENV transmission [46], further studies should address the effect of temperature and humidity in CHIKV transmission, a factor that could add to the small number of CHIKV infections observed in Panama as CHIKV emerged in Panama during the drought in El Niño season.
The early recognition of CHIKV with the subsequent measures could have limited the autochthonous cases in Panama. The first imported cases were detected during the viremic phase, with one case detected at the airport [13]. The available epidemiological data shows that around half of CHIKV cases in Panama were detected during the first days of symptoms. This suggests that the surveillance program was able to detect clinical chikungunya cases during the acute phase. DENV and CHIKV laboratory surveillance plays an important role in Panama. All suspected chikungunya cases were laboratory tested, whereas around 50% of Dengue cases are laboratory confirmed (MINSA-NED.). This differs from the situation in other countries during 2014, like El Salvador or Colombia that had only 157 confirmed cases from the 135,226 CHIKV suspected cases, or 611 confirmed from 90,481 suspected respectively [14]. Moreover, ICGES not only tested all chikungunya suspected cases, in addition its sensibility to detect viral RNA in acute samples was increased by using specific CHIKV detecting methods complemented with genus-specific alphavirus RT-PCR in dengue suspected cases that were dengue negative. The epidemiological surveillance algorithm establishes testing of convalescent paired samples to confirm positive and negative CHIKV results, to avoid false negatives results by RT-PCR due to an inaccurate onset symptoms description, low viral load or sample handling. Although patients were encouraged to attend a second medical visit to obtain the convalescent sample, few participated, especially as symptoms disappeared. For the convalescent paired samples obtained, serological tests were performed as a confirmation of acute samples. PRNT was suggested by PAHO for diagnosis confirmation and would be helpful to confirm seroconversion, especially when circulation of chikungunya has to be proven in a new country[17], this technique was not used at the beginning of the outbreak in Panama as Chikungunya first detected cases were acute, thus Chikungunya circulation was proven in Panama by molecular methods and viral isolation[13]. Once Chikungunya circulation was confirmed in Panama, the CHIKV surveillance program was adapted to the capacity of GMI laboratory and Chikungunya PRNT was not available to confirm serologic results at the time of the outbreak. However, this limitation in the chikungunya surveillance was also shared by most Latin American countries.
Vector control responses at the community level began before the laboratory confirmation, with the notification of a suspected case by the physician to MINSA. After case confirmation and notification, the vector control personnel applied intradomiciliary fumigation in order to complete the response. The algorithm set for the notification of confirmed cases includes notification (around 24 to 72 h) from the laboratory to the treating physician as well as at the local level where the cases are reported and at the National Epidemiology Department at MINSA that informs the Vector Control Department, which in less than 48h organized a complete vector control response with implementation of effective control measures, such an isolation and mosquito control in domiciliary, peridomiciliary and the surrounding community [47]. It is possible that the low number of CHIKV cases in Panama was influenced by the early response after detection of each new case [48]. Quality controls of vector control measures were done every six months. During each evaluation, the calibration of the spraying equipment, calibration of the diameter of the mouthpiece and the drop size used in fumigation was done, as well as control of the security of equipment and chemical usage. Human resource was also retrained every year. Population acceptance of fumigation was high as there are familiar with this procedure since Panama sanitation measures during Panama Canal construction [49], however there are always some closed houses that can not be analyzed intradomiciliary. Vector resistance pilots were done every year by GMI entomologists.
Our study has several limitations. First, the chikungunya suspected case definition was based in fever and arthralgia or arthritis in order to differentiate chikungunya from dengue infections, this may represent a selection bias, as the symptoms may be variable in some chikungunya infections [50]. However, samples from patients with dengue-like disease were also tested even if the physician did not request chikungunya test. Underreporting of cases is also possible, even if notification is mandatory; this is not a total guarantee that all suspected cases were reported. However, all medical institutions in Panama have epidemiologists that follow notification of cases daily and increase the surveillance during outbreak response. A small number of samples from remote areas were received at ICGES during the study. This is possibly due to the long distance and logistical difficulties that arise for transport of samples. DENV surveillance is able to detect an increase of cases in remotes areas and transport samples to ICGES during an outbreak response, however an increase of cases was not reported during our study. Most cases were treated as ambulatory cases in primary care centers where the majority of information was obtained from the epidemiological notification form. Little information on the clinical evolution is currently available, and the rate of chronic CHIKV infections in Panama is unknown, as most patients did not return to the health facility for follow-up after their first visit. This information contrasted with the high proportion of patients with sequelae that were reported in hospitals from countries like Dominican Republic and El Salvador [51]. Similarly, in this study only the clinical symptoms described in the notification form were reported, and not chemical or hematologic test were performed, in consequence correlation between clinical presentation, thrombocytopenia or leucopenia was not possible. The Panamanian vector surveillance during the outbreak was performed using Breteau index that is a larval based index, this index have been found to be inadequate as indicators for dengue virus transmission [52–54]. Indeed, a study have been proposed that adult mosquitoes index correlate better with positive dengue cases in Iquitos, Peru [53], however no study has been done to correlate these differences indices in Panama with dengue or chikungunya cases and larvae index have also shown to correlate with dengue epidemics in Cuba [55]. This discrepancy in mosquito estimator to predict outbreak, suggest variations of correlation among countries. MINSA are considering the use of pupa index instead of larvae; nevertheless, this has not yet been implemented. Finally, detection of the circulating virus in mosquitos was not performed during the response; in consequence, there is no estimate on the rates of infected mosquitos for DENV and CHIKV that could help define risk areas of transmission.
In summary, our data suggest that Panama presented a small and limited outbreak compared with others Latin American countries. Panama was able to maintain low to medium vector infestation levels in the majority of the counties or communities with imported CHIKV cases, and this, along with the interventions after identification of CHIKV infections, as well as other epidemiological conditions, could play a role in the low numbers of cases observed. The constant possibility of new imported CHIKV cases and the introduction of Zika virus in 2015 [56] will continue to be a threat for the surveillance and vector control programs. The capacity to maintain the observed pattern will depend on the preservation of low to medium infestation levels and the sustainability of the early case detection system of dengue-like arboviral diseases, as well as the subsequent implementation of vector control measures after detection of new cases before bigger epidemics.
|
10.1371/journal.pcbi.1005573 | NetNorM: Capturing cancer-relevant information in somatic exome mutation data with gene networks for cancer stratification and prognosis | Genome-wide somatic mutation profiles of tumours can now be assessed efficiently and promise to move precision medicine forward. Statistical analysis of mutation profiles is however challenging due to the low frequency of most mutations, the varying mutation rates across tumours, and the presence of a majority of passenger events that hide the contribution of driver events. Here we propose a method, NetNorM, to represent whole-exome somatic mutation data in a form that enhances cancer-relevant information using a gene network as background knowledge. We evaluate its relevance for two tasks: survival prediction and unsupervised patient stratification. Using data from 8 cancer types from The Cancer Genome Atlas (TCGA), we show that it improves over the raw binary mutation data and network diffusion for these two tasks. In doing so, we also provide a thorough assessment of somatic mutations prognostic power which has been overlooked by previous studies because of the sparse and binary nature of mutations.
| The transition from a normal cell to a cancer cell is driven by genetic alterations, such as mutations, that induce uncontrolled cell proliferation. With the advent of next-generation sequencing technologies (NGS) in the last decade, thousands of tumours have been sequenced and their mutation profiles determined. However, the statistical analysis of these mutation profiles remains challenging. Indeed, two patients usually do not share the same set of mutations and can even have none in common. Moreover, it is difficult to distinguish the few disease-causing mutations from the dozens, often hundreds of mutations observed in a tumour. To alleviate these challenges, it has been proposed to use gene-gene interaction networks as prior knowledge, with the idea that if a gene is mutated and non-functional, then its interacting neighbours might not be able to fulfil their function as well. Here we propose NetNorM, a method that transforms mutation data using gene networks so as to make mutation profiles more amenable to statistical learning. We show that NetNorM significantly improves the prognostic power of mutation data compared to previous approaches, and allows defining meaningful groups of patients based on their mutation profiles.
| Tumourigenesis and cancer growth involve somatic mutations which appear and accumulate during cancer progression. These mutations impair the normal behaviour of various cancer genes, and give cancer cells an often devastating advantage to proliferate over normal cells [1–3]. Systematically assessing and monitoring somatic mutations in cancer therefore offers the opportunity not only to better understand the biological processes involved in the disease, but also to help rationalise patient treatment in a clinical setting. Rationalising treatment involves finely characterising the genomic abnormalities of each given patient to discover which may be treatable by a targeted therapeutic agent, as well as improving prognosis using molecular information [4–6]. The development of fast and cost-effective technologies for high-throughput sequencing in the last decade has triggered the launch of numerous data collection projects such as The Cancer Genome Atlas (TCGA) [7] or the International Cancer Genome Consortium (ICGC) [8], aiming at characterising at the molecular level, including genome-wide or exome-wide somatic mutations, thousands of cancer samples of multiple origins. By systematically comparing the molecular portraits of the resulting cohorts, one might expect to be able to detect frequently mutated genes or groups of genes, and find associations between particular mutations and cancer phenotypes, response to treatment, or survival [9–12].
The analysis of somatic mutation profiles is however challenging for multiple reasons. First, most somatic mutations detected by systematic sequencing are likely to be irrelevant for biological or clinical applications. This is due to the fact that only a few driver mutations are required to confer a growth advantage to the cancer cell, and therefore most somatic mutations are likely to be passenger mutations which do not contribute to the cancer phenotype [3, 13]. Second, sequencing efforts have shown that while a few genes are frequently mutated, the vast majority of genes are mutated in only a handful of patients [14, 15]. As a result, the mutation profiles of two tumours often only share a few if any genes in common. Third, even if originating from the same tissue, tumours may exhibit widely varying mutation rates. The overall mutational burden of a tumour constitute a strong and informative signal [16–18] but can however complicate the retrieval of more subtle signals. Combined with the inherent high dimensionality of somatic mutation datasets, this makes any statistical analysis of cohorts of whole-exome somatic mutation profiles extremely challenging.
In order to make somatic mutation profiles more amenable to statistical analysis, several studies have used gene networks as prior knowledge [19, 20]. Considering genes in the context of networks instead of analysing them independently allows sharing mutation information among neighbouring genes and identifying disruptions at the level of pathways or protein complexes instead of single genes. A popular method to leverage this prior knowledge consists in using a diffusion process on the gene network. This technique first appeared for the analysis of gene expression and GWAS data [21–25], and has more recently been used for mutation profiles [26–31]. Network diffusion processes allow smoothing binary vectors of somatic gene mutations into non-negative real-valued vectors of mutational statuses, where the mutational status of a gene increases when it is close to mutated genes in the network. This approach led to state-of-the-art methods for the discovery of driver pathways or complexes [30] and for the stratification of patients into clinically relevant subtypes [31] using whole-exome mutation profiles.
In this work we propose NetNorM, a new method to enhance mutation data with gene networks. NetNorM transforms a patient’s binary mutation profile by either removing mutations or creating “proxy” mutations based on the gene network topology, until all patients reach a consensus number of mutations. The resulting mutation matrix is binary like the initial one, nonetheless we establish that it encodes new information reflecting both local network neighbourhood mutational burdens and the overall tumour mutational burden.
We evaluate the relevance of NetNorM on two tasks: survival prediction and patient stratification from exome somatic mutation profiles. In doing so, we also provide a thorough assessment of somatic mutations prognostic power which has been overlooked by previous studies because of the sparse and binary nature of mutations [32]. We show that NetNorM produces state-of-the-art results for these two tasks compared to the raw binary mutation data and to network diffusion-based methods. By comparing results obtained with real versus randomised networks, we further show that the increase in relevance is actually partly driven by the gene’s network prior knowledge. However, we observe that considering interactions between mutated genes and their network neighbours only is enough do achieve state-of-the-art results, thereby shedding light on which are the network features that are the most informative.
NetNorM takes as input an undirected gene network and raw exome somatic mutation profiles and outputs a new representation of mutation profiles which allows better survival prediction and patient stratification from mutations (Fig 1). Here and in what follows, the “raw” mutation profiles refer to the binary patients times genes matrix where 1s indicate non-silent somatic point mutations or indels in a patient-gene pair and 0s indicate the absence of such mutations. The new representation of mutation profiles computed with NetNorM also takes the form of a binary patients times genes mutation matrix, yet with new properties. While different tumours usually harbour different number of mutations, with NetNorM all patient mutation profiles are normalised to the same number k of genes marked as mutated. The final number of mutations k is the only parameter of NetNorM, which can be adjusted by various heuristics, such as the median number of mutations in the original profiles, or optimised by cross-validation for a given task such as survival prediction. In order to represent each tumour by k mutations, NetNorM adds “missing” mutations to samples with less than k mutations, and removes “non-essential” mutations from samples with more than k mutations. The “missing” mutations added to a sample with few mutations are the non-mutated genes with the largest number of mutated neighbours in the gene network, while the “non-essential” mutations removed from samples with many mutations are the ones with the smallest degree in the gene network. These choices rely on the simple ideas that, on the one hand, genes with a lot of interacting neighbours mutated might be unable to fulfil their functions and, on the other hand, mutations in genes with a small number of interacting neighbours might have a minor impact compared to mutations in more connected genes.
In this study, we compare NetNorM-processed profiles with the raw mutation data and with profiles processed with network smoothing (NS) [33] (also called network diffusion, or network propagation) followed by quantile normalisation (QN) as implemented in [31]. We refer to this method as NSQN below. Mutation profiles, either raw or processed with NetNorM or NSQN, are restricted to the genes present in the network used. While both NetNorM and NSQN leverage gene network prior knowledge to enhance mutation data, the two methods have fundamental differences. First, NetNorM leverages information about first neighbours in the network only while NSQN spreads mutation information at a more global scale on the gene network. Second, with NetNorM the normalised profiles all have the same value distribution by construction, since they are all binary vectors with k ones, removing the need for further quantile normalisation which, as we discuss below, is critical for NSQN.
To assess the relevance of NetNorM, we first explore the capacity of somatic mutations to predict patient survival. We collected a total of 3,278 full-exome mutation profiles of 8 cancer types from the TCGA portal (Table 1), censored survival information and clinical data. In parallel we retrieved a gene network to be used as background information for NSQN and NetNorM: Pathway Commons, which integrates a number of pathway and molecular interaction databases [34]. For each cancer type, we use these data to assess how well survival can be predicted from somatic mutations. For that purpose, we perform survival prediction with a sparse survival SVM (see Methods) using either the raw mutation profiles or the profiles processed with NSQN or NetNorM, respectively, and assess their performance by cross-validation using the concordance index (CI) on the test sets as performance metric.
Fig 2 summarises the survival prediction performances for the 8 cancer types, when the sparse survival SVM is fed with the raw mutation profile, or with the mutation profiles modified by NSQN or NetNorM using Pathway Common as gene network. For two cancers (LUSC, HNSC), none of the methods manages to outperform a random prediction, questioning the relevance of the mutation information in this context. For OV, BRCA, KIRC and GBM, all three methods are significantly better than random, although the estimated CI remains below 0.56, and we again observe no significant difference between the raw data and the data transformed by NSQN or NetNorM. Finally, the last two cases, SKCM and LUAD, are the only ones for which we reach a median CI above 0.6. In both cases, processing the mutation data with NetNorM significantly improves performances compared to using the raw data or profiles processed with NSQN. More precisely, for LUAD the median CI increases from 0.56 for the raw data and 0.53 for NSQN to 0.62 for NetNorM. In the case of SKCM, the median CI increases from 0.48 for the raw data to 0.52 for NSQN, and to 0.61 for NetNorM. For SKCM, both NetNorM and NSQN are significantly better than the raw data (P < 0.01).
In our experiments, silent mutations are systematically filtered out. To evaluate whether this preprocessing step is actually detrimental or beneficial for the survival prediction task, we performed further experiments where silent mutations are not filtered out (S1 Fig). We find that considering silent mutations does not improve survival prediction performances compared to the case where they are filtered out. In fact, the performance of NetNorM on LUAD is significantly decreased when silent mutations are taken into account.
To assess the influence of the gene network used on the survival prediction performances, we also repeated our experiments with four gene networks instead of Pathway Commons: BioGRID [35], HPRD [36], HumanNet [37] and STRING [38] (S2 Fig). For HumanNet and STRING, only the 10% most confident interactions were retained. We observe that no gene network clearly stands out as the best network for all cancers. For two cancers, LUSC and HNSC, performances remain very low, close to a concordance index of 0.5, whatever the method or network used. For three cancers, OV, BRCA and KIRC, NetNorM is the only method to significantly outperform the raw data with at least one network (HumanNet and STRING for OV, HPRD for BRCA, and STRING for KIRC) with a median concordance index above 0.55. For GBM, NSQN is the only method to outperform the raw data (with HumanNet and STRING) with a median concordance index above 0.55. For the two remaining cancers, LUAD and SKCM, the best performances are those obtained with NetNorM using Pathway Commons, with median CI of 0.62 and 0.61 respectively. Across all cancers, methods, and networks combinations, these two cases are the only ones where the median CI obtained exceeds 0.60.
Finally, as mutations in some genes are known to be associated with survival, such as TP53 in BRCA and HNSC which is associated with worsened survival [39], we evaluate the prediction ability of individual genes’ mutation status. For each cross-validation fold, the gene giving the best concordance index on the training set is selected and its performance evaluated on the test set. We find that for 5 cancers, the performances of individual genes are similar to those of the survival SMV applied to the whole raw mutations datasets (S3 Fig). However for BRCA and HNSC, better survival predictions are obtained using a single gene than the whole raw mutational profiles. Yet these predictions are not better than those obtained with NetNorM. For these two cases, TP53 is the gene selected in the majority of folds (17/20 for HNSC and 19/20 for BRCA), which is in accordance with existing literature (S1 Table). Lastly, the survival SVM applied to the whole dataset yields significantly better performances than the single gene approach for LUAD. This means that contrary to the BRCA and HNSC cases, the linear combinations of genes which are found for LUAD have a predictive power that generalises well to unseen data.
In summary, these results show that for at least 6 out of 8 cancers investigated, somatic mutation profiles have a prognostic value, and that for two of them (SKCM and LUAD) it is possible to improve the prognostic power of mutations by using gene networks and to reach a CI above 0.6. In both cases, NetNorM is significantly better than NSQN.
To test whether the biological information contained in the gene network plays a role in the improvement of survival predictions for LUAD and SKCM, we evaluate again NetNorM and NSQN using 10 different randomised versions of Pathway Commons for these two cancers. Random networks were obtained by shuffling the nodes’ labels of the real network while keeping the structure unchanged. The results, shown on Fig 3, demonstrate that NetNorM performs significantly better with a real network. More precisely, the real network significantly outperforms all random networks for SKCM and 8 out of 10 random networks for LUAD (Wilcoxon signed-rank test with correction for multiple hypothesis testing, FDR ≤ 5%). NSQN also performs significantly better with a real network for SKCM (7 out of 10 cases) but not for LUAD (0 out of 10 cases). This last observation is not surprising since NSQN does not improve over the raw data for LUAD, which suggests that the method may have failed to leverage network information in this case. In summary, these results indicate that the improvements obtained with NetNorM and NSQN compared to the raw data do rely on biological information encoded in the network.
In order to shed light on the reasons why NetNorM outperforms the raw data and NSQN on survival prediction for SKCM and LUAD, we now analyse more finely the normalisation carried out by NetNorM on the mutation profiles, and why they lead to better prognostic models. For that purpose, we focus on the genes that are selected at least 50% of the times by the sparse survival SVM during the 20 different train/test splits of cross-validation, after NetNorM normalisation. This leads to 21 frequently selected genes for LUAD and 10 for SKCM (Fig 4). Remembering that NetNorM either removes mutated genes for patients with many mutations, or adds proxy mutations for patients with few mutations, we can assess for each frequently selected gene whether it tends to exhibit proxy mutations or whether it tends to be actually mutated in the tumour. This is done by comparing how frequently it is marked as mutated on the raw data and after NetNorM normalisation (Fig 4, top plot). For both cancers, we observe two clearly distinct groups of frequently selected genes: those that concentrate proxy mutations (which we will call proxy genes, in red in Fig 4), and those to which NetNorM brings only few modifications compared to the raw data, meaning they are usually actually mutated in the tumours (in black in Fig 4).
We assess whether the combination of both mutations and clinical features can improve performances for LUAD and SKCM compared to using clinical data alone. For this purpose, two sparse survival SVM models are trained independently: one on the raw mutation data or mutations preprocessed with NSQN or NetNorM and one on the clinical data. Then the survival predictions from both models are simply averaged (after being standardised to unit variance). The resulting predictions are again evaluated in a 4 times 5 folds cross-validation setting. First, the results show that mutations preprocessed with NetNorM and the clinical data yield similar performances (P = 0.52, Wilcoxon signed rank test) for LUAD while the clinical data performs significantly better than NetNorM in the case of SKCM (P ≤ 1 × 10−2) (Fig 6). Moreover, we observe that combining mutations preprocessed with NetNorM with clinical features allows improving survival predictions compared to the clinical data alone for both LUAD (P = 4.8 × 10−2) and SKCM (P = 5.7 × 10−2). More precisely, the median CI increases from 0.64 with the clinical data to 0.66 with the combination of NetNorM and the clinical data for LUAD and from 0.66 to 0.70 in the case of SKCM. We also tried to concatenate the mutation profiles with the clinical data before training a unique model and observed that it did not improve the results compared to the previous strategy (S5 Fig). Overall, these results suggest that mutations could provide useful prognostic information that is complementary to the clinical information available.
We now assess the possibility to stratify patients into a small number of groups in an unsupervised way, meaning without using survival information, in order to identify distinct subgroups of patients in terms of mutational profiles. For that purpose, we use a standard unsupervised clustering pipeline based on nonnegative matrix factorisation (NMF), and apply it to the different cohorts of patients represented by the raw mutation profiles, or the profiles normalised by NSQN or NetNorM. The hyperparameters k (NetNorM) and α (NSQN) were set to default values chosen as the median number of mutations in a cohort for k and α = 0.5 as recommended in [31]. As we have no ground truth regarding “true” groups of patients, we assess the quality of clustering by two factors: (i) the stability of the clusters, assessed by the proportion of ambiguous clustering (PAC) which is the rate of discordant cluster assignments across 1,000 random subsamples of the full cohort; and (ii) the significance of association between clusters and survival.
With the raw data, NMF tends to stratify patients into very unbalanced subtypes with typically one subtype gathering the majority of patients (Fig 7b). LUSC, HNSC and SKCM are extreme cases where one cluster contains 95% of the patients, whatever the number of clusters. In addition, in cases where the obtained clusters are reasonably balanced as for KIRC, the clustering stability is low. These results are coherent with [31] who highlighted the difficulty to cluster raw mutation profiles. These undesirable behaviours disappear with both NSQN and NetNorM (Fig 7). With NetNorM the obtained clusters are reasonably balanced across all cancers and the clusters are stable (PAC ≤ 30%). NSQN also provides stable clusters (PAC ≤ 30%) when the number of clusters is set between 4 and 6 however for 2 or 3 clusters the stability is not as good (PAC ≤ 50%). To assess the clinical relevance of the obtained subtypes, we test whether they are associated with significantly distinct survival outcomes (Fig 7a). With the raw data, patient stratification is never significantly associated with clinical data. With NetNorM, significant associations of patient subtypes with survival times are achieved for HNSC, OV, KIRC and SKCM (Fig 7c), while with NSQN, a significant association is only achieved for OV. The stratification based on NetNorM remains prognostic beyond clinical data for SKCM (Likelihood ratio test, P = 2.4 × 10−2 (SKCM, N = 5)). It can be surprising at first sight that no signal is recovered for LUAD with NetNorM and for SKCM with NSQN since some signal was observed in the survival prediction setting in these cases. We hypothesized that this could be due to a bad choice of the hyperparameters k and α for these cancer types. Therefore additional experiments were run for LUAD and SKCM with k and α set to their values learned by cross-validation for the survival prediction task (S3 Table). This corresponds to k = 315 and α = 0.6 for LUAD (instead of k = 189 and α = 0.5 as defaults) and k = 140 and α = 0.25 for SKCM (instead of k = 243 and α = 0.5 as defaults). With these new values for the hyperparameters, significant associations with survival are detected for LUAD with NetNorM (for 4, 5 and 6 clusters) and for SKCM with both NetNorM (for any number of clusters) and NSQN (for 4 clusters) (S6 Fig). The recovery of a signal in these cases is in accordance with the results in the supervised setting. Overall, these results confirm the findings of [31] that network-based normalisation with NSQN allows stratifying patients better than the raw mutation profiles, and also show that the stratification obtained from NetNorM normalisation is both more stable and more clinically relevant than the one obtained with NSQN.
We now assess whether the biological information contained in Pathway Commons is crucial to obtain subtypes with significantly distinct survival outcomes. For that purpose, we carry out patient stratification with NSQN and NetNorM using 10 randomised versions of Pathway Commons for HNSC, OV, KIRC and SKCM. As for the survival prediction experiment, the randomisation involves shuffling the vertices’ labels so as to keep the structure of the network unchanged. Surprisingly, network randomisation does not affect the log-rank statistic obtained for HNSC and SKCM. This suggests that although NetNorM generates subtypes with more distinct survival times than NSQN for HNSC and SKCM, it does not benefit from Pathway Commons gene-gene interaction knowledge. Rather it exploits the prognostic information contained in the raw mutation profiles as well as the overall mutational burdens as captured by proxy mutations. Regarding KIRC and OV, NetNorM produces subtypes with significantly different survival times with 4 and 5 clusters for KIRC and for any number of clusters for OV. In the case of KIRC, the real network yields the subtypes with the most distinct survival times (N = 5) (Fig 8) while in the case of OV, most randomized networks (at least 15 out of 20 for each number of clusters) produce subtypes with worse association to survival time. This indicates that for KIRC and presumably for OV, NetNorM takes advantage of gene-gene interaction knowledge to stratify patients into clinically relevant subtypes. This is also clearly the case for LUAD with NetNorM when the hyperparameter k is set to its value learned by cross-validation in the survival prediction setting (S6 Fig).
To interpret biologically the subgroups of patients identified by automatic stratification after NetNorM normalisation, we look at differentially mutated genes and pathways across subtypes. We focus on LUAD with N = 5 groups as a proof of principle with k set to its value learned by cross-validation in the supervised setting. This choice is motivated by the fact that LUAD is the most promising cancer type for supervised survival prediction and produces interesting results in the unsupervised setting. As the basis vectors or “metapatients” yielded by the NMF summarise the mutational patterns found in the different subtypes, we analyse genes in terms of their weight in the different metapatients, and restrict our attention to the approximately 900 genes displaying highest variance (variance greater than 0.01) across basis vectors since these genes are expected to be the most differentially mutated across subtypes. Interestingly, this gene list comprises most significantly mutated genes in LUAD including TP53, KRAS, KEAP1, EGFR, NF1, RB1 [40, 41]. To analyse these genes we cluster them into groups with similar weights across basis vectors using hierarchical clustering (Fig 9b), and we test for enrichment in known biological pathways the 20 gene clusters (GCs) obtained.
One first observation is that the 5 patient subtypes have distinct overall mutational burdens with groups 4 and 5 (resp. 2 and 3) gathering patients with many (resp. few) mutations (Fig 9e). This confirms the fact that NetNorM-normalised profiles contain information about the initial number of mutations, although they are normalised to a fixed number of mutations. More importantly, most GCs exhibit high weights in one metapatient and low weights in others, suggesting that they are mainly enriched in mutations in one single patient subtype (Fig 9b). χ2 contingency tests (see methods) for each GC confirms that for most of them (17/20), the distribution of the mutations across patient subtypes is not that expected according to subtypes’ overall mutational burdens (P < 5 × 10−2) (S4 Table). The contribution of each subtype to the test statistic for each GC also confirms that GCs are often enriched in mutations in mainly one patient subtype (Fig 9d). Subtypes could thus easily be associated with one or several GCs, and therefore pathways through pathway enrichment analysis using the KEGG database [54] (see Methods).
Consequently, subtype 3 is characterised by an enrichment in mutations in genes associated with ribosomes and spliceosomes (GCs 2, 3, 4, 5, 6, 7, 8, 17, 18, 19) (S4 Table). Subtype 1 is enriched in mutations in two very small gene clusters (GCs 11 and 16): the first one consists of four genes including KRAS and the second one only includes MUC16. These two subtypes are those with poorest survival probability. Subtype 4 is mainly enriched in late replicating genes (GC 10) (Fig 9c). This reflects the fact that subtype 4 is enriched in highly mutated patients as there exists a positive correlation between somatic mutation frequency and genes replication time [16]. Subtype 2 is enriched in mutations in genes related to endocytosis and phagosomes (GCs 16, 1, 11). Finally, subtype 5 is very strongly associated with gene clusters 9 and 13. Gene cluster 9 is enriched in genes from the cAMP and PI3K-Akt signaling pathways. Gene cluster 13 could not be significantly associated to a known biological pathway. However it contains FANCD2 (Fanconi Anemia Complementation Group D2) which is involved in double-strand breaks DNA repair and the maintenance of chromosomal stability [55]. We note that 12 of the 15 patients in subtype 4 present the same 4-nucleotides splice site deletion in FANCD2, whereas across the rest of the 430 patients FANCD2 is mutated in 6 patients only, and only one of these 6 mutations is the same as that observed in subtype 4 patients.
Exploiting the wealth of cancer genomic data collected by large-scale sequencing efforts is a pressing need for clinical applications. Somatic mutations are particularly important since they may reveal the unique history of each tumour at the molecular level, and shed light on the biological processes and potential drug targets dysregulated in each patient. Standard statistical techniques for unsupervised classification or supervised predictive modelling perform poorly when each patient is represented by a raw binary vector indicating which genes have a somatic mutation. This is both because the relevant driver mutations are hidden in the middle of many irrelevant passenger mutations, and because there is usually very little overlap between the somatic mutation profiles of two individuals. NetNorM aims to increase the relevance of mutation data for various tasks such as prognostic modelling and patient stratification by leveraging gene networks as prior knowledge.
One important aspect of NetNorM is the property that, after normalisation, all patients have the same number of 1’s in their normalised mutation profile. Although there is no biological rational for this constraint, we believe that the fact that all normalised samples have the same distribution of values is an important property for many high-dimensional statistical methods such as survival models or clustering techniques to work properly. To support this claim, we notice that the Network-based stratification (NBS) method proposed in [31] performs a quantile normalisation step after network smoothing. To investigate whether the quantile normalisation step in NSQN plays an important role, we applied network smoothing without quantile normalisation (called NS) and performed survival prediction and patients stratification with this representation of the mutations. Surprisingly, NS does not improve over the raw mutation profiles for both LUAD and SKCM (Fig 10c). Moreover just as the raw data, NS is unable to stratify patients into approximately balanced clusters (Fig 10b). This suggests that quantile normalisation plays a crucial role in the performances obtained with NSQN, in spite of non obvious biological justification for this step.
Another important difference between NSQN and NetNorM is the fact that NetNorM only exploits mutation information about direct neighbours in the network, while NSQN can potentially diffuse a mutation further than the direct neighbours. However, we found that NSQN does not benefit from this possibility. Indeed, we tested a simplified version of NSQN where the network propagation is stopped after one iteration, and assessed the performance of the corresponding method which we call SimpNSQN. For survival prediction, we observe no significant difference between NSQN and SimpNSQN (Fig 10c). For patient stratification, SimpNSQN produces subtypes that are vey similar to those produced by NSQN (Fig 10d). Therefore the subtypes generated by both methods associate equally well to clinical data, and even slightly better for SimpNSQN in the case of LUAD (Fig 10a). Overall, these pieces of information indicate that the useful information created by NSQN is mostly concentrated on shared mutated order 1 neighbourhoods, and explain why we observe no loss in performance with NetNorM which explicitly restricts the diffusion of mutations to direct neighbours only. More generally, these elements also indicate that diffusion to indirect neighbours is still difficult with current methods. This is a likely consequence of the small world property of biological graphs [56]. Because the path between any two genes is usually short, diffusion even to order-2 neighbours reaches a substantial number of genes, and therefore the resulting signal observed for one gene is the superposition of a large number of signals originating from close mutations.
NetNorM encodes information about patients’ total number of mutations in the raw data, and potentially can exploit it if this information is relevant for the problem at hand. However we found that the total number of mutations is a poor predictor or survival (Fig 10c), and a poor feature for LUAD patient stratification (Fig 10a). This confirms that NetNorM conserves useful information regarding both the total mutational burden of a patient and the distribution of the mutations on the gene network, and manages to leverage both types of information. In addition to mutational burdens, NetNorM also encodes information about genes’ NMB which proved to carry some prognostic power. The fact that NMB might reveal new insights into mutation profiles is an emerging idea supported by this study. Further support has been formalised with two recently published methods [57, 58] which rely on NMB to achieve state-of-the-art performances for cancer gene discovery.
We emphasize that randomised gene networks lead to significantly worse performances than the real network for survival prediction as well as for patient stratification for several cancers. While it is not always clear whether incorporating gene networks as prior knowledge does help for a given task, this provides a sound argument that such prior knowledge is effectively leveraged with NetNorM.
Increasing the relevance of mutation data to various tasks is a broad project and NetNorM could be extended in many ways. First, although NetNorM was successful for LUAD and SKCM, we note that the method brings few improvements compared to the raw data for the remaining cancer types. Therefore extensive efforts are needed to determine whether it is possible to design representations of mutations that would increase the statistical power of models learned on these datasets. Second, NetNorM does not integrate further information about mutations such as their predicted functional impact. A possible extension could therefore include this type of information. Finally, the distribution of values for the normalised profiles is defined as the mean distribution of the original profiles in the case of NSQN, and simply a binary vector with a fixed number of 1’s in the case of NetNorM, however these choices are empirical. This suggests that an interesting future work may be to assess more precisely the effect of this distribution and, perhaps, optimise it for each specific task.
Whole exome somatic mutation calls (MAF files) were downloaded from TCGA data portal (https://tcga-data.nci.nih.gov/tcga) for 8 cancer types (LUAD, SKCM, GBM, BRCA, KIRC, HNSC, LUSC, OV) (Table 1). The data include point mutations (single nucleotide polymorphism as well as di/tri/oligo-nucleotide polymorphism) and indels. Silent mutations were filtered out and mutations profiles were defined as binary vectors with ones whenever a patient is mutated in a given gene and zeros otherwise.
Pathway Commons (http://www.pathwaycommons.org/pc2/downloads) was used throughout this work (Pathway Commons v6, SIF format). It integrates gene networks from several public databases and aggregates both genetic and protein-protein interactions (PPIs). PPIs refer to physical contacts established between proteins while genetic interactions refer to interactions through regulatory and signalling pathways. To remove interactions involving small molecules in Pathway Commons, the following interaction types were filtered out: “consumption-controlled-by”, “controls-production-of”, “controls-transport-of-chemical”, “chemical-affects”, “reacts-with”, “used-to-produce”, “SmallMoleculeReference”, “ProteinReference;SmallMoleculeReference”, “ProteinReference”. We obtained a network with 16,674 nodes (genes) and 2,117,955 edges (interactions). For the survival prediction task, we also tested the following gene networks: BioGRID v3.4.131, HPRD release 9, HumanNet v1 and STRING v10. For HumanNet and STRING, only the top 10% most confident interactions were retained.
NetNorM is a method that integrates patients mutation profiles with a gene network to produce normalised mutation profiles where all patients have the same number k of mutations. The target number of mutations k is a tuning parameter. In the context of survival prediction (supervised setting), it is learned by cross-validation while for patient stratification (unsupervised setting), it is set as the median number of mutations in a cohort, or alternatively to the median best k learned across cross-validation folds for survival prediction. Concretely, NetNorM defines a ranking over genes separately for each patient and then use this ranking to normalise mutation profiles. The ranking defined in NetNorM is obtained with a simple two-step procedure. First, genes are ranked according to their mutation status with mutated genes ranked higher than non mutated genes. Then, mutated genes are ranked according to their degree (i.e. their number of neighbours) and non mutated genes are ranked according to their number of mutated neighbours. The normalisation is then obtained by considering the k highest ranked genes as mutated while the rest of the genes will be considered non mutated. By construction, mutated genes are always ranked higher than non-mutated genes. Therefore patients with a lot of mutations will have mutations removed while patients with few mutations will hold artificial proxy mutations. Note that when the obtained ranking contains ties, all genes are given distinct ranks according to the order in which they occur in the mutation matrix.
Network smoothing propagates the influence of mutations over gene-gene interaction networks. It was implemented according to the following update function [31]:
X t + 1 = α X t D - 1 2 A D - 1 2 + ( 1 - α ) X 0
where Xt is the patient × genes mutation matrix at iteration t, X0 is the initial binary mutation matrix, A is the adjacency matrix representing the network used and D is the diagonal degree matrix where Dii=∑jAij. α is a tuning parameter controlling the length of diffusion paths over the network. Similarly to the parameter k in the context of NetNorM, it is learned by cross-validation for survival prediction (supervised task) while for patient stratification (unsupervised task) it is set as α = 0.5 as recommended in [31] with Pathway Commons or alternatively to the median best α learned across survival prediction cross-validation folds. The update function is applied until convergence, and the resulting smoothed matrix is then quantile normalised so that all patients have the same mutation distribution.
The simplified version of NSQN does not propagate mutations further than to order 1 neighbours in the network. More precisely, the SimpNSQN score of a gene is equal to its number of mutated neighbours normalised by its degree and by the degrees of its neighbours, plus a constant if the gene is mutated. This is obtained by computing:
X = α X 0 D - 1 2 A D - 1 2 + ( 1 - α ) X 0
where X0 is the initial binary mutation matrix, A is the adjacency matrix representing the network used, D is the diagonal degree matrix where Dii=∑jAij and α ∈ R is a tuning parameter. Note that SimpNSQN uses the same update equation as NSQN but it is run only once.
To estimate a survival model from high-dimensional mutation profiles, we use a survival SVM model [59] combined with a sparsity-inducing regularisation to automatically perform gene selection. Let δi = 1 (resp. δi = 0) if patient i is deceased (resp. censored), and y i ∈ R be the observed survival time of patient i. It corresponds to either a failure or a censoring time depending on whether the patient is deceased or censored. Define Z ∈ {0, 1}n×n which indicates whether a pair of patients is comparable, i.e,
Z i j = { 1 if ( y i < y j and δ i = 1 ) or ( y j < y i and δ j = 1 ) , 1 if ( y i = y j and ( δ i = 1 or δ j = 1 ) ) , 0 otherwise .
Finally, let xi ∈ {0, 1}p be the mutation profile of patient i. The survival time of patient i is modelled as si = wT xi where w ∈ R p is the model parameter learned using ranking Support Vector Machines (rSVM) as in [59]. However to get a sparse w we introduce an ℓ1 regularisation instead of the ℓ2 regularisation in [59] and thus solve the following optimisation problem:
minimise w 1 2 | | w | | 1 + C ∑ i , j Z i j ℓ h i n g e ( w T ( x j - x i ) ) ,
where ℓhinge(u) = max(1 − u, 0) is the hinge loss and C ∈ R is the regularisation parameter. To solve this problem we used the support vector classification algorithm svm.LinearSVC from the Python package scikit learn [60]. This optimisation problem maximises a convex relaxation of the Concordance Index (CI) which measures how well the predicted survival times s are in accordance with the observed survival times y for the comparable pairs of patients. Formally, CI=1| Z |∑yi≤yjZijI(sj−si) where
I ( x ) = { 1 if x > 0 , 1 2 if x =0 , 0 otherwise ,
and | Z | = ∑ y i ≤ y j Z i j. To evaluate the CI obtained on a given dataset, samples were split in 80% train and 20% test sets 20 times using 4 five-fold cross-validation. Each time, a model was learned on the training set and tested on the test set. The CI was computed according to a python implementation of the function estC from the R package compareC. Hyperparameters were learned thanks to an inner 5-fold cross-validation on the training set. The values tested for C ranged from 1 × 10−4 to 1 × 102 included in log scale. The values tested for α ranged from 0.1 to 0.9 included with steps of 0.1. Finally the values tested for k were chosen to span a grid from kmin and kmax with steps of 2, where kmin and kmax are the first and third quartiles of the distribution of patients’ total number of mutations. kmin and kmax differ for each cohort (S2 Table).
Let X ∈ R n × p be the matrix with patient mutations profiles as rows. To cluster the patients we perform a non-negative matrix factorisation (NMF) on X, i.e., solve the following optimisation problem:
minimise W , H > 0 | | X - W H | | 2 2 ,
where H ∈ R N × p defines N basis vectors or “metapatients” and W ∈ R n × N defines basis vectors loadings. Patient i was then assigned to the group j ∈ {1‥N} that represents him best i.e. a r g m a x j W i j. To promote robust cluster assignments, NMF was applied 1000 times to subsamples of the dataset composed of 80% of the samples and 80% of the features chosen at random without replacement. A consensus matrix C ∈ R n × n was then derived from the 1000 cluster assignments obtained where each entry Cij corresponds to the frequency at which two patients where clustered in the same group over all samplings where both patients were retained. The final cluster assignment was obtained by applying hierarchical clustering to the consensus matrix with euclidean distance and average linkage.
To assess the stability of the obtained clusters, we computed the proportion of ambiguous clustering (PAC) which is the proportion of discordant cluster assignments obtained through consensus clustering. Cluster assignments for a pair of patients (i, j) were considered discordant when 0.25 ≤ Cij ≤ 0.75.
In the case where only the total number of mutations was used for stratification, NMF is not applicable and kMeans was used instead with 1000 restarts and initialisation by kMeans++ [61].
Several proxy genes have a prognostic power according to log-rank tests performed for each gene separately and which compare patients with mutations (proxy or not) versus those without (P ≤ 1 × 10−2). The difference in survival outcomes observed may be due to at least two types of information encoded in proxy genes: patients’ overall mutational burden and genes’ neighbourhood mutational burden (NMB). To clarify the contributions of each effect, we investigate whether such distinct survival outcomes can be obtained with proxies for the total number of mutations only, regardless of NMBs. To this end, we simulate proxy mutations for each gene separately according to a model that only depends on patients’ total number of mutations. Let T i ∈ N be the total number of mutations of patient i, i ∈ {1, …, n}. Let M o ⊂ { 1 , . . . , n } and M p ⊂ { 1 , . . . , n } indicate which patients have original and proxy mutations respectively. For a given proxy gene whose mutations are described by the sets Mo and Mp, we leave the original mutations untouched and reallocate the proxy mutations according to
P ( i ∈ M p | T i ) = { 0 if ( T i ≥ k ) or ( i ∈ M o ) k - T i α otherwise
where α is chosen so that the probabilities sum to 1. Proxy mutations are drawn from this model 1000 times. Each time we compute the log-rank statistic between the mutated and non mutated patients which yields a distribution of the log-rank statistic under the null hypothesis. The actual log-rank statistic obtained using NetNorM is then compared to this distribution to accept or reject the null hypothesis. Rejecting the null hypothesis means that the difference in survival outcomes observed between the patients with and without artificial mutations is not only driven by patients’ total number of mutations.
To determine whether the obtained patient subtypes are predictive of survival beyond clinical data, we fitted a Cox proportional hazards regression model to the clinical data and to the clinical data augmented with a variable describing patients’ subtypes. We then performed a likelihood ratio test to compare the two models. The clinical variables used were downloaded from TCGA. It includes age, gender, stage, extent of spread to the lymph nodes, presence of metastasis, histology for both LUAD and SKCM and further variables such as smoking history, history of prior malignancy, residual tumour after surgery, tumour dimensions for LUAD and clark level at diagnosis, primary melanoma mitotic rate, new tumour event after initial treatment (yes/no), primary melanoma tumour ulceration (yes/no), primary melanoma known (yes/no) for SKCM.
We obtain gene clusters by applying hierarchical clustering with centroid linkage and Euclidean distance to the columns of the metapatients matrix (restricted to high variance genes). To obtain a reasonable number of gene clusters to analyse, we cut the hierarchical cluster tree at a distance threshold of 5.5, yielding 20 clusters. Gene clusters can be categorised into two types: those that contain a lot of proxy mutations (≥ 80% of the total mutational load of the cluster) and whose genes form a dense subgraph, and those that have neither of these two features. The presence of dense subgraphs with many proxy mutations results from the fact that NetNorM tends to add proxy mutations to all genes in a dense subgraph or none since they all have roughly the same number of mutated neighbours. The association of a gene cluster with one subtype can therefore indicate two things: either the subtype is expected to be enriched in proxy mutations in the corresponding gene cluster, which in turn indicates that the subgraph in which the cluster lies is expected to be enriched in mutations, or the gene cluster itself is expected to be enriched in mutations in the corresponding subtype. The enrichment or depletion in mutations of one gene cluster across patient subtypes was therefore tested slightly differently according to the gene cluster type. In the first case, we first define the neighbourhood of the gene clusters as all genes lying in the same dense subgraph. Specifically, we include in the subgraph all genes sharing an edge with at least 90% of the genes in the cluster, thus keeping subgraphs very dense. The obtained set of genes is the one tested for enrichment in mutations across subtype. In the second case, the gene cluster is directly tested for enrichment. Enrichment is assessed with a χ2 contingency test, where the contingency table is defined by the following marginals: the total number of raw mutations in each subtype, and the total number of raw mutations in and outside the gene cluster (generalised to the embedding of a dense subgraph if it is relevant).
Gene clusters are searched for pathway enrichment using DAVID online tool [62] (https://david.ncifcrf.gov/summary.jsp) with the KEGG database [54]. They are also tested for enrichment in late replicating genes thanks to a permutation test using data downloaded from http://www.broadinstitute.org/cancer/cga/mutsig_run. For each gene cluster c of length lc, lc genes are chosen uniformly at random without replacement from the list of genes with replication time information. This sampling is performed 1000 times and the null distribution was obtained by computing the median replication time of these 1000 gene sets. The median replication time of cluster c is then compared to the null distribution to yield a p-value, i.e. the probability to observe a set of genes of length lc with median replication time at least as extreme.
|
10.1371/journal.pgen.1006525 | Functional Requirements for Heparan Sulfate Biosynthesis in Morphogenesis and Nervous System Development in C. elegans | The regulation of cell migration is essential to animal development and physiology. Heparan sulfate proteoglycans shape the interactions of morphogens and guidance cues with their respective receptors to elicit appropriate cellular responses. Heparan sulfate proteoglycans consist of a protein core with attached heparan sulfate glycosaminoglycan chains, which are synthesized by glycosyltransferases of the exostosin (EXT) family. Abnormal HS chain synthesis results in pleiotropic consequences, including abnormal development and tumor formation. In humans, mutations in either of the exostosin genes EXT1 and EXT2 lead to osteosarcomas or multiple exostoses. Complete loss of any of the exostosin glycosyltransferases in mouse, fish, flies and worms leads to drastic morphogenetic defects and embryonic lethality. Here we identify and study previously unavailable viable hypomorphic mutations in the two C. elegans exostosin glycosyltransferases genes, rib-1 and rib-2. These partial loss-of-function mutations lead to a severe reduction of HS levels and result in profound but specific developmental defects, including abnormal cell and axonal migrations. We find that the expression pattern of the HS copolymerase is dynamic during embryonic and larval morphogenesis, and is sustained throughout life in specific cell types, consistent with HSPGs playing both developmental and post-developmental roles. Cell-type specific expression of the HS copolymerase shows that HS elongation is required in both the migrating neuron and neighboring cells to coordinate migration guidance. Our findings provide insights into general principles underlying HSPG function in development.
| During animal development, cells and neurons navigate long distances to reach their final target destinations. Migrating cells are guided by extracellular molecular cues, and cellular responses to these cues are regulated by heparan sulfate proteoglycans. Heparan sulfate proteoglycans are proteins with long heparan sulfate polysaccharide chains attached. Here we identify and study previously unavailable viable mutants that disrupt the elongation of the heparan sulfate chains in the nematode C. elegans. Our analysis shows that these HS-chain-elongation mutations affect the development of the nervous system as they result in misguided migrations of neurons and axons. Furthermore, we find that heparan sulfate chain elongation occurs in numerous cell types during development and that the coordinated production of heparan sulfate proteoglycans, in both the migrating cell and neighboring tissues, ensures proper migration. Our findings highlight the critical roles of heparan sulfate proteoglycans in nervous system development and the evolutionary conservation of the molecular mechanisms driving guided migrations.
| Cell migration is key to animal development and physiology. To reach their targets, migrating cells rely on guidance factors and morphogens, which can be regulated by heparan sulfate proteoglycans (HSPGs) [1]. HSPGs are cell-surface or extracellular proteins characterized by the attachment of heparan sulfate (HS) polysaccharide chains to the extracellular domain of their core protein [2]. HSPGs interact with molecules at the cell surface and in the extracellular matrix via both their HS chains and core proteins, and can function as co-factors that regulate the distribution of morphogens and that modulate the interactions between extracellular ligands and their receptors [1, 3]. HSPGs have been shown to be part of multiple signaling pathways across species and to be key to multiple developmental events, including those elicited by guidance cues such as Slit and Netrin, and morphogens such as Hhg, FGF, Sonic Hedgehog, Wnts, and BMPs [1, 2, 4, 5].
The importance of HSPGs during animal development has been extensively studied using mutations that disrupt individual HSPG core proteins. A number of HSPGs have been characterized in C. elegans using mutations that affect specific core proteins, such as mutations in sdn-1/syndecan, lon-2/glypican, cle-1/collagen type XVIII, unc-52/perlecan, gpn-1/glypican, and agr-1/agrin [6–26]. These studies have uncovered precise roles of individual HSPGs in morphogenesis and nervous system development. For instance, loss of cle-1/collagen type XVIII leads to synaptic defects at neuromuscular junctions, as well as specific neuronal cell and axon guidance defects [6, 7]; unc-52/perlecan promotes ectopic presynaptic bouton growth and affects the 4° dendritic branching of the neuron PVD [24, 25]; sdn-1/syndecan mutants exhibit a number of neuronal cell and axon guidance defects [12–14, 21]; and lon-2/glypican is important for cell and axon guidance [8, 12, 17, 21], particularly for netrin-mediated guidance events [8, 13]. Moreover, studies where two or three HSPG core proteins have been simultaneously mutated in double and triple mutants demonstrated that the combined actions of precise HSPGs ensure proper guidance of neurons and axons during development [8, 12, 13, 21]. All these analyses of mutations affecting specific HSPG core proteins have been instrumental to address the roles of HSPGs in development. However, mutations that remove particular HSPG core proteins inevitably also remove the HS chains that would have been attached to the missing core proteins. Therefore, the phenotypic consequences of such mutations in HSPG core proteins can be due to the absence of either the core protein or the HS chains, or both, making it difficult to extract the functional contribution of the HS chains per se with such analysis of HSPG core protein mutants.
To address the roles of the HS chains that are attached to HSPG core proteins, various mutations that affect HS chain biosynthesis have been analyzed [1]. HS chains are linear glycosaminoglycan (GAG) polysaccharides composed of alternating repeats of D-glucuronic acid (GlcA) and N-acetylglucosamine (GlcNAc) [27]. HS chain biosynthesis in the Golgi apparatus can be divided into three phases: (1) initiation, (2) elongation, and (3) chemical modifications.
First, a HS chain is initiated by the addition of a tetrasaccharide linker (synthesized by the step-wise addition of a xylose residue, a galactose residue, a galactose residue, and a GlcA residue) on a specific Ser residue of the HSPG core protein. This initiation step is catalyzed by a set of four initiation enzymes encoded by the glycosyltransferases genes sqv-6, sqv-3, sqv-2, and sqv-8 in C. elegans [28–30]. Mutations in these four initiation genes have been characterized, revealing important morphogenetic roles in embryogenesis and vulva development [28, 29, 31]. However, these four initiation enzymes add the same tetrasaccharide linker also to the core proteins of chondroitin sulfate proteoglycans (CSPGs), as they also catalyze the initiation of chondroitin sulfate (CS) chains. Thus, specific roles for HS chains cannot be addressed in these mutants in which the initiation of both HS and CS chains is affected, with phenotypes resulting from the combined disruption of both HSPGs and CSPGs.
Once initiated, the second phase of HS biosynthesis is the elongation of HS chains. HS chain elongation is catalyzed by the HS copolymerase, a heterodimer composed of two glycosyltransferases of the EXT family. HS chain elongation has been shown to be crucial to animal development across metazoans, as its dysfunction results in pleiotropic consequences including abnormal morphogenesis and tumor growth. In C. elegans, null mutations in the exostosin glycosyltransferases rib-1 and rib-2 are embryonic lethal, indicating that HS elongation is essential for morphogenesis [32–34]. In Drosophila, null mutations in the exostosin genes tout-velu, brother of tout-velu and sister of tout-velu are lethal, and loss of their function leads to severe patterning defects with abnormal morphogen signaling in many developmental contexts [35–39]. For example, tout-velu mutants exhibit a lack of diffusion of Hh in the wing imaginal disc [35]. In zebrafish, mutations in EXT family members ext2 (dackel) and extl3 (boxer) are also lethal [40]. In mice, complete loss of function of EXT1 induces defective gastrulation and embryonic lethality [41]. Mutations in glycosyltransferases genes EXT1 and EXT2 in humans result in a dominant disorder called hereditary multiple exostoses, characterized by cartilage-capped skeletal tumors known as osteochondromas, which results in skeletal abnormalities and short stature [42–47]. Although the osteochondromas are most often benign tumors, malignant transformation into chondrosarcomas or osteosarcomas occurs in ~2% of HME patients [48]. However, their exact roles in bone development and homeostasis are not well understood.
Given the lethality associated with the complete loss of HS elongation across species, genetic analysis of such mutations has been possible for early developmental roles or for conditions where the gene function is only partially lost. For instance, conditional knockouts have been used to study later developmental roles of the HS copolymerase in mice [49, 50], and partially maternally rescued mutants (displaying partial phenotypes) have been studied in zebrafish [40] and in C. elegans [21, 32–34]. Indeed, C. elegans deletion mutations in the HS copolymerase genes rib-1 and rib-2, namely rib-1(tm516 or ok556) and rib-2(tm710 or qa4900) die as embryos: homozygous mutant progeny from a homozygous mutant mother (animals genotypically m-/- z-/-, where “m” and “z” indicate the maternal and zygotic genotypes, respectively), all die as embryos. In contrast, first generation homozygous mutant animals from a heterozygous mother (that is to say animals that are genotypically rib-1m+/- z-/- or rib-2m+/- z-/-) are “maternally rescued”; they complete development and become adults, due to wild-type gene product inherited from their heterozygous mothers (Table 1) [32–34]. Such adult rib-1(tm516)m+/- z-/- and rib-2(qa4900)m+/- z-/- animals are largely but not completely maternally rescued, as they display locomotion defects and cannot lay eggs normally, becoming filled with their dead m-/- z-/- progeny [32–34]. It has been shown that one quarter of rib-1(tm516)m+/- z-/- maternally rescued animals exhibit axon guidance defects for the neuron HSN, while no HSN defect was seen in rib-2(tm710)m+/- z-/- maternally rescued animals [21, 32]. Thus, a thorough phenotypic analysis of rib-1 and rib-2 deletion mutants has been limited by both (a) the early embryonic lethality of rib-1m-/- z-/- and rib-2m-/- z-/- mutants, where later developmental stages cannot be examined, and (b) the presence of wild-type maternal product in rib-1m+/- z-/- and rib-2m+/- z-/-, which profoundly rescues development, masking HS functional requirements [21, 32–34]. The function of HS elongation in C. elegans has also been examined by RNAi knockdown of rib-1 and rib-2, where the HSN axon guidance defects were more penetrant than in maternally rescued animals [17]. Yet, developmental defects are likely partially penetrant as there is no phenocopy of embryonic lethality by RNAi knockdown of rib-1 and rib-2 [17]. Thus, the availability of partial loss-of-function mutations for the genes rib-1 and rib-2 would allow, if viable, a systematic study of developmental roles for HS elongation during the development of the nervous system in C. elegans.
After HS chain elongation, the third phase of HS biosynthesis is the chemical modification of HS chains by modifying enzymes such as epimerases and sulfotransferases [1]. Research in C. elegans has elegantly addressed the requirements for HS chain modifications in nervous system development. Specific roles in the guidance of neuronal cell and axon migrations, including the modulation of distinct guidance cues, have been uncovered using mutations in the HS modifying enzymes epimerase hse-5 and in the sulfotransferases hst-2, hst-3.1, hst-3.2, and hst-6 [9–12, 14, 15, 18–22, 26]. For instance, in the contexts of PVQ axon guidance and D-type motoraxon guidance, slt-1/Slit signaling acts in the same pathway as hse-5 and hst-6, suggesting HSPGs modified by hse-5 and hst-6 may function with slt-1/Slit to guide the axons of both PVQ and motorneurons [10]. In addition, ectopic hypodermal expression of hst-6 disrupts the guidance of the axon of the motorneuron DB7, and is dependent upon both lon-2/glypican and slt-1/Slit, suggesting that ectopic hypodermal 6O-sulfated LON-2/glypican may disrupt axon guidance through impacting slt-1/Slit signaling [11]. Also, interactions between HS modification enzymes and ephrin and integrin signaling have also been investigated in the contexts of PVQ and motoraxon guidance [10]. In contrast, whereas the core protein LON-2/glypican has been shown to function in netrin-mediated guidance [8], it remains to be determined which specific HS chain modifications may be important for the unc-6/Netrin signaling pathway.
Collectively, all of this remarkable prior work on the roles of HSPGs for nervous system development in C. elegans, targeting either core proteins or HS chain chemical modifications, has yielded a view in which sets of specific HSPGs, with distinct HS chemical modification patterns, interact with specific guidance pathways and contribute to context-dependent guidance decisions during the nervous system assembly. However, a general outlook on the functions of HSPGs in nervous system development by specifically disrupting the presence of HS chains across all HSPGs has been unavailable.
Here we report the identification of viable partial loss-of-function mutations in the two HS copolymerase glycosyltransferase genes rib-1 and rib-2 of C. elegans. We show that these mutations reduce HS levels and affect cell and axonal migration during nervous system development. We find that the HS copolymerase is expressed dynamically during morphogenesis, and that expression is sustained throughout life in specific cell types, consistent with HSPGs playing both developmental and post-developmental roles. Our findings indicate that proper axon guidance during the development of the nervous system requires coordinated HS chain elongation in both the migrating neuron itself and adjacent cells that secrete the extracellular matrix along which the growth cone extends. Our analysis highlights the functional importance of HSPGs during animal development.
The analysis of uncoordinated mutants of C. elegans has uncovered key genes underlying nervous system development and function over the past decades [51–53]. In order to identify new genes required for neuronal development, a genetic screen searching for maternally-rescued uncoordinated mutants was carried out by Hekimi et al. [54]. Briefly, an F2 clonal screen was performed, where P0 animals were mutagenized, from which individual F1 hermaphrodites were isolated and allowed to self-fertilize, followed by the selection of individual wild-type F2 hermaphrodites that were allowed to self-fertilize, to finally screen for and select broods where all the F3 animals were uncoordinated/abnormal [54]. Such a scheme allowed for the isolation of maternally rescued mutants, where F2 animals appear phenotypically normal even though they are genotypically homozygous mutant (m+/- z-/-), due to wild-type gene product provided by the heterozygous F1 mother. Only in the next generation of homozygous mutants derived from a homozygous mutant mother, namely F3 m-/- z-/- animals, does the abnormal phenotype manifest. Several maternally rescued uncoordinated mutants were isolated in this screen [54–56], two of which, qm32 and qm46, had similar phenotypes and were mutations in genes mum-1 and mum-3, respectively (mum stands for maternal-effect uncoordinated and morphologically abnormal) [54].
mum-1(qm32) and mum-3(qm46) homozygous mutant animals from a homozygous mutant mother have severe defects: 32% of mum-1(qm32)m-/- z-/- and 15% of mum-3(qm46)m-/- z-/- die as embryos, and of the embryos that hatch, 80% of mum-1(qm32)m-/- z-/- larvae and 26% of mum-3(qm46)m-/- z-/- larvae die before reaching adulthood and display morphological abnormalities ([54]; summarized in Table 1). Fortunately, a proportion of mum-1(qm32)m-/- z-/- and mum-3(qm46)m-/- z-/- mutant animals are viable and complete development to become fertile adults [54], which allows the easy propagation of the homozygous mutant strains. In fact, 14% of mum-1(qm32)m-/- z-/- animals and 63% of mum-3(qm46)m-/- z-/- animals reach adulthood, all of which are uncoordinated and egg-laying defective [54]. Importantly, mum-1(qm32) and mum-3(qm46) mutations are recessive: animals genotypically heterozygous (z+/-) exhibit a wild-type phenotype, irrespective of the maternal genotype (m-/- or m+/-) [54]. Moreover, mum-1(qm32)m+/- z-/- and mum-3(qm46)m+/- z-/- animals are fully maternally rescued: m+/- z-/- animals develop normally, display no lethality or morphological abnormalities, and become adults that are indistinguishable from the wild type, locomoting and laying eggs normally [54]. Worth noting, this maternal rescue effect is incomplete when both mum-1(qm32) and mum-3(qm46) mutations are combined. Indeed, first generation double homozygous mutant animals from doubly heterozygous mothers, i.e. mum-1m+/- z-/-; mum-3m+/- z-/-, develop into adults that are uncoordinated and egg-laying defective, becoming bloated with a brood of dead embryos that are mum-1m-/- z-/-; mum-3m-/- z-/- double homozygous mutants. Thus, the double mutant mum-1(qm32); mum-3(qm46) could not be generated (we attempted to build it by several crossing schemes), as 100% of the animals die as embryos when there is no wild-type maternal contribution. Finally, the phenotype of both mum-1(qm32) and mum-3(qm46) over a deficiency is lethal, suggesting that the null phenotype of these two genes is lethal [54].
To uncover what genes were disrupted in the mutants qm32 and qm46 we determined the molecular identity of the lesions in these two mutants. We found that qm32 and qm46 are alleles of the genes rib-1 and rib-2, respectively, as we demonstrate here. First, we narrowed down the genetic position of mum-1(qm32) by genetic mapping and then assayed cosmids corresponding to the genetic position of mum-1(qm32) for transformation rescue (Fig 1B). We found that cosmid F12F6 fully rescued the mum-1(qm32) mutants for uncoordination, egg laying defects, and abnormal larval morphology and lethality (Fig 1A and 1B). We tested PCR products corresponding to each of the genes located on this cosmid and found that a 9 kb PCR product containing the gene F12F6.3/rib-1 fully rescued all of the above mutant phenotypes of mum-1(qm32) (Fig 1A and 1B). In addition, construct Prib-1::rib-1(+) completely rescued the cell and axon guidance defects of the neuron AVM, the axon guidance defects of the neurons PVQ in mum-1(qm32) mutants (Fig 1C), and their behavioral defects. We verified the predicted exon structure of the gene by sequencing cDNA clone yk187a9. We sequenced the genomic region of rib-1 in mum-1(qm32) mutants and found that the qm32 molecular lesion is a T to A base pair change at position 39528 of cosmid F12F6, which converts the Stop codon of rib-1 into a Lys residue (Fig 1B). Thus, mum-1(qm32) corresponds to the gene previously known in the literature as rib-1, and we will refer to mum-1(qm32) as rib-1(qm32) from now on. The gene rib-1 is homologous to exostosin family members mammalian EXT1 and Drosophila tout-velu, and thus encodes one of the two glycosyltransferases that compose the C. elegans HS copolymerase, responsible for HS chain elongation (see below, Fig 1D).
The rib-1(qm32) mutation does not affect the transcript levels of rib-1 as assayed by RT-PCR (S1B Fig). Based on the sequence, it may result in the translation of an open reading frame present in the 3’UTR, which would possibly extend RIB-1 by 114 aa residues until the next in-frame Stop codon. The activity of the mutant RIB-1 protein in rib-1(qm32), or of the complex in which it functions (see below), is affected by the mutation. The rib-1(qm32) mutation is fully recessive, fully maternally rescued, and is completely rescued by expression of wild-type transgenic copies of the gene [54] (this study), suggesting that the predicted protein extension diminishes RIB-1 activity in rib-1(qm32) mutants, rather than being neomorphic. Consistent with the notion that qm32 is a partial loss-of-function mutation, the phenotype of qm32 over a deficiency is more severe (i.e., lethal [54]), and the phenotype of null deletion alleles rib-1(tm516) and rib-1(ok556) is also more severe than that of rib-1(qm32), as 100% of rib-1(tm516)m-/- z-/- and rib-1(ok566)m-/- z-/- animals die as embryos [32, 33], compared to 32% embryonic lethality in rib-1(qm32)m-/- z-/- [54]. In sum, these data indicate that the mutation qm32 is a hypomorphic mutation of the gene rib-1, where residual function allows 14% of the rib-1(qm32)m-/- z-/- mutants to be viable and become uncoordinated and egg-laying defective adults.
The second C. elegans HS glycosyltransferase and subunit of the HS copolymerase that catalyzes HS chain elongation is encoded by the gene rib-2. Given the phenotypic similarities between the mum-1/rib-1(qm32) and mum-3(qm46) mutants and that the genetic mapping position of mum-3(qm46) corresponded to a chromosomal interval containing the gene rib-2, we determined whether mum-3(qm46) was an allele of rib-2. We tested a 5.6 kb PCR product containing rib-2(+) for rescue of mum-3(qm46) mutants and found that their defects in larval development, locomotion, and egg laying were fully rescued by this transgene (Fig 2A and 2B). In addition, construct Prib-2::rib-2(+) completely rescued the cell and axon guidance defects of the neuron AVM and axon guidance defects of the neuron PVQ in mum-3(qm46) mutants (Fig 2C), as well as their developmental and behavioral defects. We verified the predicted exon structure of the gene by sequencing cDNA clone yk3c1. We sequenced the genomic region of the gene rib-2 in the mum-3(qm46) mutants and found that the qm46 molecular lesion changes a G to A at position 4366 of cosmid K01G5. The qm46 mutation results in an Arg to Gln amino acid substitution at conserved residue 434, which is near the exostosin domain in the 814 amino acid long RIB-2 protein (Fig 2D). Thus, mum-3(qm46) corresponds to the gene previously known in the literature as rib-2, and we will refer to mum-3(qm46) as rib-2(qm46) from now on. The gene rib-2 encodes the second glycosyltransferase subunit of the HS copolymerase and is most homologous to exostosin family members mammalian EXTL3 and Drosophila brother of tout-velu, and also fills the functional roles of EXT2 and Drosophila sister of tout-velu in C. elegans (Figs 2D and 3A).
Consistent with qm46 being a missense mutation, the levels of rib-2 transcript are comparable to wild type (S1B Fig). The rib-2(qm46) mutation is fully recessive, fully maternally rescued, and is completely rescued by expression of wild-type transgenic copies of the gene ([54]; this study), consistent with rib-2(qm46) being a partial loss-of-function mutation. Moreover, the phenotype of qm46 over a deficiency is more severe (i.e., lethal [54]), and the phenotype of null deletion alleles rib-2(tm710)m-/- z-/- and rib-2(qa4900)m-/- z-/-animals is also more severe as all embryos die [32–34]. In contrast, only 15% of rib-2(qm46)m-/- z-/- animals die as embryos [54]. Taken together, qm46 is a viable hypomorphic mutation in the gene rib-2, where residual function allows 63% of the rib-2(qm46)m-/- z-/- mutants to be viable and become adults that are uncoordinated and egg-laying defective.
The genes rib-1 and rib-2 each encode one of the two C. elegans HS glycosyltransferases that elongate HS chains (Fig 3A) [32, 57], which are composed of alternating GlcA and GlcNAc residues. After a tetrasaccharide linker has been synthesized on the HSPG core protein, the first step for HS chain elongation is the addition of a GlcNAc residue (Fig 3A) [57]. The addition of the first GlcNAc residue is catalyzed by RIB-2 in C. elegans, as demonstrated biochemically [57], similar to EXTL3 in mammals [58], and Brother of tout-velu in Drosophila [59]. HS chain elongation then proceeds by the repeated addition of disaccharide units of GlcA and GlcNAc (Fig 3A) [32]. This alternating addition of GlcA and GlcNAc is catalyzed by a heterodimer of RIB-1 and RIB-2, as demonstrated biochemically [32, 57], similar to EXT1 and EXT2 in mammals [60–63], and Tout-velu (Ttv) and the Sister of tout-velu in Drosophila [59].
Given the biochemically demonstrated roles of RIB-1 and RIB-2 [32, 57], the rib-1(qm32) and rib-2(qm46) mutations are expected to reduce HS chain elongation and result in decreased levels of HS in these mutants. To directly determine total HS levels in the rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- mutants, we performed Western blot analysis. For this, we extracted proteins from wild-type (N2) animals, rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- single mutants. We treated the protein extracts from these strains with a mix of heparinases I and III, and performed Western blot analysis using an antibody that specifically recognizes heparinase-digested HS chains (3G10, [64]). As expected, no signal above background was detected in untreated control samples (three left lanes, Fig 3B), compared to heparinase-digested samples (three right lanes, Fig 3B). Indeed, among the heparinase-digested samples, we found that compared to wild type, the HS content was severely reduced in rib-1(qm32) and rib-2(qm46) mutants (Fig 3B), confirming that the rib-1(qm32) and rib-2(qm46) mutations decrease HS biosynthesis, as predicted from their known biochemical functions [32, 57]. We examined whether the HS level reduction observed in these mutants could be rescued by transgenic expression of rib-1(+) and rib-2(+), respectively. For this, we extracted proteins from a strain of rib-1(qm32)m-/- z-/- mutants carrying a rib-1(+)-containing extrachromosomal array, and from a strain of rib-2(qm46)m-/- z-/- mutants carrying a rib-2(+)-containing extrachromosomal array. We found that transgenic expression of rib-1(+) and rib-2(+) re-elevates HS levels in the rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- mutants, respectively (right most lanes in Fig 3B). HS levels rescue appears incomplete likely due to the fact that by the time that the worms populations from these strains were collected, only ~10–20% of the animals actually carried the rescuing transgene (extrachromosomal arrays are lost at some frequency during cell divisions and over the course of generations [65]). Nevertheless, our results clearly indicate that the alleles of rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- strongly reduce the levels of HS compared to the wild type and that copies of the wild-type transgene re-elevate the HS levels. These results support that the rib-1(qm32) and rib-2(qm46) mutations reduce the function of the genes which are important for HS elongation, consistent with prior biochemical demonstration of their function [32, 57].
Having examined how rib-1(qm32) and rib-2(qm46) mutations impact global HS levels, we further examined their consequences on two specific HSPGs, LON-2/Glypican and SDN-1/Syndecan. In these experiments, to detect LON-2/Glypican, we expressed green fluorescent protein (GFP)-tagged LON-2 (LON-2::GFP, [66]). We carried out Western blot analysis using anti-GFP antibodies as a probe. Whereas two high molecular weight bands corresponding to LON-2::GFP and HS-modified LON-2::GFP are detected in wild-type lysates, only one of the bands is detected in lysates of rib-1m-/- z-/- and rib-2m-/- z-/- mutants (Fig 3C), indicating that HS synthesis onto LON-2/Glypican is affected by loss of rib-1 or rib-2 function. Consistent with this interpretation, wild-type worms expressing a mutant version of LON-2 in which the HS attachment sites are mutated (LON-2ΔGAG::GFP, [67]) displayed a single high molecular weight band that migrated to the same molecular weight as LON-2::GFP when expressed in rib-1(qm32)m-/- z-/- and rib-2(qm46) mutantsm-/- z-/- (Fig 3C). We next analyzed HSPG SDN-1/Syndecan using a similar strategy. We expressed GFP-tagged SDN-1/Syndecan (SDN-1::GFP, [14]) in wild-type and rib-1(qm32)m-/- z-/- mutant worms, and probed for GFP in lysates of these worms. In wild-type lysates, we detected two high molecular weight bands corresponding to SDN-1::GFP and HS-modified SDN-1::GFP, but only detected a single band in lysates of rib-1(qm32)m-/- z-/- mutants (Fig 3D), indicating that loss of rib-1 impairs HS synthesis onto SDN-1/Syndecan. Thus, our results provide compelling evidence that rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- mutations drastically reduce HS content, consistent with these mutations impairing HS chain elongation. As a note, using fluorescence microscopy, we observed that the expression of LON-2::GFP in rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- mutants was similar to wild type, and that SDN-1::GFP was similar to wild type in rib-1(qm32)m-/- z-/- mutants (S2 Fig).
rib-1(qm32) m-/- z-/- and rib-2(qm46) m-/- z-/- mutants are uncoordinated and egg-laying defective. To gain insight into the impact of HS chain elongation on neuronal development, we set out to characterize the neuroanatomy of rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/-mutants. For this, we built strains of the rib-1(qm32)m-/- z-/- and rib-2(qm46 m-/- z-/- mutants carrying integrated transgenes to drive the expression of fluorescent proteins and allow the visualization of specific neurons (see S9 Table). We examined the nervous system of rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- adult viable mutants and found that the overall organization of the nervous system is grossly normal, as neuronal ganglia, axon fascicles and isolated neurons were generally well laid out. Examination with single-cell resolution revealed that numerous neuronal migrations are affected in both mutants. For instance, the CAN neuron cell body, which migrates from the head region towards the midbody region in wild-type animals, is frequently positioned too anteriorly or too posteriorly in both mutants (Fig 4A). Also, the HSN neuron cell body, which migrates from the tail region to the midbody region in wild type, is often located too posteriorly in both rib-1(qm32) and rib-2(qm46) mutants (Fig 4B). Moreover, the AVM neuron cell body is frequently located in the posterior of the body instead of being anterior to the vulva (Fig 4C). The penetrance and expressivity of these defects is similar in both mutants. Thus, loss of function of the genes rib-1 or rib-2 impairs the guidance of diverse neurons that undergo long-range migrations during development.
We also found that axonal projections are defective in rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- single mutants. For example, the axon of the interneuron PVQ, which projects into the ipsilateral fascicle of the ventral nerve cord in the wild type, frequently projects in the contralateral fascicle or even laterally in rib-1(qm32) and rib-2(qm46) mutants (Fig 4D). Similarly, the axon of the motorneuron HSN, which projects ventrally and into the ipsilateral fascicle of the ventral nerve cord in the wild type, is misguided in the rib-1(qm32) and rib-2(qm46) mutants as it projects into the contralateral fascicle or laterally in these mutants (Fig 4B). These defects in HSN axon guidance are consistent with those reported for RNAi knockdown of rib-1 and rib-2, as well as in maternally rescued rib-1 animals [17, 32]. Another example is the axon of the mechanosensory neuron AVM, which extends ventrally towards the ventral nerve cord in the wild type, projects laterally in the mutants (Fig 4C). The axons of cholinergic and GABAergic motorneurons are also misguided in rib-1(qm32) and rib-2(qm46) mutants: contrary to the wild type, where most motorneuron axons exit the ventral midline on the right side to migrate along on the right side of the worm’s body wall, many motorneuron axons abnormally project to the left side in rib-1(qm32) and rib-2(qm46) mutants (Fig 5A). Finally, the dorsal nerve cord, which is composed of several motoraxons that run as a single fascicle in the wild type, is frequently defasciculated into several bundles in rib-1(qm32) and rib-2(qm46) mutants (Fig 5B). Thus, the guidance of numerous axons is disrupted upon loss of function of the genes rib-1 and rib-2.
In a similar way, the migration of mesodermal cells, which share guidance mechanisms with neurons [68], is defective in rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- single mutants. For instance, the canals of the excretory cell (two anterior and two posterior canals) run laterally in the wild type but are frequently too short or extend along the ventral or dorsal aspect of the body in rib-1(qm32) and rib-2(qm46) mutants (Fig 6A). Another example of misguided mesodermal cells in the rib-1(qm32) and rib-2(qm46) mutants is that of the distal tip cell (DTC), whose path determines the shape of the gonad. In wild-type animals, the anterior gonad arm is located on the right side of the animal, as the anterior DTC migrates along the right side, first anteriorly, then turning dorsally, and migrating back posteriorly towards the midbody region. Similarly, the posterior gonad arm is located on the left side, as the posterior DTC migrates posteriorly, turns dorsally and migrates back towards the midbody region. In rib-1(qm32) and rib-2(qm46) mutants, the sidedness of the gonad arms is often abnormal, with the anterior arm of the gonad on the left side of the animal and the posterior arm on the right side, or even having both gonad arms on the opposite side of the animal (Fig 6B). Lastly, rib-1(qm32) and rib-2(qm46) mutants display abnormal positioning of the excretory glands (Fig 6C). Thus, loss of function of the genes rib-1 or rib-2 disrupts the guidance of migrations of neuronal and mesodermal cells during development.
To gain insight into the roles of HSPGs during development, we determined the expression pattern of the HS copolymerase. For this, we designed a transcriptional fusion, Prib-1::gfp, between the upstream regulatory region of rib-1 and gfp. Since rib-1 is the second gene in a two-gene operon [69], we included the region upstream of the first gene in the operon, as well as the intergenic region of the operon that lies immediately upstream of rib-1 (see Materials and Methods). Second, we constructed a translational fusion for rib-1 using the same upstream regulatory region as for Prib-1::gfp, and fusing the coding region of rib-1 with venus, a gfp variant that fluoresces in acidic cellular environments [70], to make the translational fusion Prib-1::rib-1::venus. We generated at least five transgenic lines for each of these two rib-1 reporters and examined transgenic animals by fluorescence microscopy. We observed that the GFP signal from the transcriptional fusion Prib-1::gfp fills the cytoplasm of expressing cells, whereas the VENUS signal displays a punctate cytoplasmic pattern in cells expressing the translational fusion Prib-1::rib-1::venus, consistent with RIB-1 being localized to the Golgi apparatus (Fig 7A). Moreover, we found that both the transcriptional and translational fusions have a very similar spatial and temporal expression pattern during development: expression was visible in neurons, hypodermal cells, muscles of the digestive system, and reproductive tissues (Fig 7A). Importantly, we found that the translational fusion Prib-1::rib-1::venus is functional, as it fully rescues the defective locomotion, egg-laying, morphology, and axon guidance of rib-1(qm32)m-/- z-/- mutants. Our observations indicate that the observed expression pattern of the translational reporter and of the very similarly expressed transcriptional reporter are functionally relevant and largely reflect sites of endogenous rib-1 expression.
Since the transcriptional and translational rib-1 fusions have similar expression patterns, we used Prib-1::gfp, which has a stronger expression level, to characterize the expression pattern of rib-1 in more detail. We found that Prib-1::gfp is broadly expressed in ectodermal and mesodermal cells during embryogenesis. A salient feature of the rib-1 expression pattern is that it is very dynamic in hypodermal cells during development. In embryogenesis, Prib-1::gfp is detected along the entire layer of hypodermoblasts that surrounds the gastrulating embryo at about 200 minutes after fertilization. By the early comma stage of embryogenesis, Prib-1::gfp is expressed at high levels in hypodermal cells of the elongating embryo (Fig 7C), including hypodermal cells extending ventrally during ventral closure and in the two rows of dorsal hypodermal cells undergoing dorsal intercalation. Following these embryonic morphogenetic events, expression of Prib-1::gfp in the hypodermal cells of the body wall is no longer visible during larval and adult stages, except for seam cells undergoing fusion during larval development. Also, hypodermal cells of the developing vulva express Prib-1::gfp (Fig 7D), at a low expression level in L3 larvae and at a stronger level in L4 larvae and just molted young adults, and vanishing in vulval cells in the adult. These dynamic expression patterns in cells undergoing dramatic changes during morphogenesis suggest a potential for rapidly changing needs for particular HSPGs in specific tissues at different time points during development.
The nervous and digestive systems express Prib-1::gfp stably and continuously from embryogenesis throughout adulthood. Strong and sustained expression is seen in motorneurons, interneurons, sensory neurons (including AVM), neurons in the head and tail ganglia, with the GFP signal filling axons running along the ventral and dorsal nerve cords, commissures, and sublaterals. Expression in neurons of the ventral nerve cord and of the head ganglia is visible in 1.5-, 2-, and 3-fold embryos, and persists into adulthood (Fig 7A and 7B). Strong expression of Prib-1::gfp is also observed in the pharynx from the 2-fold stage of embryogenesis onwards and remained strong in adults (procorpus, metacorpus, terminal bulb, grinder, and pharyngeal-intestinal valve). The anal depressor, the anal sphincter, the two enteric muscles, the spermathecae and the uterine muscles maintain expression in adults (Fig 7B). The continued expression of rib-1 in the nervous, digestive and reproductive systems suggests that HSPGs play not only developmental, but also post-developmental roles in these cell types.
A prominent site of expression of the HS copolymerase is the nervous system, including during axon migration in embryonic and larval development (Fig 7), and disruption of the HS copolymerase in rib-1(qm32)m-/- z-/- or rib-2(qm46)m-/- z-/- mutants leads to numerous axon guidance defects (Fig 4 and Fig 5). To determine in which cells HS production is required for axon guidance, we provided rib-1(qm32)m-/- z-/- mutants with wild-type copies of rib-1(+) in subsets of cells and assessed rescue of the PVQ axon guidance defects. The axon of PVQ extends along the ventral nerve cord during embryogenesis, following the path of other axons, and is in proximity with the hypodermis and body wall muscles. We built constructs to express rib-1(+) in neurons including PVQ (using the heterologous promoter Prgef-1), in the hypodermis (using the heterologous promoter Pdpy-7), or in body wall muscles (using the heterologous promoter Pmyo-3). We then generated transgenic rib-1(qm32) worms expressing rib-1(+) in these tissues and analyzed PVQ axon guidance. As a control, we determined that expression of rib-1(+) under its own promoter completely rescued the guidance defects of the PVQ axon (Fig 8A). Targeted expression of rib-1(+) only in neurons, only in the hypodermis, or only in body wall muscles did not rescue the rib-1 mutant PVQ axon guidance defects. However, co-expression of rib-1(+) simultaneously in the hypodermis, neurons, and body wall muscles led to a significant rescue of these defects (Fig 8A), suggesting that HSPGs derived from specific cell types together contribute to proper PVQ axon guidance. The rescue of the PVQ axon was strong but incomplete, possibly due to the inappropriate rib-1 expression level or timing under these heterologous promoters. Nonetheless, expressing rib-1 simultaneously in these three tissues yielded a significant rescue of the PVQ defects, indicating a simultaneous functional requirement for HSPGs in distinct cell types to regulate the guidance of the PVQ axon.
We next turned to elucidating the spatial requirements of HS biosynthesis for guidance of the mechanosensory neuron AVM. During the first larval stage, the AVM axon pioneers its own ventral migration through a basement membrane along the body wall, sandwiched between the hypodermis and body wall muscles. We expressed rib-1(+) in the hypodermis (using the heterologous promoter Pdpy-7), in body wall muscles (using the heterologous promoter Pmyo-3), or in AVM itself (using the heterologous promoter Pmec-7) in rib-1(qm32)m-/- z-/- mutants, and analyzed AVM axon guidance. As a control, we determined that expression of rib-1(+) under its own promoter completely rescued the guidance defects of the AVM axon (Fig 8B). We found that the AVM axon guidance defects of rib-1 mutants were rescued by expression of rib-1(+) in AVM itself, or by expressing rib-1(+) in the hypodermis (Fig 8B). These results suggest that HSPGs derived from both AVM and the hypodermis crucially impact AVM axon guidance. Taken together, our results support the notion that HSPGs synthesized in distinct cell types coordinate guided axonal migration during development.
Prior studies have addressed roles of specific HSPG core proteins and HS chain modifications in Netrin- and Slit-mediated axon guidance in worms [8, 10–15, 18, 71], flies [72, 73], and mice [49]. HS chain elongation per se has been implicated in Netrin- and Slit-mediated axon guidance using in vitro spinal cord and retinal explant assays [74, 75]. However, an in vivo study of the impact of HS chain elongation in Netrin- and Slit-mediated guidance events has been lacking. To address how globally disrupting HS chain elongation affects guidance events that require the unc-6/Netrin or slt-1/Slit pathways, we studied the guidance of the AVM axon (Fig 9A). Two complementary and highly conserved guidance pathways guide the AVM axon: attraction mediated by the UNC-40/DCC receptor towards ventral UNC-6/Netrin, and repulsion mediated by the SAX-3/Robo receptor away from dorsal SLT-1/Slit (Fig 9B) [68, 76–81]. Simultaneous complete loss of both unc-6/Netrin and slt-1/Slit function leads to fully penetrant AVM axon guidance defects, where ~95% of AVM axons fail to extend ventrally, as demonstrated in [77] (also reproduced by [8]). We found that rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- mutants are defective in AVM ventral axon guidance (Fig 4C), a result that is consistent with the notion that loss of HS chain elongation may affect signaling through either unc-6/Netrin signaling, slt-1/Slit signaling, or both. It is also possible that an unidentified pathway, also involving HS chains, may help guide AVM ventrally.
To test whether HS elongation contributes to both Netrin and Slit signaling pathways, we constructed double mutants of rib-1(qm32) and rib-2(qm46) with mutations in unc-6/Netrin and slt-1/Slit. Because rib-1 and rib-2 null alleles are lethal, we used the hypomorphic alleles rib-1(qm32) and rib-2(qm46). For unc-6, we used the hypomorphic allele e78 [82], as we found that double mutants with the null allele unc-6(ev400) [83] rib-1(qm32);unc-6(ev400) and rib-2(qm46);unc-6(ev400) died as embryos. For slt-1, we used the presumptive null allele eh15 [77], as well as a condition of altered slt-1/Slit signaling, where misexpressing slt-1/Slit in all body wall muscles (using Pmyo-3::slt-1), leads to AVM axon guidance defects [84]. We found that all four double mutants (a) rib-1(qm32)m-/- z-/-;unc-6(e78), (b) rib-2(qm46)m-/- z-/-;unc-6(e78), (c) rib-1(qm32)m-/- z-/-;slt-1(eh15), and (d) rib-2(qm46)m-/- z-/-;slt-1(eh15), have AVM axon guidance defects that are more pronounced than the respective single mutants (Fig 9B). Similarly, animals expressing Pmyo-3::slt-1 in the rib-1(qm32)m-/- z-/- or rib-2(qm46)m-/- z-/- mutant backgrounds have severe AVM guidance defects compared to the rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- single mutants or to animals misexpressing Pmyo-3::slt-1 in the wild-type background (Fig 9B). While hypomorphic alleles complicate the interpretation of results, the finding that disruption of HS chain elongation alters axonal guidance in an additive manner to dysfunctional unc-6/Netrin and slt-1/Slit signaling, highlights the importance of HS chain elongation in axon guidance and is consistent with a role of HS chain elongation in unc-6/Netrin and slt-1/Slit signaling to guide AVM.
To directly test the functional importance of HS chain elongation in unc-6/Netrin signaling, we used a gain-of-function approach that specifically assays a unc-6/Netrin-dependent guidance event. Similar to AVM, the PVM axon is attracted ventrally towards secreted UNC-6/Netrin via the UNC-40/DCC receptor. However, misexpression of the repulsive UNC-6/Netrin receptor, UNC-5/UNC5, using the transgene Pmec-7::unc-5 ([85], Fig 9C) in PVM, results in an unc-6/Netrin- and unc-40/DCC-dependent abnormal extension of the PVM axon towards the dorsal side of the animal [86]. As controls, dorsal extension of the PVM axon is never observed in wild type or mutants in the unc-6/Netrin or slt-1/Slit signaling pathways (Fig 9C). We focused on the PVM axon in this assay as both the AVM and ALMR axons extend dorsally in Pmec-7::unc-5 transgenic animals rendering AVM indistinguishable from ALMR. We generated rib-1(qm32)m-/- z-/- or rib-2(qm46)m-/- z-/- mutant strains carrying the Pmec-7::unc-5 transgene [85] to misexpress unc-5 in PVM. If loss of rib-1 and rib-2 functionally disrupts unc-6/Netrin signaling, we would expect to see a decrease in the unc-5-mediated dorsal migration of PVM. Indeed, we found that rib-1(qm32) and rib-2(qm46) loss of function significantly suppressed the unc-6/Netrin-dependent unc-5-mediated dorsal migration of PVM (Fig 9C, S8 Table), indicating that unc-6/Netrin signaling requires HS chain elongation. This suppression of the dorsal extension of the PVM axon by mutations in rib-1 and rib-2 is specific, as individual loss of function of other genes required for guidance, such as sdn-1/Syndecan, slt-1/Slit and sax-3/Robo, did not suppress these abnormal dorsal extensions (Fig 9C). Taken together our observations support the notion that HS chain elongation plays a critical role in unc-6/Netrin-mediated guidance. Furthermore, if unc-6/Netrin and slt-1/Slit signaling pathways were indeed the sole two key pathways guiding the AVM axon ventrally, as the fully penetrant defects of unc-6 slt-1 double null mutants suggest [8, 77], then our results would support that HS chain elongation is important for both unc-6/Netrin and slt-1/Slit signaling. This notion is consistent with prior work implicating the HSPG syndecan in slt-1/Slit-mediated guidance [8, 14, 72, 73], the HSPG glypican in unc-6/Netrin-mediated guidance [8], and HS modifications in slt-1/Slit signaling [10, 11].
Once synthesized, HS chains become extensively modified by epimerases and sulfotransferases (reviewed in [1], Fig 3A). In C. elegans, key modifying enzymes have been studied, including glucuronyl C5-epimerase encoded by hse-5, 2O-sulfotransferase encoded by hst-2, and 6O-sulfotransferase encoded by hst-6 [9–11, 15, 87]. These HS modifying enzymes are required for axon guidance as mutations disrupting their function impair this process in a number of developmental contexts [1, 9–15, 21]. However, the roles of these HS modifying enzymes in the guidance of the AVM axon, which is mediated by unc-6/Netrin- and slt-1/Slit, are unknown. To determine the functional importance of specific HS modifications in AVM axon guidance, we first analyzed single, double, and triple null hse-5, hst-2 and hst-6 mutants, and found that loss of each single HS modifying enzyme led to minimal AVM axon guidance defects (Fig 9D), as has previously been reported [10]. However, hse-5; hst-6 and hst-2 hst-6 double mutants, in which the 6O-sulfotransferase and either the 2O-sulfotransferase or the C5-epimerase are mutant, display significant AVM guidance defects (Fig 9D). The defects of these two double mutants are not further enhanced by the loss of the third key HS modifying enzyme in hse-5; hst-2 hst-6 triple mutants (Fig 9D). These observations indicate some level of compensation between HS chain modifying enzymes, which has been observed at the biochemical level [87], and suggest that combinations of types of HS chain modifications impact the guidance of AVM, which relies on unc-6/Netrin- and slt-1/Slit-signaling. Our results add to prior work showing that specific HS chain modifications regulate precise cell and axon migration events in several other contexts, including of migration events that are slt-1/Slit- and unc-6/Netrin-dependent, and through interactions with the slt-1/Slit signaling pathway [10–15, 18].
Next, we analyzed AVM ventral axon guidance in double mutant animals lacking just one of the HS modifying enzymes, hse-5, hst-2, and hst-6, and unc-6/Netrin or slt-1/Slit. We found that loss of function of any of the HS modifying enzymes enhanced the AVM guidance defects of unc-6/Netrin null mutants. Similarly, loss of any of the HS modifying enzymes hse-5, hst-2, or hst-6 enhanced the AVM guidance defects of presumptive null mutants for slt-1/Slit [77] (Fig 9D). These results show that HS chain sulfations and epimerizations carried out by hse-5, hst-2, or hst-6 enzymes participate in AVM axon guidance likely through the two key signaling pathways unc-6/Netrin- and slt-1/Slit. These findings are in agreement with prior studies that demonstrated the importance HS chain modifications to neural development in other contexts [9–15, 18, 21].
Animal development and tissue homeostasis rely on the regulation of molecules that instruct cellular responses. HSPGs regulate morphogens and guidance cues in the extracellular environment, but their mechanisms are still not well understood, including how multiple HSPGs function together to coordinate cellular responses. Here, we identify viable hypomorphic mutations in the genes rib-1 and rib-2 encoding the HS copolymerase of C. elegans. These mutations severely reduce HS levels and disrupt morphogenesis and nervous system development. We show that the coordinated action of HSPGs from various tissues contributes to guide cellular migrations during development.
In this study we have molecularly identified and characterized two mutations that were previously isolated in a forward genetic screen for maternally rescued uncoordinated mutants [54–56]. We show that mum-1(qm32) and mum-3(qm46) are partial loss-of-function mutations in the genes rib-1 and rib-2, respectively, which encode the two exostosin glycosyltransferases that compose the HS copolymerase in C. elegans. The enzymatic activities predicted by sequence homology have been corroborated using bacterially expressed RIB-1 and RIB-2 [32, 34]. RIB-2 functions as an alpha1,4-N-acetylglucosaminyltransferase that has both GlcNAc transferase I and II activities and is involved in the addition of the first GlcNAc residue onto the tetrasaccharide linker, as well as in the elongation of HS chains [34, 57]. RIB-1 and RIB-2 glycosyltransferases together form a functional heterodimer that catalyzes HS chain elongation [32]. Moreover, HS levels have been shown to be reduced in maternally rescued worms of rib-1(tm516)m+/- z-/- and rib-2(qa4900)m+/- z-/- null mutations [32, 34]. Here we show that homozygous hypomorphic single mutants rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- have profoundly disrupted HS levels: we found that the global HS levels are severely reduced in these single mutants, and that high molecular species of LON-2/Glypican and SDN-1/Syndecan, likely corresponding to the core protein with HS chains attached, are undetectable in the rib-1(qm32) and rib-2(qm46) single mutants. These findings indicate that rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- result in a loss of function of the genes rib-1 and rib-2. It is noteworthy that rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- single mutants each display severe mutant phenotypes, indicating that rib-1 and rib-2 cannot substitute for each other, consistent with their specific biochemical roles in HS chain elongation. In sum, we provide further evidence that the function of RIB-1 and RIB-2 is required for HS biosynthesis in C. elegans.
Several observations indicate that rib-1(qm32) and rib-2(qm46) are partial loss-of-function mutations: (a) their phenotype is less severe than deletion alleles; (b) the phenotype of rib-1(qm32) and rib-2(qm46) over a deficiency is more severe that in homozygous mutants [54]; and (c) simultaneously disrupting both genes results in complete embryonic lethality. Thus, despite the severe reduction in HS in rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- single mutants, residual HS copolymerase activity appears to be sufficient for viability in the single mutants. However, that disrupting both genes in rib-1(qm32)m-/- z-/-; rib-2(qm46)m-/- z-/- leads to a more severe phenotype is likely because the HS copolymerase, a heterodimer of RIB-1 and RIB-2 proteins, is more drastically impaired in the double mutants. We propose that HS copolymerase dimers composed of one mutant protein and one wild-type protein might be stabilized by the presence of one normal protein in the complex in single rib-1(qm32) and rib-2(qm46) hypomorphic mutants (i.e. mutant RIB-1 and wild-type RIB-2 in rib-1(qm32), and mutant RIB-2 and wild-type RIB-1 in rib-2(qm46)).
In contrast to the single mutants rib-1(qm32)m+/- z-/- or rib-2(qm46)m+/- z-/-, which are completely maternally rescued [54], double mutant animals rib-1(qm32)m+/- z-/-;rib-2(qm46)m+/- z-/- are not, and instead become severely uncoordinated and egg-laying defective adults (Table 1). Thus, maternal product deposited in the oocyte is sufficient to support HS copolymerase activity and allow for normal development and behavior in single hypomorphic mutants rib-1(qm32)m+/- z-/- and rib-2(qm46)m+/- z-/-, but is insufficient for double hypomorphic mutants. Incomplete maternal rescue effect is observed for single null mutants rib-1(tm516)m+/- z-/- and rib-2(qa4900)m+/- z-/- [32, 34]. High levels of rib-1 and rib-2 transcripts are detected in the germline of C. elegans (http://nematode.lab.nig.ac.jp/). These observations highlight the importance of HS copolymerase activity, and therefore HSPGs, from the earliest stages of development.
RIB-1 and RIB-2 are not expected to affect the biosynthesis of glycosaminoglycans other than HS. In C. elegans, both HS and CS, but not hyaluronate nor dermatan sulfate, have been detected [88, 89]. Whereas HS and CS chains share the same tetrasaccharide linker to couple the HS or CS chain to their respective core proteins, the elongation of HS and CS chains are carried out by different enzymes. The elongation of CS chains, a polymer of alternating GlcA and N-acetylgalactosamine (GalNAc) residues, is catalyzed by a bifunctional glycosyltransferase encoded by the sqv-5 gene [90]. Thus, rib-1(qm32) and rib-2(qm46) mutations facilitate the study of the consequences of globally and specifically disrupting HS elongation in live animals.
Complete disruption of HS chain elongation in the deletion alleles of rib-1(tm516)m-/- z-/- and rib-2(qa4900)m-/- z-/- affects the mutant organism in a pleiotropic fashion, leading to fully penetrant embryonic lethality [32, 34], which had limited the systematic study of the impact of HS chain elongation in later developmental processes. We found that a basal level of the required enzymatic activities in the hypomorphic mutants rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- is sufficient to bypass major pleiotropic effects, allowing a proportion of the animals to fully develop and reach adulthood [54]. In these animals that complete development, major morphogenetic movements, such as gastrulation, ventral closure and organogenesis, occurred normally, and their anatomy, including the specification of neuronal identities and the layout of ganglia and major axon fascicles, was grossly normal. This hypomorphic condition allowed us to study the influence of HS chain elongation on the guidance of cell and axon migration in viable animals. We found that disrupting HS chain elongation affects the migrations of neurons and axons, including migrations that occur during embryonic and post-embryonic development, and along both body axes (antero-posterior and dorso-ventral) [78]. It is worth noting that the motility per se of migrating cells is not lost in rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- mutants as soma and axons often overshoot their targets. Rather, it is the guidance of migrations during development that is disrupted by the loss of function of rib-1 and rib-2. The critical role of HS elongation in axon guidance during nervous system development is evolutionarily conserved, as disruption of HS elongation in mice and fish leads to defective axonal guidance [40, 49].
Previous analyses of the consequences of disrupting HS chain elongation on neuronal development reported weaker defects than those we describe here using rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/-. This is not surprising since only maternally rescued animals rib-1(tm516)m+/- z-/- and rib-2(tm710)m+/- z-/-, or partial knock down of gene activities with rib-1(RNAi) and rib-2(RNAi), could be studied before. For example, HSN soma migration is fully normal in maternally rescued rib-2(tm710)m+/- z-/- [21] and 27% of maternally rescued rib-1(tm516)m+/- z-/- animals exhibit HSN axon guidance defects [32]. Similarly, rib-1(RNAi) and rib-2(RNAi) lead to a 40–45% penetrance of combined HSN cell and axon guidance events [17]. In contrast, 77% and 74% of rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- mutant animals exhibit abnormal HSN soma migration, and 100% and 83% of rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- mutant animals display defects in the guidance of the HSN axon, respectively. This highlights that the newly identified hypomorphic rib-1(qm32) and rib-2(qm46) mutants reported here enable the study of HS chain elongation-dependent biological processes.
The guidance of multiple migrating neurons and axons is altered in rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- mutants, suggesting that disruption of HS chain elongation impacts several guidance pathways. In particular, our analysis of the ventral axon guidance of AVM shows that HS chain elongation and HS chain modifications, are important for signaling via the unc-6/Netrin and the slt-1/Slit signaling pathways, and perhaps a yet to be identified HS-dependent pathway. That loss of function of rib-1 or rib-2 was able to suppress the unc-6/Netrin-dependent dorsalization of PVM upon ectopic expression of unc-5/UNC5 supports the model that unc-6/Netrin signaling requires HS chains. We have previously shown that HSPG lon-2/Glypican functions with, and is required for, unc-6/Netrin signaling in axon guidance [8]. Interestingly, the core protein of LON-2/glypican, but not its HS chains, functions in the unc-6/Netrin pathway. In fact, two versions of LON-2/Glypican lacking the HS chains are able to function in unc-6/Netrin-mediated axon guidance (one where the HS attachment sites were deleted [67], which indeed prevents the addition of HS onto LON-2/Glypican (Fig 3C), and one where LON-2/Glypican is truncated in a way that removes all HS attachment sites, [8]). Taken together, the observation that the core protein of LON-2/Glypican, but not its HS chains, functions in unc-6/Netrin signaling, and that HS elongation is required for unc-6/Netrin signaling (loss of rib-1 or rib-2 suppresses the unc-6/Netrin-dependent effect of unc-5/UNC5 ectopic expression), raises the possibility that an additional unidentified HSPG functions in unc-6/Netrin signaling. One HSPG that has a role in unc-6/Netrin guidance in other contexts is unc-52/Perlecan, which affects the guidance timing defects of distal tip cells upon ectopic early expression of unc-5/UNC5 [16]. However, loss of unc-52 alone does not affect AVM axon guidance and does not enhance defects of sdn-1/Syndecan mutants [8], indicating that unc-52 likely does not participate in AVM guidance. Thus, another unidentified HSPG may function in unc-6/Netrin signaling through its HS chains.
Also, the notion that HS chains are important for slt-1/Slit signaling is consistent with prior reports that (1) HS chain elongation is required for retinal explant axon outgrowth in vitro [74], (2) the HSPG gene sdn-1/Syndecan functions in slt-1/Slit signaling to guide the AVM axon [8] (Fig 10B) and other axons (PVQ, [14]), and (3) HSPG Syndecan is key to Slit signaling and distribution in flies [72, 73]. Furthermore, studies have demonstrated roles for HS chains in slt-1/Slit-mediated guidance in other contexts [10, 11], through the study of HS modifying enzymes mutants, however whether it is the modifications of the HS chains on SDN-1/Syndecan specifically that are required for guidance is not known.
We analyzed the spatial requirements for the HS copolymerase by focusing on two specific migrating neurons, namely the embryonically migrating PVQ axon and the AVM axon that extends during the first larval stage. In both cases we found that rib-1 expression in several cell types restored function during development of rib-1 mutants. For the guidance of the migrating PVQ axon, combined HS copolymerase expression in the hypodermis, neurons, and body wall muscles of rib-1 mutants was required to rescue PVQ axon guidance defects. This observation indicates that HS chains synthesized onto HSPGs from multiple tissue types cooperate to properly pattern the ventral midline and suggests that distinct HSPGs from specific tissues may contribute to properly guide the PVQ axons at the ventral midline (Fig 10A). Interestingly, the combined loss of two HSPGs, a glypican and a syndecan, in the lon-2 sdn-1 double mutant leads to a penetrance of defects in PVQ guidance similar to rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- mutants [12], suggesting that PVQ axon guidance defects observed in rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- mutants may reflect a disruption of HS chains onto LON-2/Glypican in the hypodermis and SDN-1/Syndecan in the PVQ neurons. Consistent with this interpretation, lon-2/Glypican has been found to function non-cell autonomously in the hypodermis to guide the axon of AVM [8], and sdn-1/Syndecan has been shown to function cell autonomously in the migrating neuron in a variety of contexts, such as AVM axon guidance [8], PVQ axons, HSN soma, and ALM soma [14]. Furthermore, PVQ axon guidance likely requires that the specific HS chains on core HSPGs not only be synthesized but also modified, as the combined loss of HS modifying enzymes also leads to PVQ axon guidance defects [10]. Indeed, loss of the C5-epimerase hse-5 and the 6-O-sulfotransferase hst-6 in hse-5; hst-6 double mutants, or loss of the 2-O-sulfotransferase hst-2 in double mutants hst-2 hst-6 leads to PVQ axon guidance defects [10] with a similar penetrance to that of rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- mutants. Together, our studies within the context of the literature suggest that the coordinated action of specific HS chains synthesized and modified in different tissues onto distinct HSPG core proteins function to properly guide the PVQ axons at the ventral midline.
Similarly, the defective guidance of the axon of the AVM neuron is strongly rescued by expression of the HS copolymerase in the AVM neuron itself, but expression in the underlying hypodermis also contributes to normal AVM development. In this case too, HS chains synthesized onto core HSPGs functioning to guide AVM may be sdn-1/Syndecan in AVM and lon-2/Glypican in the hypodermis, as previously identified by analysis of core protein mutants in the context of AVM axon guidance [8] (Fig 10B).
Given that HSPGs decorate most cells in metazoans and are implicated in numerous cellular processes at the cell surface, including cell-matrix, cell-cell, and ligand-receptor interactions during development and tissue homeostasis, it could be expected that the HS copolymerase may be expressed ubiquitously. Functional rib-1::gfp was detected in virtually all cell types at some point during development. Interestingly, the HS copolymerase expression pattern was found to be dynamic, with levels changing and differing across tissues and developmental stages. Expression is strong and transient in hypodermal cells of the embryo, the larva, and the developing vulva, likely reflecting the developmental requirements for cell migration during the formation of complex tissue shapes during morphogenesis. Our observations support a model in which high expression of the HS copolymerase in cells that secrete a basement membrane, such as hypodermal cells, is pivotal for cell migration along this basement membrane. In addition, it is quite likely that migrating cells themselves might dynamically regulate surface HSPGs to modulate their adhesion properties and regulate the signaling of guidance cues. This is reflected by HS copolymerase expression peaking during periods of cell migration during embryogenesis and vulva formation.
In addition to this dynamic expression pattern, the HS copolymerase shows sustained expression in a number of structures throughout the life of the animal, including the pharynx, the pharyngeal-intestinal valve, the anal depressor, sphincter, and enteric muscles, as well as the nervous system. These are morphologically complex cells that are under continuous mechanical stress. For instance, the pharynx is constantly pumping bacteria, thus exerting variable pressure on the pharynx itself and the pharyngeal-intestinal valve. Similarly, the enteric muscles, the anal depressor, and the sphincter, all contract to expel waste, and, as the animal moves, the relatively long neuronal axons within the nerve cords are constantly subjected to stretch and relaxation. Other cell types expressing the HS copolymerase are the spermatheca, which stretches to welcome oocytes to be fertilized and contracts to expel the zygotes, and the uterine muscles, which contract to lay embryos in reproducing adults. Our observations point to a role for HSPGs in maintaining the integrity of tissues, possibly by regulating the attachment of cells that undergo considerable mechanical stress from repeated body contractions, and thus contribute to tissue homeostasis. That rib-1 expression persists post-developmentally is consistent with studies in other model systems that describe post-developmental roles for HS and HSPGs [4, 50].
The rib-1 expression pattern overlaps with known expression patterns of specific HSPG core proteins. For example, membrane bound SDN-1/syndecan is expressed in neurons, hypodermis, and pharynx [14], GPI-linked LON-2/glypican shows expression in the intestine and hypodermis [66], and UNC-52/perlecan, a secreted HSPG, is expressed in body wall muscles, digestive system muscles, and pharynx [91, 92]. Overlap between expression of HS biosynthetic machinery, such as RIB-1, and the localization of specific HSPGs, suggests that HSPGs may remain near cells where HS synthesis occurs. This may have functional relevance, as glypicans in fibroblast cells were shown to be internalized through endocytosis, returned to the Golgi, and then transported back to the membrane with HS chains altered both in length and modification pattern [93, 94]. Whether an internal recycling of HSPGs back to the HS biosynthetic machinery in the Golgi occurs in C. elegans, or whether it has functional relevance to guidance, remains to be determined.
In conclusion, our studies have identified viable mutations in each of the two subunits of the HS copolymerase in C. elegans, which severely disrupt HS biosynthesis, leading to profound developmental defects. Our findings offer a model system to dissect the functions of HSPGs in C. elegans and uncover general principles of their roles during development and tissue homeostasis.
Nematode cultures were maintained at 20°C on NGM plates seeded with OP50 bacteria as described [51]. mum-1/rib-1(qm32) and mum-3/rib-2(qm46) alleles were outcrossed five times before building strains with reporters. 14% of mum-1/rib-1(qm32)m-/- z-/- animals reach adulthood, as 68% of the embryos hatch into larvae and 20% of larvae reach adulthood; and 63% of mum-3/rib-2(qm46)m-/- z-/- animals reach adulthood, as 85% of the embryos hatch into larvae, and 74% of larvae reach adulthood [54]. Alleles used in this study are listed in S1 Table. Strains were constructed using standard genetic procedures and are listed in S9 Table. When needed, genotypes were confirmed by genotyping PCR or sequencing, using primers listed in S10 Table.
Neuroanatomy was examined in animals of rib-1(qm32)m-/- z-/- and rib-2(qm46)m-/- z-/- that are had completed development to L4 larvae and adults (14% and 63% of the respective populations) using specific reporters. Animals were mounted on agarose pads, anaesthetized with 100 mM sodium azide, and examined under a Zeiss Axio Scope.A1 or a Zeiss Axioskop 2 Plus.
For mapping mum-1, a three-point mapping experiment was carried out by picking Unc-non-Dpy and Dpy-non-Unc recombinants from heterozygous mothers of the genotype mum-1/unc-24 dpy-20, and the presence of mum-1 was assessed among the progeny of the homozygosed recombinants. Two-point mapping was carried out by picking Dpy worms from mum-1 dpy-20/++ heterozygous mothers, and the presence of mum-1 was assessed in the next generation. Also, Lin-non-Dpy recombinants were picked from heterozygous mum-1/lin-3 dpy-20.
As the rib-1 and rib-2 mutants are severely morphologically abnormal, cosmids, constructs, and PCR products were injected into strains carrying the mum-1/rib-1(qm32) or mum-3/rib-2(qm46) mutations in a heterozygous state, balanced by flanking visible markers. For rib-1, we used rib-1(qm32)/unc-24(e138) dpy-20(e1282ts) and for rib-2, we used rib-2(qm46)/unc-32(e189) dpy-18(e364) (see S9 Table). Transgenic F1s were isolated and lines homozygous for rib-1 or rib-2 were established.
Transgenic animals were generated by standard microinjection techniques [95]. Each construct or PCR amplicon was injected at 5 to 25 ng/μl with one or two coinjection markers which included pRF4-rol-6(su1006d) (100–150 ng/μL), Pttx-3::mCherry (50 ng/μL), Pceh-22::gfp (50 ng/μL), pCB101.1 Prgef-1::DsRed2 (50 ng/μL), and Punc-122::rfp (50 ng/μL). pBSK+ (90–100 ng/μL) used to increase total DNA concentration if needed. For coinjection markers used for each rescued transgenic line, see S9 Table.
The gene rib-1/F12F6.3 is downstream of the gene srgp-1/F12F6.5 in an operon of two genes. The nearest gene upstream of the operon is transcribed in the opposite direction. The genomic region between the operon of srgp-1 and rib-1, and the upstream neighboring gene is 4352 bp, corresponding to coordinates 22290–26642 on cosmid F12F6.
Prib-1::rib-1 (PCR product): A PCR product containing bases 34593–39595 of cosmid F12F6 of the rib-1 locus was amplified with Pfu polymerase.
Prib-2::rib-2 (PCR product): A PCR product containing bases 581 to 6196 of cosmid K01G5 of the rib-2 locus was amplified with Phusion polymerase.
Prib-1::gfp (pCB78): A PCR generated piece containing bases 23701 to 26662 of cosmid F12F6 corresponding to the promoter region of the rib-1 operon, as well as the initial 7 codons of srgp-1, was cloned upstream of gfp in the pPD95.77 vector using enzymes PstI and XbaI.
Prib-1::rib-1::Venus (pCB221): The rib-1 promoter region containing bases 23701–26580 of cosmid F12F6 was PCR amplified and cloned upstream of gfp in the pPD95.77 vector using enzymes SphI and PstI. A PCR generated piece containing bases 34452–39527 of cosmid F12F6 corresponding to the intergenic sequence between the genes rib-1 and srgp-1, as well as the rib-1 coding sequence, was cloned downstream of the rib-1 promoter and upstream of gfp using enzymes PstI and AvrII. Then, gfp was replaced with a PCR amplified Venus and cloned in frame with rib-1 using enzymes MscI and ApaI.
As a note, for rib-2, we constructed several transcriptional Prib-2::gfp and translational Prib-2::RIB-2::Venus reporters with different sizes of promoter region, injected at a range of concentrations (10–150 ng/μL). At least three transgenic lines were examined for each condition, but gave no or a very weak expression level in transgenic worms carrying these constructs. A very faint level of Prib-2::gfp was broadly detected in comma-stage embryos, and in the head and vulva area at later developmental stages.
Pdpy-7::rib-1 cDNA (pCB186): The rib-1 cDNA was amplified from yk1228g12 and ligated into a Pdpy-7 vector with a pPD95.75 backbone using enzymes XmaI–NcoI.
Pmyo-3::rib-1 cDNA (pCB196): A Pmyo-3 HindIII–XbaI fragment was ligated upstream of the rib-1 cDNA in a pCB186 HindIII–XbaI fragment in place of Pdpy-7.
Pmec-7::rib-1 cDNA (pCB204): The rib-1 cDNA was amplified from yk1228g12 and cloned into the pPD96.41 vector using enzymes AgeI–BglII.
Prib-1::rib-1 cDNA (pCB225): The rib-1 cDNA was ligated downstream of the rib-1 promoter (bases 23,701 to 26,580 of cosmid F12F6) using enzymes XmaI–ApaI in the pPD95.77 backbone.
Prgef-1::rib-1 cDNA (pCB199): The rib-1 cDNA was ligated downstream of Prgef-1 in place of DsRed2 using enzymes XmaI–ApaI in the pCB101.1 vector.
All inserts of finalized clones were verified by sequencing.
The genomic regions of mum-1/rib-1 and mum-3/rib-2 were PCR amplified using Pfu polymerase and sequenced on two independent PCR products amplified from genomic DNA of qm32 and qm46, respectively, using primers to cover the entire genomic region. Primers listed in S10 Table sequence over the mutation in each of the two mutants.
Mixed-stage wild type (N2), SDN-1::GFP (opIs171), rib-1; SDN::GFP (rib-1; opIs171) and rib-1 GFP control (rib-1; lqIs4) worms were collected from plates devoid of bacteria in buffer and protease inhibitors (Roche). Worm pellets were subjected to repeated freeze-thaw cycles. Protein concentration was measured using the Pierce 660 nm Protein Assay on a Nanodrop. 80 μg of samples mixed with 2x Laemmli sample buffer (Bio-Rad) were frozen in liquid nitrogen, then boiled, separated by SDS-PAGE on a 4–20% Mini-Protean TGX gel (Bio-Rad), and transferred to PVDF membrane. Membranes were incubated in 1:3000 rabbit anti-GFP primary antibody (Millipore #AB3080) and 1:9000 goat anti-rabbit HRP secondary antibody (Bio-Rad #166-2408EDU). For the loading control, membranes were incubated in 1:5000 rabbit anti-HSP90 antibody (CST #4874) and 1:10000 goat anti-rabbit HRP secondary antibody (Bio-Rad #166-2408EDU). Signal was revealed using Clarity Western ECL Substrate (Bio-Rad), and imaged using film (LabScientific).
Mixed-stage wild type (N2), GFP control (lqIs4), LON-2::GFP (TLG257), LON-2ΔGAG::GFP (TLG199), rib-1; LON-2::GFP (VQ525), rib-2; LON-2::GFP (VQ528), rib-1 GFP control (rib-1; lqIs4) and rib-2 GFP control (rib-2; lqIs4) worms were collected from plates devoid of bacteria in buffer and protease inhibitors (Roche), mixed with 2x Laemmli sample buffer (Bio-Rad), and frozen in liquid nitrogen. Samples were boiled and spun down, separated by SDS-PAGE on a 4–20% Mini-Protean TGX gel (Bio-Rad), and transferred to PVDF membrane. Membranes were incubated in 1:3000 rabbit anti-GFP primary antibody (Millipore #AB3080) and 1:9000 goat anti-rabbit HRP secondary antibody (Bio-Rad #166-2408EDU). For the loading control, membranes were incubated in 1:5000 rabbit anti-HSP90 antibody (CST #4874) and 1:10000 goat anti-rabbit HRP secondary antibody (Bio-Rad #166-2408EDU). Signal was revealed using Clarity Western ECL Substrate (Bio-Rad), and imaged using film (LabScientific).
Worm RNA was extracted using Trizol (Invitrogen) according to manufacturer’s instructions. RNA (500 ng) was reverse transcribed using the High Capacity cDNA Reverse Transcription Kit (Applied Biosystems) and random primers. PCR reactions were carried out with cDNA template, and 0.25 μM of each primer in 10 mM Tris pH 8.3, 1.5 mM MgCl2, 50 mM KCl, 0.2 mM deoxynucleotides, and 1 U Phusion DNA polymerase for 30 cycles of 94°C for 10 seconds, 55°C for 20 seconds, and 72°C for 45 seconds. Primers used to detect rib-1 transcript: oCB1533 (TGGAATCGACACAACGGATCG), oCB1534 (CAAGCAGTTCGTCGTATTCCC), oCB1535 (GAATACGACGAACTGCTTGCC), oCB1536 (TCCAGCTCAATCTTGTTGTCG) and oCB1537 (AGATGTGATGAGGGGAGAACG). Primers used to detect rib-2 transcript: oCB1538 (CAGTTCGTTTGGAATTGACGG), oCB1539 (CTGCTATATGATTGACATCCACAGG), oCB1540 (CACGTCATCACGCCAGATACG), and oCB1541 (TGATTCTGTGGGAGACGCGTC). The transcript for Y45F10D.4 was used as control using the primers oCB992 (TCGCTTCAAATCAGTTCAGC) and oCB993 (GCGAGCATTGAACAGTGAAG).
|
10.1371/journal.ppat.1000377 | Elite Suppressor–Derived HIV-1 Envelope Glycoproteins Exhibit Reduced Entry Efficiency and Kinetics | Elite suppressors (ES) are a rare subset of HIV-1–infected individuals who are able to maintain HIV-1 viral loads below the limit of detection by ultra-sensitive clinical assays in the absence of antiretroviral therapy. Mechanism(s) responsible for this elite control are poorly understood but likely involve both host and viral factors. This study assesses ES plasma-derived envelope glycoprotein (env) fitness as a function of entry efficiency as a possible contributor to viral suppression. Fitness of virus entry was first evaluated using a novel inducible cell line with controlled surface expression levels of CD4 (receptor) and CCR5 (co-receptor). In the context of physiologic CCR5 and CD4 surface densities, ES envs exhibited significantly decreased entry efficiency relative to chronically infected viremic progressors. ES envs also demonstrated slow entry kinetics indicating the presence of virus with reduced entry fitness. Overall, ES env clones were less efficient at mediating entry than chronic progressor envs. Interestingly, acute infection envs exhibited an intermediate phenotypic pattern not distinctly different from ES or chronic progressor envs. These results imply that lower env fitness may be established early and may directly contribute to viral suppression in ES individuals.
| The majority of HIV-1–infected individuals experience high plasma viral loads and CD4+ T cells loss in the absence of antiretroviral therapy. However, a very rare and important subset of individuals termed elite suppressors is able to maintain HIV-1 plasma viral loads below the limit of viral detection in the absence of treatment. The reasons behind this ability to control the virus are poorly understood, but they likely involve both an effective host immune response against HIV-1 and factors related to the virus itself. Here, we analyze the function of the HIV-1 coat protein or envelope glycoprotein from a group of elite suppressors. HIV-1 envelope mediates entry into the host cell via interaction with the cellular receptors CD4 and CCR5. Envelopes from elite controllers interacted with these receptors inefficiently compared to those from individuals with detectable viral loads. These inefficient interactions by elite suppressor envelopes led to slow rates of entry into host cells. Envelopes from acutely infected individuals were not significantly different from elite suppressors or chronically infected individuals. These findings suggest that the decreased envelope efficiency may contribute to viral control in elite suppressors.
| A minor subset of HIV-1–infected individuals maintains stable CD4+ T cell counts in the absence of antiretroviral therapy. A small proportion of these long-term nonprogressors (LTNPs), termed elite suppressors (ES), control plasma viral loads to <50 copies/ml [1]. Mechanism(s) responsible for this elite control are poorly understood but likely involve host and viral factors. Studies have explored the contributions of the innate and adaptive immune responses, host genetic polymorphisms, and viral dynamics (reviewed in [2]). For example, the major histocompatibility complex class (MHC) I group B alleles HLA-B27, -B51, and –B57 have been strongly associated with slower rates of HIV-1-associated disease progression [3]–[6]. Although these HLA-B alleles are overrepresented in ES and LTNPs, they are only expressed in a subset of these individuals indicating that the presence of these alleles is not necessary to suppress viremia and that other factors are likely involved [4],[7].
Although much previous work on ES has focused on host factors, less is known about viral fitness in these individuals. The impact of viral attenuation on disease progression was first described in a cohort of LTNPs infected by a common donor with virus containing a deletion in the nef gene [8],[9]. Investigation of other LTNP cohorts has shown both the presence [10],[11] and absence [12],[13] of defective nef genes. In other cohorts, the presence of viruses with reduced replication capacity has been associated with slower disease progression [14]–[19]. This viral attenuation could be the result of divergent evolution as a result of direct selective pressure by the host immune response [16]–[19]. However, recent work has shown that replication-competent viruses can be recovered from ES individuals indicating that ES harbor functional virus [20]. Furthermore, large scale sequencing of ES viruses yielded no identifiable common genetic defects [21]. Investigating the relative fitness of viral quasispecies in ES will help determine whether viral fitness is influencing disease outcome in these individuals.
Low HIV-1 genetic diversity in ES may be indicative of the presence of lower fitness variants [22]. Sequence analysis of functional envelope glycoprotein ES clones showed significantly decreased env diversity compared to individuals with chronic viremia suggesting that viruses in these patients experience minimal viral replication and diversification [23]. Lack of env diversification suggests that ES envs may be closely related in genotype and phenotype to the founder virus establishing infection.
In this study we have performed rigorous phenotypic analysis on subtype B env clones from ES plasma virus to determine whether env fitness may be contributing to viral suppression in ES. A novel cell line was utilized to show that ES env clones exhibit low CD4 receptor and CCR5 co-receptor usage and slow fusion kinetics compared to chronic infection envs. Analysis of control viruses indicated that these characteristics directly correlated to reduced replication capacity in vitro. Acute infections envs were intermediate in their entry efficiency and not significantly different from either chronic or ES envs. This study provides the first direct evidence that decreased env function is a property of ES and that this may contribute to viral suppression.
This study was conducted according to the principles expressed in the Declaration of Helsinki. The study was approved by the Institutional Review Board of Johns Hopkins School of Medicine and Rockefeller University hospitals. All patients provided written informed consent for the collection of samples and subsequent analysis.
The elite suppressor and chronic progressor patients have been previously described [23] (Table 1). Patients identified with acute/early HIV-1 infection have been previously described [24]. The estimated duration of infection was calculated 2 weeks prior to the onset of acute retroviral illness unless the patient could identify a precise high risk event. Table 1 contains relevant enrollment data for all acute/early infection patients. Elite suppressors were defined as individuals who maintained viral load to below 50 copies of RNA/ml plasma in the absence of retroviral therapy yet were Western blot positive for infection. Informed consent was obtained prior to phlebotomy. The protocol for ES/CP or acute/early infection was approved by an institutional review board of the Johns Hopkins University School of Medicine and the Aaron Diamond AIDS Research Center, respectively.
Envelope expression vectors were generated as previously described [23]. Envelope pseudotypes were generated by cotransfection of 293T cells with the 1 µg of the luciferase-encoding pseudotyping vector pNLLuc.AM and 1 µg of envelope expression vector. Cells were washed after 24 h, and pseudoviruses were collected after a subsequent 48 h. Relative particle numbers were determined by limiting dilution reverse transcriptase assay. Viruses were characterized as exclusively CCR5-utilizing by comparison of infectivity of U87-CD4/CCR5 and U87-CD4/CXCR4 cells, as previously described [25]. Affinofile cells were generated by selection of 4 vector stable cells (Johnston et al., submitted). CCR5 expression is controlled by a two vector ecdysone-inducible promoter. pVgRXR encodes the VgEcR fusion protein under control of the CMV promoter, and the RXR open reading frame under control of the RSV 5′ long terminal repeat. pIND-CCR5 encodes CCR5 under control of the minimal heat shock promoter with inducible control provided by five repeats of the glucocorticoid receptor DNA binding domains (5×E/GRE). Addition of the ecdysone derivative ponasterone A (the inducer) results in recruitment of a transcriptional coactivator to the 5×E/GRE element and activation of transcription of the CCR5 ORF. CD4 expression is inducibly regulated by the TREx expression system (Invitrogen). Cells contain pcDNA5-TO-CD4 and transcription of the CD4 ORF is controlled by the addition of the tetracycline analog minocycline. Single cell clones were isolated to generate cell populations with consistent levels of induction upon stimulation of CD4 and CCR5 expression.
Affinofile cells were plated at a density of 10,000 cells per well in a 96-well plate and allowed to adhere for 48 hours. Cells were induced in a matrix pattern to express CD4 and CCR5. Minocycline was added to cells in 2-fold dilutions over 6 separate dilutions (5 ng/ml–0 ng/ml) to induce CD4 expression. Ponasterone A was added in 2-fold dilutions over 6 separate dilutions from a final concentration of 4 µM to 0 µM to induce CCR5 expression. This matrix results in 36 unique CD4 and CCR5 induction surface concentrations. Each drug concentration was induced in triplicate. Cells were induced for 24 hours prior to infection. Cells were then exposed to pseudovirus for 48 h, washed with PBS, and lysed with Glo lysis buffer (Promega, Inc.). Maximal infection was considered luciferase activity generated by infection at the highest CD4 and highest CCR5 concentration. To control for effects caused directly by minocycline and/or ponasterone A on viral infectivity, U87-CD4/CCR5 cells were treated with a similar matrix of both drugs. CCR5 and CD4 expression levels were unchanged by flow cytometry, and no changes in infectivity of Yu-2 and SF162 envelope pseudoviruses were noted, thus variation in infectivity was assumed to be due to variations in receptor expression levels (Johnston et al, submitted).
For the kinetic fusion assay, HIV-1 pseudoviruses bearing either ES or chronic envelopes were spinonculated onto U87-CD4/CCR5 cells. 2.5×106 cells were spin-infected with pseudovirus-containing supernatant for 90 min at 1,200×g at 4°C. The cells were washed twice with cold phosphate buffered saline (PBS) to remove unbound virions. Cells were resuspended in cold medium and split into 96-well plates (50 µl/well). Virus-cell mixes were synchronized for entry by addition of 130 µl of 37°C medium, and then ENF at 10 µM was added to each well in 20 µl of medium at fixed-time intervals after addition of warm medium, which is defined as tim = 0 for synchronization of viral replication. Cells were incubated for 48 h and then treated with lysis buffer and luciferase activity was determined. For the kinetic fusion assay 6 hours was used as 100% or maximal luciferase activity. For the reverse transcription assay, Affinofile cells were induced with 5 ng/ml minocycline 24 hours prior to infection. ES or chronic pseudoviruses were synchronously added to cells. Efavirenz (EFV) was added at a concentration of 1 µM to each well at fixed time intervals after the addition of virus. For reverse transcription assay, 12 hours was used as 100% or maximal luciferase activity.
Affinofile cells were induced 24 hours prior to infection with 5 ng/ml minocycline. Cells were incubated with serial 10-fold dilutions of either chemokine (CCL5 [50 nM to 0.1 nM]) or drug (ENF (T-20) [1 µM to 0.1 nM], TAK-779 [1 µM to 0.1 nM]) for 1 h prior to the addition of virus. Cells were incubated for 48 h, washed with PBS, lysed, and luciferase activity determined. Plots of luciferase activity versus drug concentration were used to determine IC50 values for each pseudovirus. Luciferase activity without drug was used as maximal or 100% infection value.
Peripheral blood mononuclear cells (PBMCs) were isolated from heparin-treated venous whole blood from HIV–seronegative donors by Ficoll-paque density gradient centrifugation (GE Healthcare, Piscataway, NJ). Isolated PBMCs were washed twice in wash buffer [phosphate-buffered saline (PBS) supplemented with 2% fetal bovine serum (FBS), 0.1% glucose, 12 mM HEPES, and penicillin (100 U/m) and streptomycin (100 µg/ml)] and activated in RPMI 1640 (Mediatech, Inc., Manassas, VA) supplemented with 10% FBS, penicillin/streptomycin, and 2 µg/ml phytohemmaglutinin (PHA, Sigma-Aldrich, St. Louis, MO) and 100 U/ml interleukin 2 (IL-2, Invitrogen, Carlsbad, CA) for 3 days at 37°C and 5% CO2. Total PBMCs were subsequently maintained in RPMI 1640 supplemented with 10% FBS, pen/strep, and 100 U/ml IL-2. For flow cytometry experiments, a total of 10° PBMCs were collected 4 days post-stimulation by centrifugation at 2500 rpm×10 minutes and washed once in FACS staining buffer (PBS with 2% FBS, 0.5% bovine serum albumin, and 0.02 sodium azide). The cells were incubated in either an anti-CD4 antibody [Fluorescein isothiocyanate(FITC)-conjugated anti-CD4, BD Biosciences Pharmingen, San Jose, CA] or FITC-conjugated IgG1 isotype control (BD Biosciences Pharmingen), or an anti-CCR5 antibody [Phycoerythrin(PE)-conjugated anti-CCR5 clone CTC5, R&D Systems, Minneapolis, MN] or PE-conjugated IgG2B isotype control (R&D Systems). All antibodies were incubated at a final concentration of 12.8 µg/ml for 30 minutes at room temperature in FACS staining buffer. Stained PBMCs were washed again in FACS staining buffer and samples were analyzed on a FACSCalibur flow cytometer (Becton Dickinson, Franklin Lakes, NJ) with Cellquest software. For assessment of Affinofile cell receptor expression levels relative to PBMCs, 5.0×105 cells were added to a 6-well plate and allowed to adhere for 48 hours in Dulbecco's Modified Eagle's Medium (DMEM, Mediatech, Inc.) supplemented with 10% FBS, pen/strep, and 50 µg/ml blasticidin (Sigma). Cells were stimulated with 20 ng/ml minocycline (Sigma) for 24 hours and subsequently recovered from plates with 3 mM EDTA in PBS. Cells were washed, stained, and analyzed by flow cytometry as reported above for PBMCs. Flow cytometry of PBMCs and Affinofile cells was performed in the same experiment. Results were analyzed by Flow-Jo software and receptor expression levels reported as events relative to mean fluorescence intensity.
Envelope-pseudotyped viruses prepared with ES or CP envelopes by transfection of 293T cells were quantified by limiting dilution reverse transcriptase activity. Equivalent virion numbers were pelleted by centrifugation at 38,000×g for 2 hours at 4°C. Supernatant was removed and virion pellets were lysed in SDS lysis buffer [40 mM Tris-HCl (pH 6.8), 10% glycerol, 10% ß-mercaptoethanol, 1% SDS]. Virus lysates were separated on SDS-10% polyacrylamide gels and transferred to nitrocellulose. Membranes were blocked with gelatin and proteins were detected either with a mouse monoclonal anti-gp120 antibody that recognizes a conserved C2 region linear epitope (B13, courtesy of Dr. Bruce Chesebro, NIAID ) or HIV-Ig (courtesy AIDS Research and Reference Reagent Program). Primary antibodies were detected with horseradish peroxidase-conjugated goat-anti-mouse or goat-anti-human secondary antibodies, respectively (Pierce Biotechnology, Rockford, IL), revealed with the ECL Plus Western Detection kit (Pierce Biotechnology) and exposed to X-ray film.
Data were analyzed by the UCLA Statistical/Biomathematical Consulting Clinic using repeated measures ANOVA, simultaneously taking into account effects due to disease group (ES vs CP), CD4 level, and CCR5 level. Within each group, the intrapatient variability in relative infection was small and did not differ significantly between the ES and CP groups. Therefore, results are reported with interpatient variance. For evaluation of surface plots between groups ES and CP, P values are given using the average of clones for a given patient as a single value or using each individual clone as a single value. For groups of 3 or more (ES, CP, and acute) we evaluated independent means by One-way ANOVA using the Kruskal-Wallis test and Dunns post-test for data that did not pass a normalcy test. For drug sensitivity and kinetic analysis statistics were performed using each individual clone as a single value. We considered a P value of <0.05 as statistically significant.
The ability of HIV-1 to infect a cell is largely influenced by surface expression of CD4 and CCR5 [26]–[32]. This study evaluates a previously described cohort of 38 independent full-length plasma env clones derived from 7 ES individuals [23]. The env clones from this cohort expressed similar levels of protein by Western blot (Figure S1) and readily infected the indicator cell line TZM-bl demonstrating their functionality [23]. As described below, observed differences in entry efficiency could not be explained by any minor variations in Env levels on the virus.
As with most cell lines, TZM-bls express CD4 at levels comparable to primary activated CD4+ T cells (approximately 65,000–100,000 molecules/cell), however CCR5 expression is significantly higher than on primary T lymphoctyes (approximately 500 to 7000 molecules/cell) [26], [33]–[37]. CCR5 expression also varies widely from patient to patient not only in absolute number of cells expressing CCR5, but also in CCR5 density/cell [26], [27], [33]–[35]. This study utilizes the Affinofile system, a novel cell line with independent dual-inducible surface expression of CD4 and CCR5 (Figure 1, Figure S1) (Johnston et al., submitted). The ability to modulate receptor and co-receptor expression on the Affinofile cells provides a more physiologic measure of HIV-1 entry efficiency. Since the ability of HIV-1 to infect a cell is largely influenced by cell surface levels of CD4 and CCR5, it is important to consider expression levels when evaluating infectivity [26]–[32].
Receptor usage as measured by the Affinofile system was validated as a surrogate marker of entry fitness. Yu-2 and a V3 crown mutant of Yu-2 [Yu-2(Y318R)] known to affect CCR5 usage were evaluated for infectivity using Affinofile cells induced at each pairwise combination of [Minocycline] and [Ponasterone A] (42 unique combinations) (Figure S1). Three-dimensional surface plots were generated from luciferase activity expressed as a function of virus infectivity at each combination of CCR5 and CD4, which was confirmed by flow cytometry (Figure 1A). CD4 and CCR5 surface levels at each drug combination are given as an average level calculated from a pool of cells expressing a range of CD4 and CCR5 molecules (Figure S4). Reduced infectivity of the Yu-2(Y318R) variant over the wild type was observed over a range of CCR5 and CD4 (Figure 1A). This decreased ability of Yu-2(Y318R) to infect cells expressing low CCR5 is consistent with a 90% reduction in replicative fitness measured by competitive replication assays in peripheral blood mononuclear cells (PBMCs) (Figure 1B, p<0.01, unpaired student's t-test). The direct relationship between entry efficiency using the Affinofile system and replicative fitness in human PBMCs has been validated for multiple primary HIV-1 isolates. Generally, viruses of increased replicative fitness display increased infectivity of cells expressing low CCR5, CD4, or both CCR5 and CD4 in the Affinofile system (Johnston et al, submitted).
To assess relative infectivity of chronic and ES env clones, pseudotyped viruses carrying a non-LTR driven luciferase were generated for each clone. Pseudotyped viruses generated from 38 independent plasma virus env glycoprotein clones from 7 ES and 32 independent plasma virus clones from 7 chronic progressors (CP) were evaluated for infectivity at each pairwise drug combination described above and surface plots were generated (Figure S2 and Figure S3). The percent infection defines the infection at each surface CCR5/CD4 level relative to a 100% infection at the highest CD4 and CCR5 surface density. This method permits the direct comparison of CD4 and CCR5 usage by each env clone and provides a rapid and efficient way to measure viral env replicative fitness.
Relative infectivity of 38 independent ES env clones and 32 independent CP env clones was ascertained at multiple combinations of CCR5 and CD4 density (Figure 2A–2F and Figure 3A–3F). Values for each of the env clones tested are shown as well as for Yu-2 and SF162. Varying CD4 levels with constant CCR5 (Figure 2A–2F) or varying CCR5 levels with constant CD4 (Figure 3A–3F) consistently demonstrated that ES env clones supported lower levels of infection than the CP clones.
At the highest CD4 and lowest CCR5 expression level, ES clones averaged 36.7% while CP clones were reliably higher averaging 53.3% (Figure 2A–2C). Infectivity differences were significant for each CD4 concentration (P values ranged from 0.01 to <0.0001, repeated measures ANOVA) at a fixed high or low CCR5 level regardless if individual env clones were evaluated (Figure 2C and 2F) or if the env clones were averaged for a given individual and compared as patient averages (Figure 2B and 2E). Additionally, consistent with previous data, the neurotropic envs SF162 andYU-2 readily infected cells expressing sub-threshold levels of CD4 while the primary isolates could not (Figure 3A–3F) [38]. Taken together, these results reveal that ES clones inefficiently infect cells expressing low CCR5 in the presence of threshold or higher levels of CD4 compared to CP clones. Additionally, the discrepancy in infectivity between ES and chronic clones at fixed, high CCR5 levels indicates that ES clones also require higher levels of CD4 to achieve similar infection as chronic clones.
To further evaluate CCR5 usage independent of CD4 expression, infection was determined at minimal and maximal CD4 levels as CCR5 expression was varied. At each concentration of CCR5 below maximal examined, ES clones infected significantly less efficiently than chronic clones (Figure 3D–3F). Therefore, even in the presence of optimal CD4 concentrations, ES clones inefficiently utilize CCR5 for entry.
Differences in receptor and co-receptor utilization of ES and chronic progressor envs were most significant when infection was performed in the context of (1) low CCR5 and varying CD4 levels or (2) low CD4 and varying CCR5 levels. Thus, these conditions were repeated to examine entry efficiency of 23 pseudotyped env plasma clones from 20 acutely infected individuals (Figure 4A and 4B). At low CCR5 expression, acute envs averaged an intermediate pattern of infectivity compared to ES and chronic envs, but these differences were not significant (Figure 4A). Similar results were obtained when infections were performed at minimal surface CD4 levels (Figure 4B). Consistent with previous reports using similar systems, these results indicate that acute envs show a broad pattern of infectivity which is not significantly different from chronic or ES envs [39].
Previous studies suggest that major differences in entry efficiency may impact on susceptibility to various entry inhibitors [25], [40]–[42]. Thus ES, CP, and acute clones were tested for their susceptibility to the natural CCR5 ligand CCL5 (RANTES), the small molecule CCR5 antagonist TAK-779, and the fusion inhibitor enfuvirtide (ENF) in Affinofile cells induced to express CD4 and CCR5 to levels that closely mimic primary CD4+ T cells (approximately 125,000 molecules CD4/cell and 1274 molecules CCR5/cell) (Figure S4). Susceptibility to CCL5 did not differ significantly between chronic and ES env clones (Figure 5A). Interestingly, acute clone IC50 values ranged from 0.05 to 50 nM (1,000-fold) with an average of 8.25 nM. The range of IC50 values for acute clones was much larger relative to ES clones whose values ranged from 0.4 to 10 nM (25-fold), with an average of 3.05 nM. Although similar trends were observed with the small molecule CCR5 antagonist TAK-779, clones were not significantly different with average IC50 values of 48.8 nM for acute, 21.0 nM for chronic, and 14.7 nM for ES (Figure 5B). These results show that ES, CP, and acute envs have no remarkable differential susceptibility to CCL5 or TAK-779 however the broad range of IC50 values for acute envs highlights the variability in env phenotypes associated with acute infection.
Finally, susceptibility of clones to ENF was evaluated (Figure 5C). IC50 values for acute clones again showed a broad range from 0.10 to 623 nM (>5000 fold range) with an average of 58.13 nM. Again, ES (33-fold range) clones exhibited a significantly narrower range of IC50 values compared to acute clones. The average IC50 value for acute envs was significantly greater than for ES (P<0.05, ANOVA, Kruskal-Wallis test). Overall, ES clones showed trends towards increased susceptibility to entry inhibitors consistent with decreased entry efficiency. These susceptibility profiles suggest that ES clones have a high degree of phenotypic similarity indicated by the narrow range of susceptibility to entry inhibitors. Conversely, chronic and especially acute envs showed broad ranges of susceptibility indicative of their diverse entry phenotypes.
Infection data of cells expressing sub-maximal concentrations of CD4 and CCR5 indicates that ES-derived env clones require higher levels of receptor and co-receptor for efficient entry. This requirement for higher receptor levels could suggest that these envs also exhibit differences in the rates of HIV-1 entry into host cells [42],[43].
To assess host cell entry kinetics, U87-CD4/CCR5 cells were first spinoculated with virus at a temperature non-permissive for viral fusion with the host cell. Enfuviritide (ENF) was added once to each well at a concentration of 10 µM at various times after the cells were shifted to temperatures permissive for viral fusion (Figure 6A). Viruses that have completed the final step in HIV-1 entry (six helix bundle formation) are ENF insensitive and will continue the viral replication cycle regardless of the addition of ENF. Thus, this assay permits determination of the entry kinetics of each env clone.
Fusion kinetics were measured for each ES, CP, and acute clone. ES clones fused with an average T1/2 of 92.1 minutes while acute clones averaged a T1/2 of 67.5 minutes and chronic clones a T1/2 of 58.3 minutes (Figure 6B). This delay in ES fusion kinetics was significant when compared with acute and CP clones (P<.0001 and P<.0001 respectively, One-way ANOVA). Therefore, even in the presence of saturating levels of CD4 and CCR5, ES clones do not complete entry processes as efficiently and exhibit slower kinetics than both CP and acute env clones.
This delay in entry kinetics was maintained during subsequent steps of the replication cycle as indicated by a kinetic reverse transcription assay. For these analyses, Efavirenz (a non-nucleoside reverse transcriptase inhibitor) was added at various times post-infection to arrest infection events which have not completed reverse transcription. As expected, ES derived env clones completed reverse transcription slower (mean T1/2 of 8.89 hours) than acute and CP clones [mean T1/2 values of 7.74 (P<.0001) and 7.95 (P<.0001) respectively with One-way ANOVA] (Figure 6C). Due to the isogenic background of the pseudotyping virus these results suggest that delays in reverse transcription are the result of delays in entry processes. ES env clones exhibit a kinetic lag in entry processes which are maintained during downstream events in the viral life cycle.
This study represents an evaluation of intrinsic phenotypic characteristics of full-length functional subtype B env quasispecies derived from ES plasma. Envelope glycoproteins from ES clearly exhibited reduced capacity to support HIV-1 entry into host cells compared to CP envs. Given the wide range in entry efficiencies observed with envs derived from acute infections it is possible that relatively lower fitness env variants are selected early in infection in ES. The impact of this observed entry deficiency with ES clones is still not fully understood but decreased replicative fitness and lack of diversification in ES viruses suggests these individuals may have contracted a less fit HIV-1 variant or these low fitness variants are selected for early in infection[7],[44],[45].
To date, phenotypic studies of ES viruses have been difficult to perform due to the low amount of virus in these individuals. Analysis of minor differences in env function has been confounded by the use of cell lines expressing non-physiologic amounts of co-receptor (CCR5). Given the high degree of variability in expression of CCR5 among patients it is important to evaluate env function over a wide range of CCR5 levels [26],[27],[34],[35]. Detailed analyses of env function in the presence of physiologic levels of CD4 and CCR5 was possible in this study through the use of the novel Affinofile system. Although ES and chronic individuals each harbored quasispecies with different receptor utilization phenotypes, ES clones from each individual showed an average decreased entry efficiency compared to chronic clones over almost all CD4 and CCR5 expression levels. These differences between chronic and ES clones were most dramatic at low CCR5 surface levels. Thus, this low fitness phenotype could be further accentuated in vivo in an individual expressing low CCR5 levels or possibly higher levels of CCR5 ligands which have been associated with viral control [46]–[48].
Several reports have suggested a correlation between susceptibility to entry inhibitors and relative env fitness [25],[40]. ES, CP, and acute clones exhibited a diverse range of IC50 values consistent with previous data [40],[41],[43],[49],[50]. However, acute clones showed consistently the most variation in susceptibility indicating the diverse phenotypes associated with early infection [40],[43],[49]. Conversely, the low range of IC50 values for ES clones underscores their phenotypic homogeneity. Due to reported differences between primary cells and cell lines entry inhibitor susceptibility assays were performed in the Affinofile cells induced to express CD4 levels and CCR5 levels that closely mimic primary CD4+ T cells (approximately 125,000 molecules CD4/cell and approximately 1274 molecules CCR5/cell) [51]. Additionally, it would be expected that variations in CD4 utilization would result in variations in susceptibility to soluble CD4 (sCD4) [52]. However, consistent with previous reports showing the relative resistance of primary isolate viruses to sCD4 [53], meaningful inhibition of ES, CP, and acute infection envelopes was not observed at maximal achievable concentrations of sCD4 (25 µg/ml).
Previous studies have also highlighted an association between receptor utilization profiles and susceptibility to neutralizing antibodies. It is possible that ES envs display altered susceptibility to broadly neutralizing antibodies given their observed entry phenotype. Previous studies suggest that antibody binding and neutralization have a kinetic component [54]. It may potentially be generalized that slow fusing viruses, independent of the mechanism, may be more susceptible to neutralizing antibodies that act with a kinetic dependence. However, it has been shown that ES individuals generate low titers of neutralizing antibodies against autologous virus and thus the role of neutralizing antibodies in maintenance of low level viremia in ES is nominal [23].
The host entry process is thought to be a rate limiting step in HIV-1 replication. It was thus important to determine if poor entry efficiency by ES clones leads to a reduced rate of entry kinetics. ES env clones were found to fuse on average over 1.5 times slower than chronic or acute clones in the presence of saturating levels of both CD4 and CCR5. As result of poor entry efficiency, this kinetic delay was maintained during subsequent steps of the retroviral lifecycle. Compounded effects of inefficient CD4 and CCR5 usage by ES clones likely contributes to kinetic delays in entry processes and thus overall decreased replicative fitness. The delayed entry kinetics may be even more important in vivo if there is a limited time frame over which entry can occur due to competing inhibitory processes such as the binding of neutralizing antibodies or the presence of CCR5 ligands.
Despite genotypic differences in both the virus and the ES host, viral quasispecies in different ES individuals are remarkably similar phenotypically. This implies that poor envelope function is a common feature in ES individuals. This result is in sharp contrast to data from acute infection envs where clones exhibited much phenotypic diversity. Potentially, HIV-1 infection in ES may be established by lower fitness env(s) which are present in a subset of acutely infected individuals. Alternatively, HIV-1 infection in ES may be established by phenotypically diverse envs and early pressure from the immune results in the outgrowth of lower fitness escape variants. At present, no data exists on the natural history of acute infection of ES. It remains unclear whether these individuals experience typical high level viremia that is subsequently reduced to an undetectable setpoint or control their viral load from the onset of infection. It would be of great interest to be able to address this significant gap in our understanding of viral dynamics in elite control of viremia. Lower env fitness is likely not sufficient to mediate absolute viral suppression. Viral control could be achieved in those individuals who are also able to mount a potent immune response and/or are genetically predisposed to better control HIV-1 viremia. In these individuals, viral replication and diversification of early infection viruses required to achieve efficient receptor utilization by env quasispecies may never be attained. Lower fitness of other viral factors may also be contributing to reduced replication and lower viral load. Full understanding of the in vivo impact of lower env fitness in ES will require further study however this data underscores the important contribution of viral factors in elite HIV-1 suppression.
|
10.1371/journal.pgen.1005874 | Six Novel Loci Associated with Circulating VEGF Levels Identified by a Meta-analysis of Genome-Wide Association Studies | Vascular endothelial growth factor (VEGF) is an angiogenic and neurotrophic factor, secreted by endothelial cells, known to impact various physiological and disease processes from cancer to cardiovascular disease and to be pharmacologically modifiable. We sought to identify novel loci associated with circulating VEGF levels through a genome-wide association meta-analysis combining data from European-ancestry individuals and using a dense variant map from 1000 genomes imputation panel. Six discovery cohorts including 13,312 samples were analyzed, followed by in-silico and de-novo replication studies including an additional 2,800 individuals. A total of 10 genome-wide significant variants were identified at 7 loci. Four were novel loci (5q14.3, 10q21.3, 16q24.2 and 18q22.3) and the leading variants at these loci were rs114694170 (MEF2C, P = 6.79x10-13), rs74506613 (JMJD1C, P = 1.17x10-19), rs4782371 (ZFPM1, P = 1.59x10-9) and rs2639990 (ZADH2, P = 1.72x10-8), respectively. We also identified two new independent variants (rs34528081, VEGFA, P = 1.52x10-18; rs7043199, VLDLR-AS1, P = 5.12x10-14) at the 3 previously identified loci and strengthened the evidence for the four previously identified SNPs (rs6921438, LOC100132354, P = 7.39x10-1467; rs1740073, C6orf223, P = 2.34x10-17; rs6993770, ZFPM2, P = 2.44x10-60; rs2375981, KCNV2, P = 1.48x10-100). These variants collectively explained up to 52% of the VEGF phenotypic variance. We explored biological links between genes in the associated loci using Ingenuity Pathway Analysis that emphasized their roles in embryonic development and function. Gene set enrichment analysis identified the ERK5 pathway as enriched in genes containing VEGF associated variants. eQTL analysis showed, in three of the identified regions, variants acting as both cis and trans eQTLs for multiple genes. Most of these genes, as well as some of those in the associated loci, were involved in platelet biogenesis and functionality, suggesting the importance of this process in regulation of VEGF levels. This work also provided new insights into the involvement of genes implicated in various angiogenesis related pathologies in determining circulating VEGF levels. The understanding of the molecular mechanisms by which the identified genes affect circulating VEGF levels could be important in the development of novel VEGF-related therapies for such diseases.
| Vascular Endothelial Growth Factor (VEGF) is a protein with a fundamental role in development of vascular system. The protein, produced by many types of cells, is released in the blood. High levels of VEGF have been observed in different pathological conditions especially in cancer, cardiovascular, and inflammatory diseases. Therefore, identifying the genetic factors influencing VEGF levels is important for predicting and treating such pathologies. The number of genetic variants associated with VEGF levels has been limited. To identify new loci, we have performed a Genome Wide Association Study meta-analysis on a sample of more than 16,000 individuals from 10 cohorts, using a high-density genetic map. This analysis revealed 10 variants associated with VEGF circulating levels, 6 of these being novel associations. The 10 variants cumulatively explain more than 50% of the variability of VEGF serum levels. Our analyses have identified genes known to be involved in angiogenesis related diseases and genes implicated in platelet metabolism, suggesting the importance of links between this process and VEGF regulation. Overall, these data have improved our understanding of the genetic variation underlying circulating VEGF levels. This in turn could guide our response to the challenge posed by various VEGF-related pathologies.
| Vascular Endothelial Growth Factor (VEGF) is secreted largely by endothelial cells and plays a key role in several physiological and pathological conditions. During growth, development, and maintenance of the circulatory system, VEGF is the principal pro-angiogenic factor and it has additionally, a neurotrophic role. High levels of circulating VEGF have been observed in individuals with various vascular diseases (myocardial infarction [1], stroke [2,3], heart failure [4], and atherosclerosis [5]), neurodegenerative conditions (age-related cognitive decline [6] and Alzheimer dementia [7]), immune inflammatory disorders (rheumatoid arthritis [8], inflammatory bowel disease [9], and Behçet’s disease [10]) and cancers (breast [11,12], uterine [13], gastrointestinal [14,15], lung [16] and prostate [17]). An increase of VEGF levels has also been found in patients with diabetes [18] and various reproductive disorders [19–21]. Reduced circulating VEGF levels have been observed in amyotrophic lateral sclerosis [22] and spinal bulbar muscular atrophy [23]. Moreover, since VEGF levels are pharmacologically modifiable, understanding the determinants of circulating VEGF could support efforts directed at risk prediction, prevention and therapy. Circulating VEGF levels are highly heritable [24–27] leading to a search for specific genetic determinants within the Vascular Endothelial Growth Factor A (VEGFA) gene [27–29]. Several putative candidate genes were then identified but could not be consistently replicated [10,30–41]. A genome-wide linkage study of VEGF levels identified the 6p21.1 VEGFA gene region as the main quantitative trait locus determining variation in VEGF serum levels [27]. Specific variants at this locus were also identified as the strongest associations in the first genome-wide association study (GWAS) of circulating VEGF levels based on data from 3 large cohort studies in this consortium, wherein two addition loci, located at 8q23.1, and 9p24.2 were also identified [42]. We have now conducted a new GWAS meta-analysis using an extended sample, the largest to date, and a deeper genomic coverage based on imputation to the 1000 genomes panel to identify additional genetic variants that explain variation in circulating VEGF concentrations.
A GWAS meta-analysis of VEGF levels was performed in 16,112 individuals from 10 cohorts of European ancestry (see Materials and Methods and Section 1 in S1 Text for details): the Age Gene/Environment Susceptibility Reykjavik Study (AGES), the Cilento study (Cilento), the Framingham Heart Study (FHS), the Ogliastra Genetic Park (OGP), the Prospective Investigation of the Vasculature in Uppsala Seniors Study (PIVUS), and the Val Borbera study (VB) served as discovery cohorts; the Gioi population, the Sorbs population, the STANISLAS Family Study (SFS) and a sample of hypertensive adults (HT) served as replication cohorts. The characteristics of study participants are shown in Table 1. The mean age of the participants was 54.8 years, ranging from 30.4 years in SFS to 76.2 years in the AGES. The percentage of females in the overall sample was 54%, ranging from 37% in OGP to 64% in Sorbs. To account for differences in age distribution and gender among the studies, both age and sex were subsequently used as covariates in the association analyses. Across studies, median VEGF levels ranged from 27.0 to 393.6 pg/ml, with the lowest median levels in HT and SFS studies in which VEGF was measured in plasma rather than serum (see Section 2 in S1 Text for details). This is expected since VEGF levels are higher in serum than in plasma secondary to VEGF release from platelets during clot formation [43,44]. Differences in VEGF levels also partly reflect demographic and assay differences between the cohorts.
An overview of the study design is presented in Fig 1. Due to heterogeneity in the distribution of VEGF levels among the cohorts (Table 1), a sample size-weighted Z-score (rather than an inverse-variance) method was chosen for the meta-analysis. A discovery GWAS meta-analysis was carried out for 6,705,861 autosomal variants in 13,312 individuals from the six cohorts described in the “Characteristics of study participants” section (Stage 1). A Quantile-Quantile plot for the investigated variants revealed many more variants with lower observed p-values (P) than expected (S1 and S2 Figs).
There were 920 variants in 5 chromosomal regions (6p12.1, 8q23.1, and 9p24.2, which have been previously described and two novel regions at 5q14.3 and 10q21.3) that reached genome-wide significance (P<5x10-8) in the discovery sample (S2 Table). To identify independently associated variants within these 5 genome-wide significant genomic regions, conditional analyses were carried out in the study with the largest number of samples (FHS). This approach was selected since our use of a Z-score meta-analysis, which does not yield effect size estimates, precluded the use of aggregate results for conditional analyses. The conditional analyses revealed 10 independent signals (4 previously known and 6 novel variants). These 10 Stage 1 variants were carried forward to in-silico (Stage 2) and subsequent de-novo (Stage 3) replication.
Further, 57 variants in 13 loci were suggestively associated at 5x10-8<p-value<1x10-5. At each locus, a single independent signal was identified using a clumping procedure, and the most strongly associated variant at each of these 13 loci was also tested in the in-silico replication. Among them, 2 variants reached a genome-wide level of significance in the joint meta-analysis of discovery and in-silico replication samples and these two were also carried forward for the de-novo replication. So a total of 12 variants were carried forward to the de novo replication.
Overall, 10 of these 12 variants, 8 of the 10 independent variants identified in Stage 1 and the 2 variants identified in Stage 2 (combined discovery and in-silico replication), were successfully replicated in the Stage 3 meta-analysis of the combined discovery, in-silico, and de-novo replication samples (Fig 2 and Table 2).
For these variants, an additional inverse variance-weighted meta-analysis was performed as a secondary analysis on the Stage 3 data, including the discovery and both replication cohorts. These secondary meta-analysis results, reported in the Table 2, are concordant with our original analysis results.
Forest plots reporting the effects of the 10 replicated variants in all the cohorts and the cumulative effect in the inverse-variance meta-analysis are shown in the Fig 3.
Among those 10 signals, 4 were located in novel chromosomal regions (5q14.3, 10q21.3, 16q24.2, and 18q22.3) and 6 (2 novel, independent variants and 4 previously known signals) were located in previously identified chromosomal regions (6p21.1, 8q23.1, and 9p24.2).
The leading SNP on chromosome 5q14.3 was rs114694170 (P = 6.79x10-13). This new association is located in the intronic region of the myocyte enhancer factor 2C (MEF2C) gene. Conditional analyses did not identify additional independent variants in the region.
In the locus on chromosome 10q21.3, the most significantly associated variant was rs74506613 (proxy rs10761741 used for in-silico replication has r2 of 0.97, P = 1.17x10-19) located within the intronic region of the jumonji domain containing 1C (JMJD1C) gene. Conditional analyses did not identify any other independent variants in this region.
Two additional loci reached a genome-wide significance level in the meta-analysis of the combined discovery and replication samples. At the locus on chromosome 16q24.2, the most significantly associated variant was rs4782371 (P = 1.59x10-09) located within the intronic region of the zinc finger protein, FOG family member 1 (ZFPM1) gene. At chromosome 18q22.3, the leading variant was rs111939830 which along with the second leading variant rs2639990 (used as proxy for de novo replication for rs111939830, r2 = 0.48, P = 1.72x10-08) was located in the intronic region of the zinc binding alcohol dehydrogenase domain containing 2 (ZADH2) gene.
The most significant variant on chromosome 6p21.1 was rs6921438 (P = 7.39x10-1467), already identified in the previous GWAS [42]. Two additional independent variants were also identified at this locus after conditional analyses. One was rs1740073 (P = 2.34x10-17) which was in LD with rs4416670 reported in the previous GWAS (r2 = 0.15) [42]. Although the LD between these two SNPs is relatively low, rs4416670 and rs1740073 are in close physical proximity (3055 base-pair distance) and conditional analysis confirmed that rs1740073 eliminated the signal of rs4416670 (P = 4.16x10-21; before adjusting for rs1740073, P = 0.727; after adjusting for rs1740073), hence we believe the two SNPs, rs1740073 and rs4416670, both represent a single locus of genetic variation. This rs1740073 SNP is located about 22Kb downstream from rs6921438 and both are located upstream of the gene C6orf223, which encodes an uncharacterized protein. The other independent variant identified, about 221kb distant from the main signal rs6921438, was rs34528081 (P = 1.52x10-18), a novel variant, located upstream of the VEGFA gene and the mitochondrial ribosomal protein S18A (MRPS18A) gene. The values of r2 between the 3 variants at 6p21.1 are extremely low (rs6921438-rs1740073 = 0.01, rs6921438-rs34528081 = 0.007, rs1740073-rs34528081 = 0.01), suggesting that the 6p21.1 region has 3 independent variants that modulate circulating VEGF levels.
The leading variant identified on chromosome 8q23.1 was rs6993770 (P = 2.44x10-60). This SNP, located within an intron of the zinc finger protein multitype 2 (ZFPM2) gene, was already known to be associated with circulating VEGF levels [42].
On chromosome 9p24.2 the most significantly associated SNP was rs2375981 (P = 1.48x10-100, which is in strong LD with rs10738760 (r2 = 0.81) reported in the previous GWAS [42]). This variant lies downstream of the very low-density lipoprotein receptor (VLDLR) and upstream of the potassium voltage-gated channel subfamily V member 2 (KCNV2) genes. One novel independent signal also found in this region using conditional analyses was rs7043199 (P = 5.12 x10-14) located about 71kb upstream of rs2375981, in the VLDLR-AS1 gene and upstream of the VLDLR gene. No LD exists between the two variants (r2 = 0.0008). Thus, in the 9p24.2 region, there are 2 independent variants able to influence VEGF levels.
A genetic score was calculated for each individual using information on the 10 VEGF replicated variants. This genetic score explained 52% of the observed variability in circulating VEGF levels in FHS. The proportions of variance in circulating VEGF explained by these 10 replicated variants in PIVUS, Cilento, AGES, VB, HT, and SFS are 48%, 46%, 24%, 24%, 21% and 19%, respectively. The observed differences in the proportion of variance explained might be due to heterogeneity in effect sizes of some SNPs related to the trait variability in distribution of VEGF levels across the cohorts (Table 2). Accordingly, the explained variability is similar in the cohorts where a similar distribution of VEGF levels was observed (Table 1).
To identify putative functional elements at the associated loci, ENCODE data related to chromatin modifications and hypersensitivity DNAse sites (DHSs) included in HaploReg [45] were analyzed. Among the 10 replicated variants and their 126 proxies (r2>0.8), 16 variants were located in regions reported as DHSs in 5 or more different cell lines. Among these 16, 11 variants (rs114694170 on chromosome 5p14.3, rs6993770 on chromosome 8q23.1, rs7043199 on chromosome 9p24.2, 5 proxies of rs74506613 on chromosome 10q21.3 and 3 proxies of rs4782371 on chromosome 16q24.2) were also located in a promoter and/or enhancer histone mark. These results suggest a potential functional role of these variants.
A large database assembled by one of the authors (AJD) that included eQTL association results from 61 studies (detailed Section 3 in S1 Text) was queried for the 10 replicated variants identified in the GWAS and their 126 proxies (r2>0.8). Eighty-four variants in three loci (1 replicated variant and 83 proxies of two additional replicated variants) were found in the database. The variant rs6993770 on chromosome 8q23.1 was a trans eQTL for the CXCL5 gene; rs609303 (proxy of rs111939830) on chromosome 18q22.3 was a cis eQTL for the TSHZ1 gene. On chromosome 10q21.3 82 proxies for rs74506613 were identified: 2 variants were trans eQTL for 6 genes (AQP10, CXCL5, GUCY1A3, ITGA2B, MYL9, and NRGN) and 81 were cis eQTLs for 3 genes (JMJD1C, NRBF2 and REEP3); one variant rs10761779 is both a trans and cis eQTL. All 84 variants identified as eQTL in this search are listed in S3 Table.
In order to identify biological pathways involved in the modulation of VEGF protein levels two pathway analysis approaches were applied. MAGENTA software [46] was applied to the Stage 1 meta-analysis results, to identify the known biological pathways most strongly represented among all the variants associated with circulating VEGF concentrations (see Materials and Methods). Overall, 3,216 biological pathways (with at least 10 genes) and 168,932 genes were examined. This pathway analysis identified 18 biological pathways, 3 molecular functions and 2 cellular components significantly associated with VEGF levels at a nominal Gene Set Enrichment Analysis (GSEA) p-value ≤0.01. Among these, only the ERK5 pathway reached statistical significance after correction for multiple testing (FDR threshold of 0.05).
The Ingenuity Pathway Analysis software (IPA, www.qiagen.com/ingenuity) was used to explore functional relationships between genes in the VEGF associated loci. A total of 26 genes located at and adjacent to the 10 replicated variants were selected as focus genes for IPA analysis (S4 Table). Among them, 17 genes were found to be biologically linked in a unique network of 70 molecules as shown in Fig 4. The associated functions for this network were organism development, especially early embryonic and later cardiovascular system development. The probability that 17 genes would be linked in a randomly designated set of 26 genes using data from the Global Molecular Network was 1.0x10-42. Thus, it appears extremely unlikely that this network has been identified purely by chance.
In this GWAS meta-analysis of circulating VEGF levels, we identified 10 independent variants located in 7 chromosomal loci; 4 of those variants had been described in a previous GWAS [42]. We now describe 6 novel variants, 4 of which were in newly identified chromosomal regions (5q14.3, 10q21.3, 16q24.2, and 18q22.3) whereas 2 were identified through conditional analyses at previously described loci (6p21.1 and 9p24.2). These 10 variants explain about 52% of VEGF phenotypic variance in the largest cohort in this study, with the 6 novel variants increasing the explained variance by 4% compared to the 48% described by Debette et al. for the 4 previously identified loci [42]. This increase represents a valuable addition to the proportion of variance explained when compared to the results obtained from GWAS of other complex traits [47–50].
The newly identified regions include many interesting and plausible candidate genes with angiogenic and neurotrophic roles.
The leading variant on chromosome 5 was located within an intron of the MEF2C gene. This protein has a demonstrated role in cardiac myogenesis, morphogenesis and in vascular development. MEFC2 knock out is embryonically lethal due to cardiac and vascular abnormalities. MEFC2 also supports cortical development and variants in this region have been associated with severe neurodevelopmental problems in humans such as developmental retardation, cerebral malformations [51,52], stereotypic movements and epilepsy. MEF2C was also reported to be associated with retinal vascular caliber in the Cohorts for Heart and Ageing Research in Genomic Epidemiology (CHARGE) consortium [53], which is particularly interesting given the known role of VEGF in proliferative retinopathy and macular degeneration. MEF2C may be a transmitter of VEGF signaling and has been shown to be regulated by VEGF in-vitro, as a key mediator [54].
The leading variant on chromosome 10 was located in an intronic region of JMJD1C, a protein-coding gene with an intriguing role in many biological processes ranging from platelet and endothelial cell function to DNA repair [55]. Thyroiditis [56] and fatty liver disease [57] have been associated with this gene. A GWAS of plasma liver enzymes revealed an association of rs7923609 (P = 6.0x10-23, G = risk allele) with elevated enzyme levels indicating abnormal liver function. Interestingly, this SNP also showed an association with VEGF levels in our study (P = 1.15x10-12) with the G allele associated with higher levels [58]. In a mouse model, it was noted that VEGF promotes proliferation of hepatocytes through reestablishment of liver sinusoids by proliferation of sinusoidal endothelial cells; thus VEGF may mediate the genetic association observed [59] between JMJD1C variants and hepatic steatosis.
JMJD1C and MEF2C genes were found associated to platelet count and volume in a European ancestry GWAS [49]. Further, a variant (rs7896518, P = 2.93x10-15) located in an intron of the JMJD1C gene showed an association with platelet counts (P = 2.3x10-12) in an African American GWAS [60]. In a second European ancestry GWAS of platelet aggregation another SNP in the same gene, rs10761741, showed an association with epinephrine-induced platelet aggregation with the T allele being associated with greater aggregation [61]. Interestingly, this T allele of rs10761741 was also associated with higher circulating VEGF levels (P = 7.10x10-15). Because both platelets and VEGF play important roles in the development of atherosclerosis and arterial thrombosis, investigating the intricate relationships among platelet, VEGF, and JMJD1C might identify novel drug targets and biological pathways implicated in atherosclerosis and arterial thrombosis.
In a GWAS of serum androgen levels in European men a variant (rs10822184) in JMJD1C reached genome-wide significance (P = 1.12x10-8) with the C allele being associated with lower levels [62]. This variant was also associated with higher circulating VEGF levels (P = 4.06x10-11). Further, in a GWAS of sex hormone-binding globulin, the T allele of a variant in JMJD1C (rs7910927) was associated with a decrement of sex hormone-binding globulin concentrations (P = 6.1x10-35) [63]. This T allele was also associated with a decrement of VEGF levels (P = 1.31x10-12). Sex hormones influence VEGF levels [64] thus suggesting a hormone-dependent VEGF production mediated by JMJD1C.
The leading variant in chromosome 18 was located in an intergenic region downstream of the ZADH2 gene and upstream of the Teashirt Zinc Finger Homeobox 1 (TSHZ1) gene and a variant in strong LD with the lead SNP regulates expression of the latter gene. Both genes have been reported as candidate genes for congenital vertical talus [65]. TSHZ1 has been associated with increased expression in Juvenile Angiofibroma (JA) [66]. Because VEGF is secreted by JA, and VEGF contributes to vascularization in JA [67], the investigation of relationships among TSHZ1, JA, and VEGF might lead to a new therapy for JA.
The top variant in chromosome 16 was located in an intron of the ZFPM1 gene. The ZFPM1 gene is also known as Friend of GATA1 (FOG1) gene and is related to ZFPM2, which was identified in our previous meta-analysis [68]. Both proteins are transcription factors that play a role in the development of the heart and coronary vessels. Further, a mutation in the N-finger of the GATA1 gene, abrogating the interaction between GATA1 and FOG1, showed associations with X-linked macro-thrombocytopenia, non-X-linked thrombocytopenia and dyserythropoiesis [69]. It is possible that the observed association between ZFPM1 and serum VEGF levels was partly driven by variations in platelet counts.
Biological pathway exploration using IPA showed that the Ubiquitin C (UBC) gene directly interacted with 10 of the focus genes. The encoded protein is a polyubiquitin precursor [70]. This gene has been associated with progressive accumulation of ubiquitinated protein inclusions in neurodegenerative disorders that involve dysfunction of the ubiquitin-dependent proteolytic pathway [71] and with verbal memory performance [72]. The UBC gene might play an important role in the association between variants and circulating VEGF serum as either mediator or confounder. However, a direct role for the UBC gene in determining circulating VEGF levels was not identified and none of the variants within 60kb of the UBC gene were associated with circulating VEGF level even at a nominally significant level.
Gene set enrichment analysis revealed the ERK5 pathway as significantly enriched for VEGF associations. ERK5 pathway is involved in multiple processes, such as cell survival, anti-apoptotic signaling, cell motility, differentiation, and cell proliferation [73,74]. ERK5 is also involved in the angiogenic process, where it acts as regulator of VEGF expression [75,76]. More recently it has been reported that this molecule is expressed on the platelet surface, and acts as platelet activator in ischemic conditions, such as after a myocardial infarct [77].
Based on eQTL analysis, we observed that 3 of the replicated variants were themselves, or in strong LD with, variants acting as cis and/or trans eQTLs on different genes. In particular, among those identified as trans-regulated genes, there were some very interesting candidates.
The C-X-C motif chemokine 5 (CXCL5) gene was a trans-regulated gene for 3 variants in two VEGF associated regions (rs6993770 on 8q23.1 and 2 proxies of rs74506613 on 10q21.3). It encodes a protein that through the binding of the G-protein coupled receptor chemokine (C-X-C motif) receptor 2, recruits neutrophils [78,79], promotes angiogenesis [80] and is thought to play a role in cell proliferation, migration, and invasion in different types of cancer [81–85]. CXCL5 acts by activating several angiogenic signaling pathways, some of which, including JAK/STAT [86] and Src family kinases [87] pathways, are also activated by VEGF. Given the involvement of the two genes in the same pathways, it is conceivable that they could be co-regulated.
The GUCY1A3 gene encodes the alpha-3 subunit of the Soluble Guanylate Cyclase (sGC), an heterodimeric enzyme that, acting as main receptor of the nitric oxide (NO), catalyzes the conversion of guanosine-5'-triphosphate (GTP) in 3', 5'-guanosine monophosphate (cGMP) and pyrophosphate. This NO-sGC-cGMP pathway controls vascular smooth-muscle relaxation, vascular tone, and vascular remodeling, and is activated by VEGF signaling. Inhibition of sGC reduces VEGF-induced angiogenesis [88,89]. Moreover, activation of sGC inhibits platelet activation [90].
The protein encoded by the MYL9 gene is a myosin light chain that regulates muscle contraction by modulating the ATPase activity of myosin heads. In platelets, MYL9 is associated with MYH9, the major nonmuscle myosin expressed in megakaryocytes and platelets. Defects in the MYH9 gene are responsible of different autosomal dominant disorders characterized by thrombocytopenia and platelet macrocytosis [91,92]. Moreover, it has been demonstrated that MYL9 is involved in pro-platelet formation [93]. In megakaryocytic cells, MYL9 expression is regulated by RUNX1, a major hematopoietic transcription factor whose haplo-deficiency is associated with familial thrombocytopenia, platelet dysfunction, and predisposition to leukemia [94].
The ITGA2B gene encodes the integrin alpha chain 2b, a subunit of the glycoprotein IIb/IIIa, and an integrin complex expressed on the platelet surface. On the activated platelets, it acts as receptor for fibrinogen; this binding induces platelet aggregation, an essential event in thrombus formation, and permits clot retraction. Defects in the ITGA2B gene cause Glanzmann thrombasthenia, an autosomal recessive bleeding disorder characterized by failure of platelet aggregation and by absent or diminished clot retraction [95]. Moreover, a GWAS on platelet count revealed a SNP in the ITGA2B gene region associated with platelets count (rs708382, P = 1.51x10-8) [49]
As for the ZFPM1 and JMJD1C genes, the observed connection between VEGF levels and GUCY1A3, MYL9 and ITGA2B genes could be due, therefore, to a regulation of the number and/or the functionality of the circulating platelets. Overall our data suggest that studies clarifying whether the relationship between these genes and VEGF levels is mediated by platelets may be helpful to better understand the role of these genes in VEGF regulation.
In conclusion, the identification of novel genes and pathways associated with circulating VEGF levels could lead to new preventive and therapeutic strategies for a wide variety of diseases in which a pathophysiological role for VEGF has been implicated.
The major strength of this work is that it is the largest GWAS of circulating VEGF to date. A limitation is that, due to the heterogeneity in VEGF levels among the cohorts, a sample size-weighted Z-score method was used to perform the GWAS meta-analysis, which has lower power to detect associations compared to inverse-variance weighted meta-analysis, hence we may have failed to detect some real associations. Further, our analysis focused mostly on common and less frequent variants. Therefore, we could not comprehensively assess the effect of rare variants on VEGF levels. Identifying rare variants in future studies, could contribute to further increasing the proportion of variance in circulating VEGF explained. Also, our study was confined to individuals of European ancestry. The results need to be replicated in other racial and ethnic groups. Finally, a functional validation of the identified associations is needed.
Six discovery data sets including 13,312 samples were analyzed in the Stage 1. The participating discovery studies were the Age Gene/Environment Susceptibility Reykjavik Study (AGES, n = 1,548), the Cilento study (Cilento, n = 1,115), the Framingham Heart Study (FHS, n = 7,048), the Ogliastra Genetic Park (OGP, n = 897), the Prospective Investigation of the Vasculature in Uppsala Seniors Study (PIVUS, n = 945), and the Val Borbera study (VB, n = 1,759). Two additional studies, the Gioi population (Gioi, n = 470) and the Sorbs population (Sorbs, n = 659) provided data for an in-silico replication (Stage 2). Further a de-novo replication (Stage 3) was undertaken in the STANISLAS Family Study (SFS, n = 676) and in a sample of hypertensive adults (HT, n = 995) from the Biological Resources Center (BRC) IGE-PCV “Interaction Gène-Environment en Physiopathologie Cardio-Vasculaire. The participating cohorts are described further in Section 1 in S1 Text. The local institutional ethics boards for each study approved the study design. Each subject signed an informed consent before participating to the study. Further details can be found in S5 Table.
In the discovery and in-silico replication cohorts, genotyping was performed using various arrays, and imputation was carried out using the 1000 genome v3 as reference panel in all studies. Details of pre-imputation quality control parameters, genotyping platforms and imputation parameters for each study are provided in S1 Table. In all cohorts blood samples were collected after an overnight fast, and serum/plasma samples were prepared and stored as described in Section 2 in S1 Text. Serum VEGF levels (plasma VEGF were measured in SFS and HT) were measured using commercial ELISA assays as detailed in Section 2 in S1 Text. The de-novo genotyping at SFS and HT was undertaken on a competitive allele specific PCR (KASP) chemistry array and variants were called using a FRET-based genotyping system.
In each individual study, a natural log-transformation of VEGF levels was applied. To do that, in a few studies (AGES, OGP, VB, and Sorbs) where some individuals had VEGF levels below the detection threshold of the assay, half the minimum value of VEGF found in that cohort was arbitrarily assigned to each such participant [96]. The transformed trait, adjusted for age, sex and additional study-specific covariates (e.g. principal components associated with VEGF levels, study center for multi-site studies), was related to the variant dosages using a linear regression. Studies with familial correlation used linear mixed effect models to account for familial relatedness. Detailed information about the software used in each cohort is reported in the S1 Table. An additive genetic model with 1 degree of freedom was applied. Study specific results of genome-wide per-variant associations underwent additional quality control prior to meta-analysis. Checking of file formatting, data plausibility, and distributions of test statistics and quality measurements was facilitated by the gwasqc function of the GWAtoolbox package v1.0.0 in R [97]. Prior to the meta-analysis, variants with low minor allele frequency (<1%) and poor imputation quality (r2< 0.4) were removed.
Meta-analysis was performed in METAL using an effective sample size weighted Z-score method [98]. This method was chosen over an inverse-variance meta-analysis because of different covariate-adjusted mean values and standard deviations in VEGF levels among studies. The results of meta-analysis were adjusted for genomic control inflation factor. To define the effective sample size, the product of the sample size and the imputation quality for each variant was calculated in each cohort [99]. The sum of the product of each cohort divided by overall sample size represents the proportion of the effective sample size for each variant Eq (1).
[∑i=1CNi×ri2]/13,312=Effective sample size
(1)
where C is the total number of participating cohorts, i indicates the specific cohort, N is the sample size used for the variant association test, and r2 is imputation quality of the variant. After completing initial quality control checks, 6,705,861 variants, each of which was informative at an effective sample size of >70%, were included in the meta-analysis (Stage 1). The genomic control inflation factor of the metal analysis was 1.003. All variants having a p-value less than 5x10-8 were considered to be genome-wide significant.
To identify all independent associations within the loci reaching genome-wide significance, conditional analyses were performed in a forward stepwise fashion, examining the most significant association and including in successive association models the next most significantly associated variant (P<5x10-8) in a specific region at each step (referred to as the top variant in Eq (2)). We repeated this process until no more genome-wide significant associations were found. The conditional analysis model follows the formula (2).
ln(VEGF)=β0+β1variant+∑i=1nβiCovariatesi+∑j=1kβjTop variantj
(2)
where n is the number of covariates used in the primary GWAS, k is the number of steps. The conditional analysis was only performed in FHS because it represents the largest cohort in the meta-analysis. The final conditional analysis model included 10 independent variants with p-values less than 5x10-8 in FHS.
Genome-wide significant variants identified in the conditional analysis were examined in the two in-silico replication cohorts and also carried forward to de-novo replication. Furthermore, for each suggestive locus (5x10-8<P<1x10-5) the lead variant was also examined in the in-silico replication sample, and those suggestive variants that reached a genome-wide significant p-value in a meta-analysis of the discovery and in-silico replication data (Stage 2) were also carried forward to the de-novo replication phase. To check for the presence of other independent variants in the suggestive regions, a clumping procedure implemented in PLINK [100] was performed. The 1000-genome v3 genotypes were used as reference panel for LD calculation; the physical threshold for clumping was 1 Mb, and the r2 threshold for clumping was 0.1.
For selected variants that failed de-novo genotyping, a proxy variant having either the highest linkage disequilibrium (LD) value, or the variant in the same region with the next lowest p-value was genotyped instead of the lead variant. We considered as replicated, all variants that reached a genome-wide significance level in the meta-analysis of the discovery and the in-silico and de-novo replication samples (Stage 3).
For the replicated variants, an inverse variance-weighted meta-analysis was also performed as a secondary analysis, including in the analysis all the discovery and replication cohorts.
The variants identified after replication stages were used to estimate, in each cohort, a genetic score associated with circulating VEGF levels by summing the product of the beta-estimate and genotype for each variant in a given individual Eq (3).
RiskScore=∑i=110βi*Genotypei
(3)
where i is the variant, β is effect size of the variant in the cohort, and genotype is additively coded genotype of the variant. The proportion of phenotypic variance explained by the variants incorporated in the score was estimated fitting two linear mixed effect models, in which VEGF levels were regressed, respectively, on: 1) gender and age (basic model); 2) gender, age, and genetic risk score (risk score model). The variance explained by the replicated variants was estimated as the difference between the variance explained by the risk score model and that explained by the basic model. The lmekin function (R package), which uses the genomic kinship matrix to correct for relatedness between individuals, if any, was applied.
The replicated SNPs and variants in LD with them (r2>0.8) were investigated for the presence of chromatin histone marks and hypersensitive DNAse elements using data from ENCODE included in Haploreg_v3 software (http://www.broadinstitute.org/mammals/haploreg/haploreg_v3.php) [45].
A database of expression Single Nucleotide Polymorphism (eSNP) was created collecting results from multiple published sources, reported in Section 3 in S1 Text. The eSNP results from each study were included in the database if they met criteria for statistical thresholds for association with gene transcript levels as described in the original references. To search for eQTLs among the associations found in the meta-analysis we queried this database for the replicated variants and their proxies (r2>0.8).
Two different approaches were used to identify biological pathways influencing VEGF variability.
The GSEA-like statistical test implemented in MAGENTA program was used to test the over-representation of genes containing VEGF-associated variants in a given biological pathway. To do that, all data of meta-analysis results from Stage 1 were used and the gene-set annotations from the Kyoto Encyclopedia of Genes and Genomes (KEGG), PANTHER, INGENUITY, Gene Ontology, REACTOME and BIOCARTA databases were applied. Each gene in the genome was scored by the most significant association p-value among all the SNPs located within a region from 110 kb upstream to 40 kb downstream of each gene’s transcript boundaries. Confounding effects on gene association scores were identified and corrected for. This “normalized best gene score” was used to evaluate the gene enrichment against a null distribution of 10,000 gene sets of identical set size that are randomly sampled from the genome. The 95th percentile of all gene scores for the meta-analysis was used as the enrichment cutoff. Genes within the HLA-region were excluded from analysis due to difficulties in accounting for gene density and LD patterns and only gene sets with at least 10 genes were included in the analysis. Significance was determined when an individual pathway reached a false discovery rate (FDR)<0.05.
The Ingenuity Pathway Analysis software (IPA) was used to explore the functional relationship between genes of interest, selected from candidate regions. For this purpose, a candidate region was defined as comprising all variants between the first and last variants in a chromosomal region that were associated at genome-wide significance with circulating VEGF levels, either in discovery phase (Stage 1) or the combined discovery and replication meta-analysis (Stage 3). The genes of interest were chosen including all within 60kb of each of the candidate regions. A total of 26 genes (listed in the S4 Table) fit this description and served as ‘input’ genes for the pathway analysis. Direct and indirect interactions, a reasonable confidence (experimentally observed, highly predicted, or moderately predicted) and a maximum size of 70 genes/proteins per network were used as parameters in the analysis.
|
10.1371/journal.pntd.0007041 | Mechanisms underpinning the permanent muscle damage induced by snake venom metalloprotease | Snakebite is a major neglected tropical health issue that affects over 5 million people worldwide resulting in around 1.8 million envenomations and 100,000 deaths each year. Snakebite envenomation also causes innumerable morbidities, specifically loss of limbs as a result of excessive tissue/muscle damage. Snake venom metalloproteases (SVMPs) are a predominant component of viper venoms, and are involved in the degradation of basement membrane proteins (particularly collagen) surrounding the tissues around the bite site. Although their collagenolytic properties have been established, the molecular mechanisms through which SVMPs induce permanent muscle damage are poorly understood. Here, we demonstrate the purification and characterisation of an SVMP from a viper (Crotalus atrox) venom. Mass spectrometry analysis confirmed that this protein is most likely to be a group III metalloprotease (showing high similarity to VAP2A) and has been referred to as CAMP (Crotalus atrox metalloprotease). CAMP displays both collagenolytic and fibrinogenolytic activities and inhibits CRP-XL-induced platelet aggregation. To determine its effects on muscle damage, CAMP was administered into the tibialis anterior muscle of mice and its actions were compared with cardiotoxin I (a three-finger toxin) from an elapid snake (Naja pallida) venom. Extensive immunohistochemistry analyses revealed that CAMP significantly damages skeletal muscles by attacking the collagen scaffold and other important basement membrane proteins, and prevents their regeneration through disrupting the functions of satellite cells. In contrast, cardiotoxin I destroys skeletal muscle by damaging the plasma membrane, but does not impact regeneration due to its inability to affect the extracellular matrix. Overall, this study provides novel insights into the mechanisms through which SVMPs induce permanent muscle damage.
| Snakebite is a major neglected tropical disease that affects thousands of people in the rural areas of developing countries. As well as the deaths, snakebites result in a significant number of disabilities including permanent loss of limbs that alter the lifestyle of the victims. Snake venom is a mixture of different proteins with diverse functions; one of these major protein groups present in viper venoms are metalloproteases that primarily induce muscle damage. The mechanisms behind the development of snakebite (metalloprotease)-induced permanent muscle damage are poorly studied. Here, we have purified a metalloprotease (CAMP) from the venom of the Western diamondback rattlesnake, and characterised its function in mice. To determine the actions of CAMP in the development of permanent muscle damage, it was injected into the muscle of mice in a parallel comparison with cardiotoxin I (from the venom of the Red spitting cobra). The effects of these proteins on muscle regeneration were analysed at 5 and 10 days after injection. The results demonstrate that through a combination of effects on the structural scaffolds surrounding the tissues, blood vessels and regeneration, CAMP significantly affects the muscles, thereby leading to permanent muscle damage.
| Snakebite envenomation is a recently reinstated neglected tropical disease [1] that causes around 100,000 deaths annually [2, 3] and innumerable permanent disabilities predominantly on the rural population living in the lower income regions of the world [4–6]. The significant rate of mortality and morbidity occurs due to the difficulties associated with the treatment of snakebites [7], which vary depending on the species [8], geographical location [9], age of the offending snake [10, 11], the quantity of venom injected, correct diagnosis and mode of treatment [12]. Snake venoms contain proteins and small peptides with diverse functional effects [12]. Medically important snakes are generally found in two main families; Elapidae, a family with venoms mainly composed of smaller, neurotoxic proteins such as phospholipase A2 (PLA2) and three finger toxins, and Viperidae, a family with generally larger proteins such as serine and metalloproteases that primarily affect the cardiovascular and musculoskeletal systems. Snake venom serine proteases (SVSPs) mainly cause systemic envenomation effects such as the alteration of blood pressure, activation or inhibition of coagulation factors and degradation of fibrinogen [13, 14]. However, snake venom metalloproteases (SVMPs) primarily induce local envenomation effects such as swelling, necrosis and extensive tissue/muscle damage as well as the activation of certain coagulation factors and degradation of fibrinogen. SVMP-induced muscle damage is often difficult to treat due to the delay in obtaining appropriate medical treatment and poor outcome of anti-snake venom (ASV) treatment in the local tissues [15, 16]. Hence, extensive tissue damage is frequently treated by fasciotomy, a surgical procedure to remove the damaged tissues, cleaning the affected areas followed by skin graft or amputation of affected limbs or fingers when fasciotomy fails to suffice [7]. This results in permanent disabilities for victims, and significantly affects their socio-economic status following snakebites. For example, long term (persisting for over 13 years) musculoskeletal disabilities were found in over 3% of snakebite victims in a rural population of Sri Lanka and of these over 15% had to undergo amputations [17].
Skeletal muscle is composed of myofibres surrounded by the collagen-rich basement membrane. This tissue is imbued with a resident stem cell population called satellite cells (SCs), located under that basement membrane (BM), which are able to regenerate a functional tissue even after extensive damage [18]. The BM plays a key role in muscle repair by orientating the regenerating myofibres, a process mediated by SCs and acting as a scaffold for fibres to grow parallel to the existing fibres [19]. The majority of the direct myotoxic effects of venoms are attributed to PLA2 [20]. They can induce either local or systemic effects depending on their specificity to muscle cells (systemic effects) or a broader range of cells (local effects) through hydrolysis of phospholipids in plasma membrane. Other myotoxic venom components include sodium channel-blocking myotoxins [21] and muscle fibre depolarising cardiotoxins [22]. SVMPs are enzymatic proteins that primarily attack the collagenous structures and various other important components of BM to induce muscle damage. It has recently been reported that SVMPs induce haemorrhage by cleaving components of the BM and extracellular matrix surrounding the smaller blood vessels [23] although as multi-domain proteins, they are capable of binding to and cleaving a range of different proteins [24, 25].
SVMPs are generally classified into four groups based on the additional domains present in their structure: PI/Group I—contains only a metalloprotease domain; PII/Group II—contains a metalloprotease and disintegrin domain, and in some cases the disintegrin domain has been reported to be processed and liberated as a free disintegrin; PIII/Group III—contains a metalloprotease, a disintegrin-like and cysteine-rich domains; PIV/Group IV—contains two lectin-like domains connected by disulphide bonds to the other domains that are found in PIII SVMPs [26]. Although disintegrin-like domains show high sequence identity to disintegrins, they lack the typical RGD motif found in the venom disintegrins, which inhibit platelet aggregation via selectively blocking integrins. Both disintegrin-like and cysteine-rich domains have been found to inhibit collagen-induced platelet aggregation and induce early events of acute inflammation [27]. Notably, disintegrin-like domains were reported to contain an ECD motif that interacts with integrins and block their functions [28]. The non-proteinase domains play key roles in determining the diverse pharmacological effects of PII, PIII and PIV classes of SVMPs including the activation of coagulation factor X [29] and prothrombin [30] amongst others. These domains have also been found to co-localise in muscles, facilitating the hydrolysis of collagen and other BM components by the metalloprotease domain and promoting its accumulation in the BM [24], exerting haemotoxic activities. Moreover, SVMPs are also known to cause ischaemia in the local tissues due to poor blood supply as a result of their haemotoxic effects [31], which may prevent phagocytic removal of necrotic debris and reduce the supply of oxygen and nutrients needed for regeneration [32]. Given the complexity of their actions, a better understanding of the molecular mechanisms through which SVMPs induce permanent muscle damage may pave the way to the development of improved therapeutic strategies for snakebites. In this study, we demonstrate novel insights into the mechanisms by which a PIII/group III metalloprotease isolated from the venom of a North American viper, the western diamondback rattlesnake, Crotalus atrox triggers permanent muscle damage. Our results establish that this SVMP induces muscle damage and also prevents muscle regeneration by acting on the BM, myofibres, blood supply and SCs.
Lyophilised C. atrox venom was purchased from Sigma Aldrich (UK) and the purified Cardiotoxin 1 (CTX), a three-finger toxin from the venom of Naja pallida was obtained from Latoxan (France).
C. atrox venom (10mg) was dissolved in 1mL of 20mM Tris.HCl buffer (pH 7.6) and centrifuged at 5000g for 5 minutes before applying to a pre-made 1mL HiTrap™ Q HP Sepharose anion exchange column. Protein elution was performed at a rate of 1mL/min using 1M NaCl/20mM Tris.HCl gradient (up to 60%) by an ÄKTA purifier system (GE Healthcare, UK) over 20 minutes. The collected fractions were analysed by SDS-PAGE using standard protocols as described previously [33] and fractions with the protein of interest were pooled. The pooled fractions were then concentrated using a Vivaspin centrifugal filter and applied to a gel filtration column (Superdex 75, 1.6cm x 70cm). Protein elution was performed at a rate of 1mL/min using 20mM Tris.HCl (pH 7.6). Following SDS-PAGE analysis, the fractions containing the protein of interest were pooled and concentrated before running through the same gel filtration column again for further purification. Finally, the fractions containing the pure protein were pooled, concentrated and stored at -80°C until further use. Protein estimation was performed using Coomassie plus protein assay reagent (ThermoFisher Scientific, UK) and bovine serum albumin as standards.
The purified protein was subjected to SDS-PAGE, and a gel section containing the pure protein was subjected to tryptic digestion and analysed by mass spectrometry at AltaBioscience (Birmingham, UK). The extracted protein (10μg) from the gel slice was added to 100mM ammonium bicarbonate (pH 8). This was then incubated with dithiothreitol (10mM) at 56°C for 30 minutes. After cooling to room temperature, the cysteine residues were alkylated using iodoacetamide (50mM). Trypsin gold (Promega, UK) was subsequently added and the samples were incubated overnight at 37°C. The digested peptides were concentrated and separated using an Ultimate 3000 HPLC series (Dionex, USA). Samples were then trapped on an Acclaim PepMap 100 C18 LC column, 5um, 100A 300um i.d. x 5mm (Dionex, USA), then further separated in Nano Series Standard Columns 75μm i.d. x 15 cm. This was packed with C18 PepMap100 (Dionex, USA) and a gradient from 3.2% - 44% (v/v) solvent B (0.1% formic acid in acetonitrile) over 30 minutes was used to separate the peptides. The digested peptides were eluted (300nL/min) using a triversa nanomate nanospray source (Advion Biosciences, USA) into a LTQ Orbitrap Elite Mass Spectrometer (ThermoFisher Scientific, Germany). The MS and MS/MS data were then searched against Uniprot using Sequest algorithm and the partial sequence was then compared to the other similar protein sequences available in the protein database.
Human plasma fibrinogen (100μg/mL) was incubated with different concentrations of the whole venom or the purified protein, and a small volume of digested samples were removed at 30, 60, 90 and 120 minutes and mixed with reducing sample treatment buffer [4% (w/v) SDS, 10% (v/v) β-mercaptoethanol, 20% (v/v) Glycerol and 50mM Tris.HCl, pH 6.8]. The samples were then analysed by 10% SDS-PAGE and stained with Coomassie brilliant blue to determine the fibrinogenolytic activity of venom and the purified protein.
The metalloprotease activity of both C. atrox whole venom and the purified protein was assessed using a fluorogenic substrate, DQ-gelatin (ThermoFisher Scientific, UK). Briefly, the whole venom or purified protein (10μg/mL) was mixed in phosphate buffered saline (PBS, pH 7.4) with DQ gelatin (10μg/mL). The reaction mix was incubated at 37°C and the level of fluorescence was measured at 60 minutes using an excitation wavelength of 485nm and emission wavelength of 520nm by spectrofluorimetry (FLUOstar OPTIMA, Germany).
Similarly, the serine protease activity was measured using a selective substrate, Nα-Benzoyl-L-Arginine-7-Amido-4-methylcoumarin hydrochloride (BAAMC) (Sigma Aldrich, UK). The whole venom or the purified protein (10μg/mL) was incubated with BAAMC (2μM) at 37°C and the level of fluorescence was measured at an excitation wavelength of 380nm and emission wavelength of 440nm by spectrofluorimetry.
The University of Reading Research Ethics Committee has approved the procedures for blood collection from healthy human volunteers and the consent forms used to obtain written consent. Experiments with mice were performed in line with the principles and guidelines of the British Home Office and the Animals (Scientific Procedures) Act 1986 (PPL70/7516). All the procedures were approved by the University Research Ethics Committee (License number: UREC 17/17).
Human blood was obtained from healthy volunteers in vacutainers with 3.2% (w/v) sodium citrate as an anti-coagulant and the platelet-rich plasma (PRP) was prepared as described previously [34–36]. Platelet aggregation assays were performed by optical aggregometry using 0.5μg/mL cross-linked collagen related peptide (CRP-XL) as an agonist in the presence and absence of different concentrations of the purified protein.
The C57BL/6 mice (8 weeks old) were obtained from Envigo, UK. Mice were anaesthetised with 3.5% (v/v) isofluorane in oxygen before maintaining at 2% for the procedure. They were then injected intramuscularly with 30μL of either PBS (undamaged control), 50μM CTX, and 8 or 16 μM of the purified protein into their right tibialis anterior muscle. Mice were then allowed to recuperate for either 5 or 10 days before sacrificing by carbon dioxide asphyxiation and cervical dislocation.
The TA muscles from mice were dissected, weighed and frozen on liquid nitrogen cooled iso-pentane prior to storage at -80°C. The EDL muscle was dissected from the undamaged contralateral hind limb of experimental mice and immediately placed in a 2mg/mL collagenase solution (Sigma Aldrich, UK) and incubated at 37°C with 5% CO2 for 2 hours to isolate the single fibres as previously described [37].
To determine the proliferation of SCs and myogenic differentiation, isolated single fibres were cultured for up to 48 hours at 37°C with 5% CO2 in single fibre culture medium (SFCM—DMEM, 10% (v/v) horse serum, 1% (v/v) penicillin-streptomycin and 0.5% (v/v) chick embryo extract) supplemented with either 0, 0.3, 1 or 3 μM of the purified protein prior to fixation in 2% (w/v) paraformaldehyde in PBS and maintained in PBS prior to immunocytochemistry.
The migration of muscle fibre SCs was analysed as described previously [37]. Briefly, the isolated single fibres were cultured for 24 hours in SFCM before transferring to SFCM containing 0, 0.3, 1 or 3μM of the purified protein and monitoring by a phase contrast microscope at 37°C with 5% CO2 using a 10X objective. A time-lapse video was captured at a rate of 1 frame every 15 minutes for a 24-hour period and analysed to determine the rate of migration.
The collected TA muscles were mounted in Tissue-TEK® OCT compound in an orientation allowing the transverse sections of 13μm thickness to be obtained using a cryo microtome. The tissue sections were incubated in permeabilisation buffer [20mM HEPES, 3mM MgCl2, 50mM NaCl, 0.05% (w/v) sodium azide, 300mM sucrose and 0.5% (v/v) Triton X-100] for 15 minutes at room temperature. To remove the excess permeabilisation buffer, 3 x 5 minute washes were performed using PBS before the application of wash buffer [PBS with 5% fetal bovine serum (v/v), 0.05% (v/v) Triton X-100] for 30 minutes at room temperature.
Primary antibodies were pre-blocked in wash buffer for 30 minutes prior to application onto muscle sections overnight at 4°C. In order to remove the primary antibodies, muscle sections were washed three times (5 minutes each) in wash buffer. The sections were then incubated with species-specific secondary antibodies that were conjugated with Alexa Fluor 488 or 594. The secondary antibodies were pre-blocked in wash buffer for minimum of 30 minutes before their application onto the slides and incubated for 1 hour in the dark at room temperature. Thereafter, the muscle sections were washed 3 x 5 minutes in PBS to remove the unbound secondary antibodies. Finally, the slides were mounted in fluorescent mounting medium, and the myonuclei were visualised using 4, 6-diamidino-2-phenylindole (DAPI) (2.5μg/mL). The images of sections were obtained using a fluorescence microscope (Zeiss AxioImager) and analysed using ImageJ. Macrophages were detected by F4.80 staining using the Vector Laboratories ImmPRESS Excel Staining Kit. A list of antibodies used in this study is provided in S1 Table.
All the statistical analyses were performed using GraphPad Prism 7 and the P-values were calculated using one-way ANOVA followed by Dunnett’s post hoc multiple comparisons test.
In order to purify a protein with a molecular weight of around 50kDa (as predicted for group III SVMPs) from the venom of C. atrox, a two-dimensional chromatography approach was employed. Following the initial fractionation of venom via anion exchange chromatography (Fig 1A and 1B), the selected fractions (14–18) with a highly abundant protein at approximately 50kDa were pooled and run through a gel filtration chromatography column (Fig 1C and 1D). The fractions (62–67) were pooled and run through the same gel filtration column again to refine the purification (Fig 1E and 1F). Finally, a pure protein with a molecular weight of around 50kDa was isolated. Mass spectrometry characterisation of the tryptic digested peptides of this protein and further Mascot analysis confirmed it to be a similar or identical protein to vascular apoptosis inducing proteins (VAP) such as VAP2, a protein with a molecular weight of 55kDa (an identical molecular weight to the purified protein) [38], which is a group III metalloprotease (Fig 1G). The identified peptide sequences of the purified protein covered around 43% of the sequence of VAP2A (highlighted in red in Fig 1G). The purified protein has been referred to as ‘CAMP’ to denote C. atrox metalloprotease throughout this article.
By using fluorogenic substrates, the protease activity of CAMP was analysed in comparison to the whole venom. CAMP displayed no serine protease activity as it failed to cleave a serine protease selective fluorogenic substrate, BAAMC although the whole venom displayed significant serine protease activity (Fig 2A). However, it showed high levels (similar to the whole venom) of collagenolytic activity (Fig 2B). Furthermore, the ability of CAMP to digest fibrinogen was analysed by incubating it with human plasma fibrinogen. The SDS-PAGE analysis of samples that were taken at different points of incubation confirmed that CAMP is capable of cleaving Aα and Bβ chains of fibrinogen although it was unable to cleave the γ chain (Fig 2C). The digestion of fibrinogen with CAMP appears to be rapid as the levels of Aα and Bβ chains of fibrinogen were reduced significantly as early as 30 minutes of incubation. These results corroborate CAMP as an SVMP with collagenolytic and fibrinogenolytic activities, which may affect the collagen in the BM around the local tissues at the bite site and fibrinogen in the blood.
The ability of CAMP to inhibit agonist-induced platelet activation was analysed using human platelet-rich plasma (PRP) by optical aggregometry. The pre-treatment of human platelets (PRP) with CAMP (50μg/mL) has significantly inhibited 0.5μg/mL CRP-XL-induced platelet aggregation (Fig 2D and 2E). This data confirms the ability of CAMP to affect human platelet activation.
In order to determine the mechanisms through which SVMPs induce permanent muscle damage, CAMP was used as a tool to determine its pathological effects in TA muscle of mice in comparison with CTX. The intramuscular injection of CAMP induced haemorrhage in the damaged muscles and thereby, caused swelling and increase in muscle weight after five days of administration (Fig 3A and 3B). However, CTX did not induce haemorrhage or swelling although the muscle weight was reduced compared to the controls at the same time point. In contrast, after ten days of administration, muscle weight in CAMP-treated mice was decreased similar to CTX-treated muscle (Fig 3A and 3C). These data demonstrate that CAMP is capable of inducing haemorrhage and swelling and thereby, increases in muscle weight initially although it decreases at a later time point.
We examined the cellular processes underpinning the morphology of skeletal muscle and assessed muscle regeneration after damage induced by CAMP. Haematoxylin (H) and eosin (E) staining facilitates the identification of cellular organisation within a tissue and also the presence of fibres containing centrally located nuclei (CLN), which is an indicator of muscle regeneration. Five days after tissue damage, muscles treated with CTX contained many large fibres with CLN (Fig 4A). Furthermore, there were regions of high cell density between fibres displaying CLN. In contrast, 5 days after CAMP damage the number of fibres with CLN was less abundant and smaller than in CTX damaged muscle (Fig 4A and 4B). Additionally, there were areas of sparsely populated regions between fibres. Ten days after CTX damage, large fibres with CLN were evident with very little space between muscle fibres (Fig 4C and 4D). The fibres appeared to be regular in terms of shape and size, evidencing robust muscle regeneration. Whereas, at the same time point, muscle damaged with CAMP displayed smaller fibres with CLN and inter-fibre regions populated with cells were prominent (Fig 4C and 4D). Next, we documented the profile of dying muscle fibres, facilitating the infiltration of circulating immunoglobulins (Ig) into the damaged fibres. Five days after CTX injection, low density of small calibre fibres displayed the infiltration by IgG (Fig 4E–4G). In contrast, at the same time point, CAMP treatment resulted in not only a higher density of fibres with infiltrated IgG, but they were also of larger size (Fig 4E–4G). By day 10, very few dying fibres were present in CTX treated muscle, however dying fibres were prominent in CAMP treated muscles (Fig 4H and 4I). We then examined the presence of regenerating muscle fibres, facilitated through the expression of embryonic myosin heavy chain protein (MYH3). Muscle regeneration was clearly evident in muscles damaged by CTX at day 5 (Fig 4J and 4K). Large numbers of evenly sized fibres expressing MYH3 featured in CTX-damaged tissue (Fig 4J and 4L). In contrast to CTX treatment, the number of regenerating fibres in CAMP-treated muscle was lower and when present were of heterogeneous size (Fig 4J–4L). By Day 10, the expression of MYH3 has been cleared in CTX-damaged muscle and when present was in very large fibres (Fig 4M–4O). In contrast, MYH3 expression was clearly evident at day 10 in CAMP-damaged muscle but in smaller, non-uniform fibres (Fig 4M–4O). Next, we examined the impact of CAMP and CTX on blood vessels through immunostaining with the endothelial cell specific antibody, CD31. At both 5 and 10 days, the number of capillaries serving each regenerating fibre was greater in the CTX treated sample compared to CAMP (Fig 4P–4S). Importantly, the number of capillaries serving each regenerating fibre in the CTX treated sample was identical to the undamaged sample. Moreover, the degree of macrophage infiltration into the damaged area was analysed, as these cells are key to effective muscle regeneration. The density of macrophages in damaged muscle was greater in the CTX treated muscle compared to CAMP at day 5 (Fig 4T and 4U). However, by day 10, the situation was reversed; there was a greater density of macrophages in the CAMP treated samples compared to CTX (Fig 4V and 4W).
Efficient regeneration of skeletal muscle following acute damage is contingent on stem cells capable of replacing damaged tissue and their highly ordered formation into myotubes/fibres, a process orchestrated by the ECM. The organisation of collagen IV, a major BM component of muscle fibres was analysed as described previously [39]. A thin circle of collagen IV surrounding muscle fibre was evident 5 days after CTX treatment (Fig 5A). In contrast, at an identical time after CAMP treatment, muscle displayed large irregular, thick depositions of collagen IV (Fig 5A and 5B) but by day 10, the picture was even more polarised, as CTX-damaged muscle showed a relatively normal distribution of collagen IV (Fig 5C and 5D). In contrast, very few fibres from CAMP-treated muscle (at day 10) displayed a ring of collagen IV, and instead this protein was localised at thick foci (Fig 5C and 5D). A near-identical pattern was documented for the distribution of laminin, another major component of the muscle fibre ECM (Fig 5E–5H).
Furthermore, the impact of CTX and CAMP on molecules that are associated with linking the contractile apparatus to the ECM was investigated. Dystrophin is normally localised under the sarcolemma of mature muscle fibres. Its expression was evident around some of the larger regenerating muscle fibres 5 days after CTX damage (Fig 5I). Whereas, very few fibres expressing dystrophin were detected at a similar time point in CAMP-treated muscles (Fig 5I). However, when present, the thickness of the dystrophin expression domain was similarly reduced by the two treatments (Fig 5J). At day 10, most of the fibres from CTX-treated muscle displayed a continuum of dystrophin expression, although at a lower thickness compared to undamaged tissue (Fig 5K and 5L). However, very few fibres with a ring of dystrophin were present in CAMP-treated muscles at day 10 (Fig 5K). Furthermore, the domain, when present was thinner than both control as well as CTX treated muscles (Fig 5L). Then, the distribution of nNOS, a protein that localises to a sub-sacrolemmal position which is dependent on its binding to dystrophin was assessed. At 5 days after treatment, very little nNOS was present in either CTX or CAMP-damaged muscles (Fig 5M and 5N). By day 10, a thin band of nNOS was evident in CTX-treated muscle but not in the muscle damaged by CAMP (Fig 5O and 5P). Lastly, the muscles were analysed to determine the presence of remaining CAMP in damaged tissues. The immunohistological profiling showed that CAMP was clearly present at both 5 and 10 days after its administration (Fig 5Q and 5R). These results show that CAMP treatment damages not only the ECM of muscle fibres but also affects intracellular components that link it to the contractile machinery.
The role of SCs adjacent to the muscle fibres is critical for muscle regeneration. In order to determine the impact of CAMP on SCs, we have isolated myofibres from intact EDL muscles and exposed them to a range of concentrations of CAMP. As early as 24 hours after CAMP treatment, it was evident that there was a concentration dependent disturbance to the collagen component of the ECM around muscle fibres. A uniform layer of collagen expression was detected in untreated fibres (Fig 6A). At the lower concentration, CAMP caused a localised denuding of the myofibre (Fig 6B), whereas the higher concentration resulted in the absence of collagen from most parts of the fibres and caused it to concentrate in specific locations (Fig 6C). The cell growth, proliferation and migration were monitored on the isolated single muscle myofibres over a 48 hour time period. SCs were immunostained using the myogenic transcription factors, Pax7 (uncommitted cells) and MyoD (activated cells) in order to monitor the progression of cells through myogenesis. The concentrations of above 0.3μM of CAMP induced hypercontraction, which is indicative of extensive fibre damage. At 0.3μM, viable fibres were present, and revealed that CAMP significantly decreased the number of associated SCs (Fig 6D). Furthermore, the number of SC clusters was reduced per fibre (Fig 6E), although each cluster at the lower concentration had more cells than untreated fibres (Fig 6F). Analysis of differentiation was only possible at the lowest concentration of CAMP (Fig 6G) as at higher concentrations, hypercontraction prevented this analysis. The migration speed was calculated between 24 and 48 hours and was found to decrease significantly as the concentration of CAMP increased (Fig 6H). These data demonstrate that CAMP is able to affect both the proliferation and migration of SCs but not the differentiation.
The swelling and necrosis at the bite site as well as permanent muscle damage are common effects of snakebite envenomation (particularly viper bites). These effects frequently lead to amputation and therefore disable victims, which adds to their inability to earn money, and exacerbates the poverty that is already experienced by the vast majority of snakebite victims [4]. Here we have purified a metalloprotease from one of the most studied venomous snake species, C. atrox. Although deaths from this snake are now uncommon, disfigurement is still a prevalent side effect for survivors. SVMPs are a predominant component in viper venoms that are involved in inducing the local envenomation effects including muscle damage. The ability of SVMPs to degrade collagen has been established, but its impact on permanent muscle damage under in vivo settings has not been previously demonstrated in sufficient detail. Therefore, we deployed a metalloprotease from the venom of C. atrox and analysed its impact on skeletal muscle damage in comparison to a three-finger toxin, CTX from the venom of Naja pallida. Mass spectrometry analysis of the purified protein suggests it to be a group III metalloprotease, which possess a metalloprotease domain as well as a disintegrin-like and cysteine-rich domains [40]. Based on the peptide sequences identified by the mass spectrometry, the purified protein is likely to be VAP2 or one of its heterodimers; VAP2A or VAP2B, both of which are vascular apoptosis inducing proteins (38) that are known to be haemorrhagic [41] and in the case of VAP2B to inhibit collagen-induced platelet activation [42]. Due to the limited peptide sequences identified by mass spectrometry for the purified protein, we are unable to conclude whether the purified protein (CAMP) is identical to VAP2 or either of its heterodimers. CAMP was characterised to be a collagenolytic and fibrinogenolytic enzyme. It also inhibited CRP-XL-induced platelet aggregation; group III metalloproteases are known to interact with the integrin α2β1, binding to the α2 subunit and causing the shedding of β1 subunits [43]. However, VAP2B (a protein described from C. atrox) has been reported to inhibit collagen induced platelet aggregation by binding to collagen [44], although whether the SECD sequence found in disintegrin-like domains is able to bind CRP-XL in the same way, is unknown.
In order to determine the impact of SVMPs in stimulating permanent muscle damage, different concentrations of CAMP were administered in mice along with CTX and control groups and the effects were analysed at five and ten days after the administration. We suggest that this occurs at two levels; by breaking down the ECM which normally acts as a scaffold for the formation of new muscle fibres and around existing blood vessels and secondly attenuating properties of resident stem cells that are essential to effective tissue repair. SVMPs are known for their collagenolytic activities and for targeting various components of the BM in the vasculature and inducing haemorrhage [45]. In line with previous studies, here we demonstrate that CAMP induces haemorrhage and affects the architecture of collagen and laminin. The destruction of the collagen based ECM may be the key to long-term tissue destruction wrought by CAMP. The disintegrin-like and cysteine-rich domains have already been identified as essential to the haemorrhagic activity and ECM degradation attributed to the PIII metalloproteases [46]. This is in contrast to myotoxic PLA2 and three-finger toxins that are well documented in causing membrane permeabilisation and consequentially myonecrosis via the hydrolysis of membrane phospholipids or imbedding directly into the membrane respectively [47–49]. Our data emphasise that CAMP in comparison to CTX significantly hindered the regeneration of skeletal muscle fibres most probably by disturbing the organisation of the ECM. The elevated levels of necrosis seen five days after administration with CAMP improved after ten days, although it was still evident. However, it must be noted that at day 10, the CTX treated muscles had almost completely regenerated with healthy fibres. Moreover, very low levels of MYH3 were detected five days after CAMP treatment. In CTX-treated muscles, this marker of regenerating fibres was clearly evident and present at a high level. This indicates that the initiation of the regeneration process was attenuated by CAMP in comparison to CTX. Muscle regeneration is dependent on blood supply and clearance of damaged fibres. We show here that both these cellular compartments are affected in a detrimental manner by CAMP. We found that the number of capillaries serving each regenerating muscle fibre was smaller in CAMP treated muscle compared to CTX. Importantly the number of capillaries serving each fibre in damaged CTX muscle was the same as in undamaged regions. These results show that capillaries as well as muscle fibres are damaged by CAMP whereas it is only the latter in CTX treated tissue. Additionally, we show that there was a greater influx of macrophages into the CTX damaged muscle compared to regions affected by CAMP. Furthermore, the density of macrophages decreased in CTX treated muscle over time, attesting to regeneration. In contrast, the density of macrophages in CAMP treated muscle was lower at day 5 compared to CTX, possibly indicating an attenuated clearance process. Importantly the density of macrophages did not change in the CAMP treated muscle over 10 days suggesting on-going muscle damage. Although the abundance of MYH3 increased in CAMP-treated muscles by day 10, its expression in CTX-injured muscles was almost undetectable, signifying advanced regeneration. This was reflected in the appearance of dystrophin and nNOS at their normal sub-sarcolemmal position. In keeping with the notion that CAMP treatment not only affects the degree of regeneration but also its timing, we showed that very few fibres expressed dystrophin in its normal position and a significantly reduced expression of nNOS was observed even at day 10. Most importantly we show that CAMP is still present at the site of injury even 10 days after its administration and that it profoundly disorganises the ECM.
The single fibre experiments highlight another aspect to explain the attenuated muscle regeneration following CAMP-mediated muscle damage. We demonstrate that the proliferation and migration of SCs was significantly reduced by CAMP treatment. Both of these factors are key in promoting muscle regeneration. Therefore, CAMP may bring about permanent impairment of muscle organisation and function by firstly destroying muscle fibres, secondly breaking down the organisation of the ECM. This is required by the SCs in order to align and fuse in a coordinated manner and lastly by diminishing the ability for SCs to expand their numbers and migrate to the site of injury to enact efficient regeneration. It is clear that current ASV treatment is not effective at preventing muscle damage. Although translating the results of this study into therapeutics might be difficult, these will improve the understanding of SVMP-induced permanent muscle damage. The ability of group III metalloproteases to bind components of the BM and prolong muscle exposure to their myotoxic effects suggests a therapeutic agent that is capable of interacting with these enzymes and non-enzymatic domains and preventing the longevity of these proteins in the area surrounding fibres may be able to speed up the rate of regeneration considerably. Moreover, any drugs aimed at treating this aspect of snakebite envenomation may struggle to reach it intravenously, and therefore, they may have to be administered via multiple local injections considering the widespread damage to microvasculature [32] and consequential lack of blood supply to affected tissues.
ASV is the only effective treatment for systemic envenoming, however local venom pathology is largely unaffected by ASV when treatment is not immediately administered [50]. ASV is composed of large immunoglobulins that appear to struggle to reach the areas affected by SVMPs. The combination of small vessel destruction combined with BM cleavage results in a poor blood supply and therefore weak neutralisation by intravenously administered ASV. Local injections of ASV have also been found to be of no benefit to the snakebite victims [51]. However, there are a range of matrix metalloprotease inhibitors that have undergone testing for their specificity to SVMPs and some promising compounds have been identified [52]. The small molecule inhibitors aimed at the metalloprotease domain such as batimastat [53] have been tested extensively and they were found to abrogate the haemorrhagic effects of venom if administered immediately after envenoming. Given their haemorrhagic effects are largely dependent upon collagen degradation, it is reasonable to postulate that this prevention of haemorrhagic effects may also apply to muscle damage. Moreover, metal chelating agents such as EDTA have also been tested in vivo at non-toxic doses and found to prevent venom-induced lethality [54].
The need for immediate administration is of course unrealistic with conventional ASV but small stable inhibitors have the potential to be spread and made available to those in areas with a high density of snakebites. Multiple local injections do bring the potential for delivery directly to the bite site and administering to multiple sites may overcome the problematic spread of drug through a site of damaged muscles and vessels. Future experiments should aim to investigate the effect of these drugs on BM components, using both pre-incubation with drugs and post envenomation delivery models. Overall, the complete destruction or loss of a range of BM and dystrophin-glycoprotein complex components as well as the effect of this SVMP on muscle regeneration highlights the significant difficulties involved in treating the necrosis and muscle damage associated with snakebite envenomation. Hence, this study provides greater insights into the understanding of SVMP-induced permanent muscle damage and local snakebite envenomation effects.
|
10.1371/journal.pgen.1003411 | The Role of the Arabidopsis Exosome in siRNA–Independent Silencing of Heterochromatic Loci | The exosome functions throughout eukaryotic RNA metabolism and has a prominent role in gene silencing in yeast. In Arabidopsis, exosome regulates expression of a “hidden” transcriptome layer from centromeric, pericentromeric, and other heterochromatic loci that are also controlled by small (sm)RNA-based de novo DNA methylation (RdDM). However, the relationship between exosome and smRNAs in gene silencing in Arabidopsis remains unexplored. To investigate whether exosome interacts with RdDM, we profiled Arabidopsis smRNAs by deep sequencing in exosome and RdDM mutants and also analyzed RdDM-controlled loci. We found that exosome loss had a very minor effect on global smRNA populations, suggesting that, in contrast to fission yeast, in Arabidopsis the exosome does not control the spurious entry of RNAs into smRNA pathways. Exosome defects resulted in decreased histone H3K9 dimethylation at RdDM-controlled loci, without affecting smRNAs or DNA methylation. Exosome also exhibits a strong genetic interaction with RNA Pol V, but not Pol IV, and physically associates with transcripts produced from the scaffold RNAs generating region. We also show that two Arabidopsis rrp6 homologues act in gene silencing. Our data suggest that Arabidopsis exosome may act in parallel with RdDM in gene silencing, by epigenetic effects on chromatin structure, not through siRNAs or DNA methylation.
| To maintain genomic stability and prevent expansion of invasive genomic sequences such as transposable elements (TEs), eukaryotes have evolved defensive mechanisms to control them. Here, we examine the role of the Arabidopsis exosome complex in such mechanisms. Evolutionarily conserved from archaea to humans, the exosome is a stable complex of RNase-like and RNA binding proteins that plays a central role in RNA metabolism in eukaryotes. Depletion of the exosome allows some repetitive sequences to escape from silencing. Most of these transcripts emanate from centromeric and pericentromeric chromosomal regions and other heterochromatic loci, and many derive from repetitive and transposable elements. In plants, TEs are targeted for de novo DNA methylation by smRNA–mediated pathways. However, we found that exosome depletion has only minor effects on smRNA populations that are acting in the main silencing mechanism in Arabidopsis, siRNAs–dependent DNA methylation RdDM. Instead, exosome depletion affects histone H3K9 dimethylation, an epigenetic mark that affects chromatin structure and thus alters transcription. Our data suggest that the exosome collaborates in gene silencing, likely acting in a parallel pathway to other mechanisms. We also propose that the Arabidopsis exosome may coordinate the transcriptional interplay of different RNA polymerases to modulate repression of some repetitive sequences.
| High-throughput analyses have revealed that eukaryotic genomes are pervasively transcribed [1]–[4], and the majority of the transcriptional activity takes place outside of protein-coding genes, producing non-coding RNAs (ncRNAs) derived from genome regions once thought to be transcriptionally silent, including intergenic and heterochromatic regions [1]–[3], [5]. Pervasive transcription constitutes a risk for the cell, as it can be associated with expansion of TEs, loss of genomic stability and defects in gene expression. However, recent studies have also shown that ncRNAs themselves can have important regulatory functions, including the establishment and maintenance of the epigenetic architecture of eukaryotic genomes. In some cases, long ncRNAs serve directly as molecular scaffolds for recruiting chromatin modifiers [6], [7], whereas in other cases ncRNAs are processed by the RNAi machinery into short interfering siRNAs that guide DNA methylation and chromatin modifications to homologous regions of the genome [8], [9]. Thus, RNA-mediated heterochromatin formation requires an affected region to be transcribed for transcriptional silencing to occur. Many of the ncRNA transcripts earned the term “hidden” because they remain invisible unless RNA degradation is prevented, for example, by inactivation of the degradation machinery [1], [3], [4], [10]–[14], raising the important question of how these ncRNAs are regulated.
The exosome complex plays a central role in RNA metabolism in eukaryotes. Evolutionarily conserved from archaea to humans, the exosome is a stable complex of RNase-like and RNA binding proteins that catalyzes 3′ to 5′ processing and decay of various RNA substrates [15]. The current view of eukaryotic exosome structure is based mostly on studies done in yeast and human. The eukaryotic exosome has nuclear and cytoplasmic forms that share ten components. The key structural feature is a nine-subunit donut-shaped structure called the exosome ring. Six of the subunits, RNase PH domain-containing proteins Rrp41, Rrp42, Rrp43, Rrp45, Rrp46 and Mtr3, are organized into a hexameric ring, capped on one side by a trimer of subunits that contain S1 and KH RNA binding domains (Rrp40, Rrp4 and Csl4) [16], [17]. The 9-subunit ring structure has no catalytic activity in yeast and human, due to amino acid replacements that disable binding of RNA, phosphate ion, or catalysis [16], [17]. The exosome active sites are contributed by the tenth protein, Rrp44 (Dis3), which has endonucleolytic and exonucleolytic activities and considered to be the tenth subunit of the exosome core [18], [19]. In addition to Rrp44, the nuclear form of the eukaryotic exosome is associated with a second active 3′ to 5′ exonuclease, Rrp6 [20], [21]. Most functions of the exosome are dependent on cofactors. One of the notable complexes associated with the nuclear exosome is the Trf-Air-Mtr4 polyadenylation (TRAMP) complex endowed with a poly(A) polymerase activity that stimulates degradation [22]–[24]. The plant exosome might differ from yeast and human exosomes, as its ring subunit Atrrp41p appears to retain an active site and was also shown to have catalytic activity in vitro [1], [25]. Our previous genome-wide study using tiling microarrays to examine exosome targets in Arabidopsis revealed that a large number of exosome substrates correspond to ncRNAs originated from promoters, 5′UTRs, intergenic regions, repetitive elements and TEs [1]. Many of these ncRNAs derive from centromeric and pericentromeric regions and other heterochromatic loci known to give rise to smRNAs that participate in silencing of these loci [26]. In Arabidopsis, the main and most-studied pathway for transcriptional gene silencing of repetitive elements and transposons is the siRNA-based silencing mechanism known as RNA-dependent DNA methylation (RdDM) [9], [27]–[29]. The effects of exosome depletion on these ncRNAs and, potentially, on smRNAs are unlikely to be attributable to indirect effects of exosome depletion on the expression of RdDM pathway components, since no genes acting in siRNA biogenesis, siRNA-mediated transcriptional gene silencing (TGS), DNA methylation or demethylation, or histone H3K9 modifications were found to be affected in these lines [1].
RdDM induces de novo methylation of cytosines in all sequence contexts at the region of siRNA–DNA or siRNA-RNA sequence homology. This silencing pathway requires two plant-specific RNA polymerases, Pol IV and Pol V, specializing in transcriptional gene silencing (TGS) [28], although transcriptional activity of Arabidopsis Pol II was also reported to be involved in siRNA-directed gene silencing [30]. The mechanistic details of RNA-dependent silencing are not fully understood and also appear to vary from one genomic location to another, but the RdDM pathway likely consists of three main steps: (i) siRNA production from transcripts that are likely transcribed by RNA Pol IV [9], (ii) synthesis of non-coding RNAs that could serve as scaffolds by RNA Pol V and/or Pol II at some of the loci [30], [31], and (iii) assembly of AGO-siRNA effector complexes to recruit methylation machinery to complementary sequences [9]. In siRNA biogenesis, RNA Pol IV transcripts are made double-stranded by RNA-DEPENDENT RNA POLYMERASE 2 (RDR2), processed into 24 nt siRNA by DICER-LIKE 3 (DCL3), and then incorporated into ARGONAUTE (AGO4 and possibly AGO6) to form an AGO-siRNA complex [32]–[35]. The AGO-siRNA complex and other RdDM effectors [31], [35]–[37], assemble on scaffold RNA to form a guiding complex that recruits DNA methyltransferases and histone methyltransferases to direct the silencing of specific genomic loci through a mechanism that is not fully understood. Pol IV is thought to initiate RdDM pathway, whereas Pol V and AGO4-associated siRNAs function downstream from Pol IV to promote DNA methylation by recruiting the silencing complex to targeted loci. RNA Pol IV, Pol V and Pol II activities in RdDM are functionally diversified and coordinated; however, it is not yet clear how they are functionally integrated in heterochromatin silencing.
The model system in which siRNA-mediated silencing is the best understood mechanistically is fission yeast. In S. pombe RNA Pol II carries out the functions attributed to Pol IV and Pol V in plants, therefore, it generates both siRNA precursors and scaffold transcripts to which siRNAs bind at loci that are subject to siRNA-mediated silencing. Exosome defects in S. pombe were reported to result in the loss of transcriptional silencing from centromeric, silent mating type, and telomeric loci [38]–[40]. In S. pombe, in the absence of exosome-mediated degradation, abundant aberrant RNA species enter the RNAi pathway and interfere with heterochromatic silencing through competition for RNAi biogenesis machinery, resulting in the dramatic decrease in centromeric siRNAs [38]–[40]. Recently, it was also shown that exosome plays an important role in remodeling of facultative heterochromatin [41]. Earlier work in plants also suggested that aberrant RNAs could enter RNAi pathways unless they are degraded by the 5′ to 3′ pathway [42]. However, the role of the exosome complex in smRNA metabolism in Arabidopsis has not been examined. It is also not known whether the Arabidopsis exosome complex interacts with the RdDM silencing pathway.
To answer these questions we employed next-generation sequencing to profile populations of smRNAs in exosome-depleted plants, and in mutants of RdDM pathway genes. Unexpectedly, we found that loss of the exosome subunits had little effect on the global populations of smRNAs and had no affect on the level of DNA methylation in examined RdDM loci; rather, it resulted in a reduction of histone H3K9 dimethylation. We propose that the Arabidopsis exosome may coordinate the transcriptional interplay of RNA polymerases Pol II, Pol V and Pol IV, to achieve the appropriate level of transcriptional repression of heterochromatic loci.
Previously, we found that the majority of transcripts upregulated in RRP4 and RRP41 exosome depletion mutants originate from the promoters, repeats, intergenic, and siRNA generating regions [1]. Most of these regions harbor repeats and TEs that are known to be silenced by RdDM through siRNAs.
Since microarray experiments allow estimation of only the length of affected regions, but not the exact length of affected transcripts, we set out to examine whether the exosome is involved in down regulation of these regions through regulating either quantity or quality of smRNAs. To characterize any changes in smRNA populations that occur in response to exosome depletion, we employed next-generation sequencing to deep sequence the smRNA populations in depletion mutants of exosome subunits RRP4 and RRP41. Null T-DNA insertion mutations in RRP4 and RRP41 are lethal; therefore, we used inducible RNA-interference (iRNAi) constructs to reduce RRP4 and RRP41. The seedlings of RRP4 (rrp4-i) or RRP41 (rrp41-i) transgenic plants grown on estradiol-containing medium to induce the RNAi constructs subsequently exhibit a growth arrest ([1], Figure 1A). We selected the earliest time-point of estradiol treatment corresponding to the accumulation of underprocessed 5.8S rRNA species (the hallmark of the exosome defect), but before growth retardation, to minimize changes in gene expression that did not result directly from exosome depletion [1]. Small RNA libraries for Illumina sequencing were generated from the seedlings of rrp4-i and rrp41-i iRNAi lines grown with and without estradiol (Table S1) and smRNAs between 15- and 32 nt in length were selected and mapped to the Arabidopsis genome (TAIR version 9).
We first examined the smRNAs from the iRNAi transgenes used for inactivation of RRP4 or RRP41 [1]. As expected, these silencing cassettes generate silencer sequences corresponding to RRP4 or RRP41 (mapping to AT1G03360 and AT3G61620 loci). Profiling silencer sequences by size and by first nucleotide revealed that the majority of the silencer sequences are 21, 22 and 24 nt and start with 5′U or 5′A (Figure S1), suggesting that they are preferentially loaded into Ago1, Ago2 and Ago4 complexes [43] to silence their target. Silencer sequences produced from iRNAi transgenes were filtered out and libraries without silencer reads were termed FLR, for filtered reads (Table S1). Each library was normalized either to the total number of mapped non-redundant reads or to the total number of non-redundant filtered reads (FLR), multiplied by 106 (RPM, reads per million). Both methods of normalization produced similar results; therefore, only data normalized using filtered reads (FLR) are presented graphically in this study.
We then classified smRNAs based on their size, the nature of their first nucleotide, and their genomic features. The majority of functional smRNAs in A. thaliana range from 21 to 24 nt. Our libraries were constructed using 15–32 nt smRNAs; therefore, we were able to detect any effect exosome depletion might have on smRNA metabolism. We found that exosome defect does not lead to changes in smRNAs in the 15–19 nt and 26–32 nt categories (data not shown). Importantly, the group of 20–25 nt smRNAs, which contains the majority of functional smRNAs, was present in similar proportions, although with minor variations, relative to the number of total reads in the libraries of both of exosome depletion mutants and in their corresponding non-induced lines, and constituted about half of total smRNAs mapped to the genome (Table S1, Figure 1B). Therefore, for simplicity we graphed only data corresponding either to 20–25 nt smRNAs, or to smRNAs of one specific length.
In addition, the depletion of either RRP4 or RRP41, which are both essential for exosome function, with slight variations, had no effect on the smRNA size distribution (Figure 1B) or the frequencies of their first nucleotide (Figure 1C). All together, these results suggest that defects in exosome function do not lead to accumulation of un-degraded smRNA fragments or to any changes in the cleavage bias of Dicer proteins. Also, exosome depletion did not change proportions of smRNAs mapped to different classes of RNAs, such as mRNAs, tRNAs, rRNAs, and snoRNAs (Figure 1D). Therefore, unlike the situation in S. pombe, where exosome acts as a negative regulator of siRNA biogenesis, Arabidopsis exosome does not act to prevent spurious RNAs from entering RNAi pathway.
In Arabidopsis, repeats and TEs are silenced by siRNAs through RdDM; therefore, we examined the effect of exosome loss on 20–25 nt smRNAs corresponding specifically to TEs and repeats. Surprisingly, we saw no changes in the groups of smRNAs mapped to tandem repeats (TR), inverted repeats (IR), dispersed repeats (DR) or the group of TEs in both exosome mutants (Figure 2A and 2B). The diverse heterochromatic siRNAs participating in TE silencing are mostly 24-mers and are Pol IV- and/or Pol V-dependent [9]. Most siRNA production relies on Pol IV, but there are also Pol V-dependent and Pol IV-independent siRNA-generating loci [44], [45]. Therefore, to examine whether the exosome complex functionally overlaps with the components of the RdDM pathway, we constructed lines containing rrp4-i or rrp41-i iRNAi and mutations affecting Pol IV, Pol V, RDR2 and DCL3, which are nrpd1, nrpe1, dcl3 and rdr2 respectively (allele numbers provided in Methods). This approach also allowed us to confirm that smRNAs observed in exosome depletion lines are siRNAs produced by components of the RdDM pathway and not short RNA degradation products accumulated in the absence of functional exoribonucleolytic complex.
Pol IV, Pol V, RDR2 and DCL3 are not essential for viability [27], [29], [46]. Combining mutations in nrpd1, nrpe1, dcl3 and rdr2 with rrp41-i iRNAi knock-down line did not exacerbate the phenotypes of single exosome depletion mutants (Figure 1A).
We next analyzed the smRNAs corresponding to repeats and TEs produced in the rrp41/nrpd1 and rrp41/nrpe1 double mutants (Figure 2C) and the rrp41/rdr2 and rrp41/dcl3 double mutants (Figure 2D). Similar to previous reports, we observed a significant reduction in the amount of smRNAs corresponding to TEs, TRs and IRs in nrpd1, nrpe1, rdr2, and dcl3 mutants [27], [44], [47], [48]. Depletion of the exosome in nrpd1, nrpe1 and rdr2 mutants had no effect on the amount of TE and repeat-associated smRNAs produced in these mutants (Table S2, Figure 2C and 2D). Depletion of rrp41 in dcl3 led to a minor restoration of this defect in all groups of repeats and TEs. In the absence of dcl3, other Arabidopsis Dicer proteins are known to process dcl3 substrates [49]; therefore this minor restoration most likely resulted from compensatory effects of other DICER proteins (Table S2, Figure 2D). Profiling repeat- and transposable element-generated smRNAs by their size confirmed that the exosome defect did not affect the group of 20–25 nt smRNAs even in Pol IV, Pol V, RDR2 and DCL3 deficient genetic backgrounds. Typically, siRNAs participating in RdDM are 24 nt long; therefore we profiled smRNAs mapping to transposable elements by length, but observed no change in abundance of 24 nt smRNAs (Figure 2E). Further analysis of the 24 nt smRNAs mapped specifically to the different transposable element superfamilies led to the same conclusion (Figure 2F and 2G). We therefore concluded that there are no significant changes in the populations of siRNAs corresponding to repeats and TE superfamilies in exosome depletion mutants. We also did not observe any significant differences in amounts of mature 21-mer miRNAs. The results of our sequencing analysis were confirmed by Northern blot analysis (Table S3, Figure 3, Figure S2). Together, these data suggest that the Arabidopsis exosome complex is not involved in siRNA metabolism on a global scale. Nevertheless, we can not exclude the possibility that exosome might control a small number of smRNA precursor transcripts at a few specific loci that would have been missed in our experiments and with the data processing approach we took while dissecting differences on genomic level.
To further investigate whether the exosome participates in gene silencing and interacts with the RdDM pathway, we examined the transcription patterns of several specific loci regulated through RdDM. solo LTR and AtSN1 are the heterochromatic loci for which the role of RdDM players in their silencing and interactions between them are best-understood [30], [31], [50]–[52]. Transcriptional silencing of solo LTR and AtSN1 heterochromatic loci are dependent on Pol IV and Pol V [30], [31], [50]–[52]. Based on previous studies, both solo LTR and AtSN1 loci can be subdivided into region A and an adjacent region B [30], [31]. Region A represents the siRNA-generating region likely transcribed by Pol IV, and region B gives rise to the ncRNAs that are proposed to serve as a scaffold for recruiting siRNA-mediated complexes that form heterochromatin (Figure 4A). Pol V was proposed to produce transcripts which serve as the scaffolds [31], although in case of solo LTR, Pol II was also shown to be involved [30].
We then used real-time RT–PCR to examine the levels of transcript produced from region A, as a measure of the silencing status of each locus. We found that exosome defects resulted in accumulation of polyadenylated ncRNA produced from both regions A and B of solo LTR (Figure 4B). We then compared the amplitudes of the region A derepression in the rrp41, with rrp41 iRNAi/nrpd1 and rrp41 iRNAi/nrpe1 double mutants relative to the respective single mutants. As previously reported by others [30], [31], we observed solo LTR to be significantly derepressed in Pol IV and Pol V single mutants (Figure 4C and 4F). Interestingly, only the combination of exosome defect with mutation of Pol V, but not with mutation of Pol IV, resulted in the synergistic increase of region A transcript (Figure 4C). Reverse transcription with oligo dT primers does not discriminate between transcripts originating from either DNA strand; thus an elevated level of polyadenylated transcript could result from transcription from either one of the DNA strands. Therefore, to find out which of the transcripts increased in abundance, we carried out strand-specific RT-PCR for the A and B regions.
Following standard nomenclature, the top transcript (also called top strand RNA) corresponds to the transcript identical to the sequence of the DNA top strand (and therefore produced from the bottom DNA strand), and the bottom transcript is identical to the sequence of DNA bottom strand. The scaffold RNAs were reported to correspond to region B top strand [30], [31].
Similar to previous results [30], [31], we observed region A top and bottom transcripts to be significantly derepressed in Pol IV and Pol V single mutants (Figure 4D and 4E), and depletion of RRP41 lead to increased accumulation of the region A top and bottom transcripts (inserts in Figure 4D and 4E). Interestingly, we found that the bottom transcript was synergistically derepressed in rrp41 iRNAi/nrpe1 double mutants relative to nrpe1 and rrp41 iRNAi single mutants, while no change was observed in rrp41 iRNAi/nrpd1 double mutants (Figure 4D). Despite the fact that the exosome defect equally affected the levels of both top and bottom region A transcripts, combining the exosome defect with either Pol IV or Pol V mutants had no additive or synergistic effect on the level of region A top transcript. Surprisingly, the level of expression of region A top transcript was even somewhat decreased in rrp41 iRNAi/nrpd1 and rrp41 iRNAi/nrpe1, compared to nrpd1 and nrpe1 single mutants, opposite to the pattern we observed for the bottom strand (Figure 4E). Production of scaffold transcripts is central in silencing of the locus and it was reported that even in the presence of functional Pol IV and siRNAs, silencing of solo LTR fails when scaffold RNAs are not produced [30], [31].
We therefore examined the scaffold-producing region B and found that the exosome also affects the amount of region B top transcript, but there is no synergistic increase of this transcript in rrp41 iRNAi/nrpe1 double mutants (Figure 4F and 4G). When we examined AtSN1, we observed a very similar synergistic increase in the level of the siRNA-producing region A of bottom strand transcript of AtSN1 in rrp41 iRNAi/nrpe1 mutants (Figure 4H and 4I).
Together, these results suggest that the exosome participates in controlling the amount of top transcripts emanating from the scaffold-producing region B of solo LTR, and thus may contribute to the repression of region A through regulating the level of region B transcripts.
The solo LTR, AtSN1 and IGN5 loci are silenced primarily by RdDM, through siRNA mediated de novo methylation of DNA [9], [30], [31]. We reasoned that if the exosome is involved in controlling the amount of RNA expressed from these loci in a siRNA-dependent manner, then the exosome defect might affect the amount of siRNAs generated from these regions. To address this question, we first compared solo LTR and AtSN1-specific smRNAs. We found that production of smRNAs from the siRNA-generating A regions was not altered in rrp4-i or rrp41-i mutants relative to WT (Figure 5A and 5B), similar to the results of the global smRNA analysis we described above. The increased amount of smRNAs observed in dcl3 mutants is because in the absence of DCL3, the other Dicer proteins process DCL3 substrates [49]. In order to make sure that the smRNAs produced from one strand of region A are not masking the smRNAs produced from the opposite strand in exosome depletion mutants, we also analyzed these smRNA populations in a strand-specific manner. However, the patterns of strand-specific siRNAs were very similar to the patterns we observed previously and siRNAs were not affected by exosome depletion (Figure 5C and 5D). We examined an additional region controlled by RdDM, the IGN5 locus [31], and found that IGN5-specific smRNAs are also not affected in exosome mutants, similar to solo LTR and AtSN1 loci (Figure S3C). This implies that the increase in accumulation of transcripts in exosome-depleted plants was not a result of siRNA defect. To verify this directly, we examined the patterns of DNA methylation in these regions by using methylation sensitive restriction enzymes (Figure 5E). The DNA of the solo LTR region was examined by two different assays (Figure 5E and 5F). We found that, consistent with the results of the region-specific siRNA analysis, de novo DNA methylation was not changed in rrp41-i plants (Figure 5A–5D). Taken together, these results indicate that an increase in transcript accumulation is not caused by the loss of de novo methylation and the region is still silenced by RdDM. It also suggests that in the examined loci, the exosome complex functions independently of RdDM. Interestingly, the increased amount of transcripts accumulated in these regions does not contribute to increased smRNA amounts in the exosome-depleted plants. This was observed regardless of whether these transcripts originated from siRNA-generating regions, or adjacent regions. Indeed, even several thousand-fold upregulation of region A transcript in iRNAi/nrpe1 mutants (Figure 4B, 4C, 4G and 4H) does not produce any increase in the amount of siRNAs (Figure 5A–5D).
DNA methylation and histone modification are two major epigenetic marks regulating gene expression and chromatin state in plants. Monomethylated histone H3 lysine 27 (H3K27me1) and dimethylated histone H3 lysine 9 (H3K9me2) are hallmarks of heterochromatin, and silencing of solo LTR, AtSN1 and IGN5 loci also involves histone modifications [30], [31]. Although de novo methylation does not directly affect the level of H3K9me2, it does affect the level of H3K27me1 [31], suggesting that in addition to histone modification pathways, which are dependent on RdDM, other, RdDM-independent, pathways also contribute to transcriptional silencing of these regions. We therefore used chromatin immunoprecipitation (ChIP) to examine whether the exosome is involved in regulation of histone modifications in these loci.
Similar to the results reported by others [30], [31], we observed a significant decrease in the level of H3K9me2 in the solo LTR locus in nrpd1 and nrpe1 mutants, which affect Pol IV and Pol V, respectively. We found that RRP41 depletion also led to a decrease in H3K9me2 but less than observed in nrpd1 and nrpe1 mutants (Figure 6A). The decrease in level of this repressive histone modification also correlated with a mild increase in RNA Pol II occupancy in the solo LTR region, as would be expected with a release of transcriptional block (Figure 6B). The rrp41 iRNAi/nrpe1 double mutant did not exhibit any additive or synergistic effect on the loss of H3K9me2 relative to respective single mutants.
When we examined AtSN1, we found that the level of H3K9me2 was mildly decreased in all mutants tested (Figure 6A). For AtSN1, it was previously suggested that RNA Pol III is the main RNA polymerase transcribing the region when the region is in a derepressed state [31], although RNA Pol II was also reported to be associated with this region [30]. We found that RNA Pol II occupancy in AtSN1 was very low but it increased significantly in rrp41 iRNAi/nrpe1 double mutants (Figure 6B), in accordance with the synergistic increase of the transcript level we observed (Figure 4H and 4I).
Depletion of another exosome subunit, RRP4, caused a similar loss of H3K9me2 at solo LTR and AtSN1 loci (Figure 6C). We then chose several additional regions, termed REG3 and REG4 (Figure S3A), that are mildly upregulated in exosome mutants according to our previous microarray analysis [1], and examined them using ChIP. REG3 harbors a MuDR transposon, and REG4 is situated in a tandem repeat area. Neither of these loci produces smRNAs or is controlled by DNA methylation (Figure 6E and data not shown). We found that the H3K9me2 in these loci was similarly affected by exosome depletion (Figure 6C), while the level of H3K27 methylation in these regions didn't show any difference (Figure 6D). These results suggest that the exosome may participate in maintaining chromatin structure in these regions as well, and does so by specifically affecting the level of H3K9me2 in addition to controlling the level of transcripts.
We then examined exosome association with ncRNA loci. Detection of some protein–nascent mRNA interactions by ChIP were reported previously for proteins working on RNA, but the results of our attempts to localize tagged exosome subunits at solo LTR locus have proven inconclusive. Transcripts from region A are normally below the level of detection in wild-type plants, but transcription from the region B adjacent to solo LTR has been previously documented in wild-type plants [1], [30], [31]. In order to confirm that the exosome directly associates with these transcripts, we conducted RNA immunoprecipitation (RIP) using plants carrying a transgene expressing RRP41-TAP, and examined the ncRNAs associated with the exosome by RT-PCR. No region A transcripts were detected in immunoprecipitates, but we found that region B transcripts were co-precipitated with exosome (Figure 7A). These data suggest that in wild-type plants, exosome physically associates with polyadenylated transcripts produced from region B of solo LTR.
In contrast to solo LTR, we did not detect a physical association of exosome with AtSN1 region B transcript (Figure 7A). This implies that exosome depletion may not directly affect the silencing of AtSN1. However, we observed that exosome depletion resulted in accumulation of transcript in the AtSN1 locus and we detected a synergistic derepression of the locus in rrp41/nrpe1 mutants, similar to solo LTR locus (Figure 4H and 4I). Most likely the regulation of AtSN1 is more complex because an additional RNA polymerase, RNA Pol III, is involved. AtSN1 is transcribed mostly by RNA Pol III [31], [53], suggesting that the double deficiency in exosome and Pol V may increase both Pol II and Pol III access to the locus. We also observed the increased Pol II association with AtSN1 in rrp41/nrpe1 mutants by ChIP assay using anti-Pol II (Figure 6B), which is consistent with the results of qRT-PCR. Therefore, it is also possible that the loss of exosome function may lead to the alteration of chromatin structure in regions adjacent to AtSN1 and thus affect the stability of silencing in AtSN1 indirectly. Nevertheless, these results are similar to the interplay between exosome and Pol V observed for solo LTR.
The 9-subunit exosome complex is catalytically inactive in yeast and human. Instead, active sites are contributed by Rrp44 (Dis3) and by the subunit Rrp6, which is substoichiometric, nuclear-specific, and not essential for viability. Degradation of S. cerevisiae nuclear ncRNAs depends on polyadenylation by the TRAMP complex and involves Rrp6, the subunit that is also responsible for elimination of heterochromatic RNAs in S. pombe [20], [22]–[24], [39]–[41]. In Arabidopsis there are three RRP6-like proteins – nuclear localized RRP6L1 and RRP6L2, and cytoplasmic RRP6L3; these were suggested to be functional homologues of RRP6 [54]. None of the RRP6-like proteins co-purified with the exosome complex in our proteomic studies [1], but may have been underrepresented in our preparations. In addition, RRP6L2 was later shown to have at least some commonalities with core exosome substrates [54]. We therefore examined whether the Arabidopsis RRP6-like proteins control the amount of ncRNA at the solo LTR locus. To determine this, we used T-DNA insertion alleles in RRP6L1, RRP6L2 and RRP6L3. We isolated the rrp6l1-2 allele from the University of Wisconsin BASTA population (Ws ecotype), and the alleles of the rrp6l2-2 and rrp6l3-1 are SALK alleles (Col-0 ecotype). To control for effects of ecotype, we compared the amount of region A transcript in rrp6l3-1, rrp6l2-2, rrp6l1-2/rrp6l2-2 mutants to Col-0 wild type plants, and rrp6l1-2, rrp6l1-2/rrp6l2-2 mutants to Ws ecotype plants (Figure 7B and 7C).
We found that, similar to depletion of the core subunits RRP4 and RRP41, rrp6l1-2 and rrp6l2-2 mutants exhibited increased accumulation of transcripts produced from region A. As would be expected based on cytoplasmic localization of RRP6L3 protein, no effect was observed in rrp6l3-1 mutants. To our surprise, we observed a dramatic derepression of region A in rrp6l1-2/rrp6l2-2 double mutants, suggesting that both RRP6L1 and RRP6L2 proteins are involved in the silencing of this region and might have a redundant function in this process.
We also examined the status of solo LTR DNA methylation in rrp6l1-2, rrp6l2-2, and rrp6l1-2/rrp6l2-2 double mutants. We found that methylation was not affected in these mutants regardless of the extent of derepression of the region (Figure 7D), consistent with the results obtained using rrp4-i and rrp41-i depletion mutants. Taken together, these results indicate that the observed increase in transcript accumulation is not caused by the loss of de novo methylation and the region is still methylated by RdDM. This further confirms that the exosome complex functions independently of the RdDM pathway.
The exosome functions in virtually all aspects of RNA metabolism and it appears to also have a prominent role in transcriptional gene silencing in different species [1], [10], [38]–[41], [55]–[59]. This study examined the role of the exosome complex in metabolism of smRNAs and explored the possible relationship between the exosome and the RdDM pathway in gene silencing in Arabidopsis.
Our results showed that exosome-mediated silencing did not produce global changes in smRNA profiles, nor in DNA methylation at specific loci. However, we did find effects on histone methylation, indicating that the exosome may regulate chromatin structure, thereby playing an important role in maintenance of gene silencing on a much broader scale than the RdDM pathway. It is clear from our results using suppression of key exosome components that plants have an exosome-dependent pathway that relies on ncRNAs to target heterochromatin.
Our finding that the increase in ncRNA transcribed from heterochromatic loci in exosome-depleted plants did not lead to an increase in levels of smRNA indicates that exosome function in Arabidopsis differs from that in fission yeast. In fission yeast, exosome defects have a dramatic effect on siRNAs leading to redistribution of the spectrum of Ago1-associated siRNAs, from mostly repeat-associated to those derived predominantly from exosome substrates such as rRNA and tRNA [39], indicative of exosome acting as a negative regulator of siRNA biogenesis. Our data indicate that the Arabidopsis exosome most likely lost this function during evolution, meaning that exosome substrates do not compete with siRNA precursors for siRNA biogenesis machinery and spurious transcripts do not enter RNAi pathways in plants. Additionally, it suggests that perhaps only very few of the ncRNA transcripts controlled by the exosome could be bona fide siRNA precursors. One of the reasons for this could be the fact that plants evolved two plant-specific RNA polymerases, Pol IV and Pol V, which specialize in siRNA-mediated TGS. Pol IV is required for biogenesis of the majority of 24-nt siRNAs and is supported by Pol V, which is responsible for production of a subset of siRNAs [31], [44], [45], [60]. It is also plausible that there might be other unknown plant-specific ribonucleases that specialize in controlling stability of siRNAs or the amount of siRNA precursors generated by Pol IV and/or Pol V in plants. We also cannot rule out the possibility that some of the transcripts controlled by the exosome in a small subset of loci are legitimate siRNA precursors; this definitely warrants further in-depth investigation.
siRNA-dependent RdDM is thought to be the main pathway for transcriptional gene silencing of repetitive elements and transposons in plants [27], [28], [31], [61], [62], although existence of other DNA methylation-independent gene silencing pathways have also been reported [63]–[71]. One of the DNA methylation-independent gene silencing pathways is mediated by MOM1 (Morpheus' molecule 1) protein [63], [65], which predominantly silences transposons and loci harboring sequences related to gypsy-like transposons. Activation of transcription in mom1 mutants occurs with no change in DNA methylation, histone modifications or chromatin condensation, and the investigation of the relationship between RdDM and MOM1 revealed a very complex interplay between these two pathways [63], [69], [72]–[74]. However, a reduction in H3K9 dimethylation was reported in some loci in mom1 mutants and it was suggested that MOM1 may transduce RdDM signals to repressive histone modifications by an unknown mechanism [75].
Also, a recent study of MORC family ATPases revealed that mutation of AtMORC1 or AtMORC6 caused derepression of DNA methylated genes and TEs without any loss of DNA methylation, change in histone methylation or alteration of siRNA levels [71]. These proteins are involved in alteration of chromosome superstructure and are likely to act downstream of DNA methylation. These results indicate that there are multiple parallel pathways for DNA methylation-independent gene silencing in Arabidopsis. The exosome-mediated silencing we observed here bears some similarities to the silencing observed for MOM1 and MORC; for example, they show effects on repetitive sequences and an absence of effects on siRNAs, although there are notable differences as well. Here we show that, similar to MOM1 and MORC mechanisms, exosome-dependent gene silencing also affects repetitive sequences and acts independent of RdDM, although our results are limited in scope. Characterization of the relationship between these pathways remains an interesting topic for future study.
The different silencing pathways likely have different functions, depending on the genomic region, the nature of the regulated sequences, and the precision and dynamics of silencing required. For example, methylated sequences can affect the expression of nearby genes. The expression of nearby genes is negatively correlated with the density of methylated, but not unmethylated TEs. Methylated TEs are preferentially removed from gene-dense regions over time and TE families that have a higher proportion of methylated insertions are distributed farther from genes [76], arguing that positional effects and the surrounding landscape most likely contributes to the choice of silencing mechanisms and the interplay between them.
There are multiple mechanisms by which the exosome can be envisioned to participate in gene silencing in Arabidopsis. Heterochromatin assembly is used by all eukaryotes in gene silencing. In addition to repressive histone modifications employed by all organisms, humans and plants widely use DNA methylation as well, and ncRNAs play a central role in the control of chromatin structure in all organisms. While ncRNA-mediated silencing proceeds through multiple mechanisms some of which are organism-specific, the end result appears to be the same repressive histone modifications. For example, budding yeast, which lacks RNAi machinery, employs strategies that include, but not limited to, the use of antisense, cryptic or read-through transcripts, as well as transcripts originating from divergent promoters to guide histone modifications. Fission yeast is more similar to higher eukaryotes and uses all of the above strategies in addition to utilizing RNAi as well. However, DNA methylation is not used by budding and fission yeast. Plants, on the other hand, evolved very sophisticated epigenetic mechanisms that include the use of both RNAi-dependent and RNAi-independent pathways to guide DNA methylation and histone modifications for gene silencing [9], [31]–[33], [44], [45], [47], [61], [68], [70], [75], [77], [78]. Exosome complex proved to be amazingly versatile in impacting gene silencing in budding and fission yeasts. In fission yeast, the organism which takes full advantage of RNAi machinery to regulate its gene expression, the exosome is involved in silencing of both facultative and constitutive heterochromatin by acting in several different pathways through smRNAs, produced in either an RNAi-dependent or RNAi-independent manner [38], [39], [79], [80]. It was also found to act through surveillance of RNA quantity and quality as well as by collaborating with termination machinery [40], [41], [57], [80], [81], similarly to the manner exosome participates in gene silencing in bakers yeast, which lacks RNAi machinery [55], [58], [59].
In Arabidopsis, silencing of repetitive elements involves siRNA-dependent DNA methylation guided by homologous siRNAs [9]. Repressive histone modifications always appear to accompany DNA methylation, however, the mechanistic link between them is not yet fully understood. In budding and fission yeasts, degradation of nuclear ncRNAs depends on polyadenylation by the TRAMP complex and involves Rrp6. We also found that mutations in two RRP6-like proteins AtRRP6 L1and AtRRP6 L2 led to significant dereperession of solo LTR (Figure 7B, 7C and 7D) and occurred in a DNA methylation-independent manner as in rrp4 and rrp41 (Figure 7F). These results suggest that Atrrp6s may be true nuclear catalytic subunits of Arabidopsis exosome, or may also work independently of core exosome. It will be interesting to examine whether another putative exosome catalytic subunit AtRrp44a [J. Lee and J. Chekanova unpublished data] is involved in this process, and whether components of the TRAMP complex also participate.
We also observed that the exosome physically associates with the polyadenylated ncRNA transcripts from scaffold producing regions. We could not reliably crosslink the exosome to the DNA of the target locus by ChIP (data not shown), although this could simply reflect the difficulty of reliably crosslinking proteins to DNA through RNA, or it could mean that the exosome binds to the transcripts after they are released from the locus and that exosome-mediated regulation of the transcripts may be important for maintenance of chromatin structure around the locus. H3K9 dimethylation was reported to be disturbed and lost when isolated Arabidopsis nuclei were treated with RNase A [82], meaning that histone modification may be affected by RNA level and/or RNA in close proximity to the target loci. In fission yeast, the mutation of Cid14, one of the subunits of the TRAMP complex, results in accumulation of aberrant heterochromatic RNA close to the target loci and leads to a mild decrease in H3K9 methylation. It was recently shown that decrease of H3K9 methylation in yeast is the result of HP1 protein (Heterochromatin Protein1), which binds to H3K9me2 heterochromatin and propagates H3K9me2 spreading, being titrated by an excess of heterochromatic RNA [83]. In our study, we also observed a combination of the transcripts accumulation in exosome mutants relative to WT with a weak decrease in H3K9me2 levels in solo LTR (Figure 6A). Taken together, these data could suggest that a similar mechanism to regulate the stability of chromatin structure might operate in plants. However, LHP1 (Like-HP1), the closest Arabidopsis homolog of yeast HP1, has specificity for H3K27me3 [84], not H3K9me2, and the rrp41 iRNAi/nrpe1 double mutant did not exhibit any additive or synergistic effect on the loss of H3K9me2 relative to respective single mutants as well, suggesting that the loss of H3K9me2 observed in the exosome mutants is unlikely to result from an unknown functional homolog of Arabidopsis HP1 simply titrating an excess of ncRNA off chromatin, as reported in fission yeast.
Our results showed that the exosome depletion produced no effect on siRNAs and DNA methylation of solo LTR, AtSN1 and IGN5 loci, arguing that the exosome complex functions independently of RdDM. However, our findings also indicated that the exosome is involved in the silencing of these loci and does interact with the RdDM pathway, possibly through its functional interaction with RNA Pol V. The converging transcripts we observed in the rrp41-i and rrp4-i mutants in solo LTR and AtSN1 suggest that the exosome is involved in regulation of either processing or level of RNA from these loci (Figure 4A–4I, and model Figure 8). We found that production of smRNAs from the siRNA-generating A regions was totally abolished in rrp41/nrpd1 double mutant (Figure 5A–5D), ruling out a possibility for these transcripts to serve as a double stranded precursors for RNA Pol IV-independent siRNAs. We also found that the exosome physically associates with the polyadenylated transcripts produced from the scaffold region (region B) and exhibits synergistic derepression of the locus (region A) when combined with a Pol V mutant, while there was no change in the derepression in rrp41/nrpd1 double mutants (Figure 4B, 4C, 4H and 4I). Based on these results, we speculate that RNA polymerase V may function in gene silencing of these loci in two ways, the first acting in the DNA- methylation-dependent RdDM pathway, and the second acting independently of a DNA-methylation. Indeed, RdDM- independent roles of Pol V in silencing of 5S rDNA [31], [85] and several other loci [82] were previously reported. A recent genome-wide study of Pol V-associated loci also hints at the possibility of Pol V having unknown functions in addition to the function it plays in the RdDM pathway [45]. The DNA-methylation-independent function of Pol V may then be in addition to its function in RdDM, and may operate in parallel to the exosome pathway. If this is the case, the depletion of both rrp41 and nrpd1 may not lead to synergistic derepression because it would be compensated by the RdDM-independent function of Pol V. However, deficiencies in exosome and Pol V would result in synergistic desilencing due to the loss of three different pathways. Both Pol II and Pol V were reported to be responsible for the transcription of scaffold RNA and be required for silencing [30], [31], although it is not known how their activities are functionally integrated. It is also not known how Pol V initiation sites are chosen, but they appear to be promoter independent [31]. Perhaps transcription by Pol II helps maintain open chromatin architecture at this site, and together with the resulting noncoding RNAs facilitates Pol V transcription initiation. Alternative possibility is that Pol II produces two distinct pools of transcripts, one of which is controlled by the exosome, and the exosome functions by regulating the Pol II transcripts that are distinct from the transcripts that are used in RdDM pathway. This possibility would be very interesting to examine, particularly in light of the yeast exosome involvement in gene silencing through regulation of cryptic transcripts, transcripts originating from divergent promoters and read-through transcripts [4], [55], [58], [59]. How the Arabidopsis exosome complex and the exosome controlled ncRNAs facilitate recruitment of chromatin modifiers in order to enforce silencing through repressive histone modifications remains an interesting topic of future studies. We suggest that the exosome may coordinate the transcriptional interplay of RNA polymerases Pol II and Pol V to achieve the right level of transcriptional repression of heterochromatic loci (Figure 8).
In summary, our data suggest that the exosome likely acts in a parallel pathway to RdDM pathways in gene silencing, possibly affecting the transcriptional interplay of different RNA polymerases to modulate repression of heterochromatic sequences. The mechanisms that link this RNA metabolic complex, the epigenetic modification of histone methylation, and heterochromatic silencing in plants remain to be elucidated. Our results indicate that there is no one-size-fits-all pathway or mechanism that exclusively governs silencing of all loci; rather, different loci and different players in RdDM interact with different pathways and are silenced by different, likely overlapping mechanisms. The positional effects and the surrounding landscape most likely also play important roles in the choice of silencing mechanisms and the interplay between them. This may reflect the crucial importance of silencing in developmental gene regulation and in maintenance of genomic stability by suppression of invasive sequences.
iRNAi lines of exosome subunits RRP4 and RRP41, RNA Pol IV (SALK_128428.20.10, nrpd1a-3, nrpd1-3), RNA Pol V (SALK_029919, nrpd1b-11, nrpe1-11), RDR2 ( SAIL_1277808, rdr2-1), and DCL3 ( SALK_005512.38.70.x0, dcl3-1) mutants were described previously [1], [27], [33], [86]. rrp41 iRNAi/nrpd1-3, rrp41 iRNAi/nrpe1-11, rrp4 iRNAi/nrpd1-3,and rrp4 iRNAi/nrpe1-11 double mutants were obtained by crossing of rrp41 iRNAi and rrp4 iRNAi with nrpd1/nrpe1-11 line. rrp41 iRNAi/dcl3-1, rrp41 iRNAi/rdr2-1 double mutants were obtained by crossing.
The alleles of the rrp6l2-2 and rrp6l3-1 correspond to SALK_011429 and SALK_122492 lines, respectively. The rrp6l1-2 allele was isolated from the University of Wisconsin BASTA population. The ecotype background is Col-0 for all Salk alleles and Ws for University of Wisconsin alleles. To induce iRNAi, seedlings were germinated and grown for 7 days on ½× MS plates with 8 mM 17β-estradiol, as described before [1].
Total RNA was isolated from 7-day-old seedlings using the mirVana miRNA isolation kit (Ambion) according to the manufacturer's protocol. The total RNA sample was used for sequencing library construction using the Small RNA sample Prep v1.5 kit and TruSeq Small RNA Sample Prep kit (Illumina, San Diego, CA) according to the manufacturer's instructions. The smRNA libraries were sequenced using the Illumina Genetic Analyzer II (by DNA Core Facility, University of Missouri) and Illumina HiSeq 2000 (by Biotechnology Center, University of Wisconsin) according to the manufacturer's instructions. HiSeq 2000 sequencing reads were demultiplexed using Casava v 1.8 (by Bioinformatic Resource Center, University of Wisconsin) before further bioinformatic analysis
Data processing was done using available tools and custom in-house UNIX shell programming [43], [75], [87]–[90]. The raw sequences in Illumina GAIIx and demultiplexed HiSeq 2000 sequencing reads were trimmed removing adapter using “fastx_clipper” in the FASTX-Toolkit (version 0.0.13) [91] and smRNAs with lengths between 15- and 32-nt were selected and mapped to the Arabidopsis genomic sequences (TAIR9 version) using BOWTIE (version 0.12.7) [92]. Reads that failed to perfectly map to the nuclear genome with no mismatches, and reads present in fewer than two counts were discarded. All Arabidopsis lines used in this study carried iRNAi cassette transgenes used for inactivation of either RRP4 or RRP41 exosome subunit genes [1]. These silencing cassettes generate a number of 21-, 22- and 24-nt silencer sequences corresponding to RRP4 or RRP41 genes (mapping to AT1G03360 and AT3G61620 loci), respectively. Therefore, silencer sequences produced from iRNAi transgenes were filtered out from each library and libraries were analyzed separately to ensure accurate interpretations. The remaining smRNA reads, termed FLR for filtered reads, were used for further analysis.
Each library was normalized either to the total number of mapped non-redundant reads or to the total number of non-redundant filtered reads (FLR), multiplied by 106 (rpm, reads per million). Both methods of normalizations were compared and found to produce results which lead to identical interpretations, therefore, only data analyzed using filtered reads are presented in this study.
Classification of small RNAs was performed by BEDTools (v2.10.0) [93] and in-house UNIX shell programming using the following databases: TAIR9 annotations for protein coding and non-coding features (tRNA, rRNA, ncNRA, miRNA, snRNA, snoRNA, and transposable elements [76]), miRBase (release 18) [94] or mature miRNA annotations. Some smRNAs match more than one annotation category; therefore the sum of the numbers is bigger than the total input number.
The small RNA reads with 20 to 25 nt length were calculated and plotted versus the sum of their normalized reads per million (rpm). The relative frequencies of each 5′ terminal nucleotide of the small RNAs were calculated (Tables S1, S2 ) and represented graphically.
Repetitive genomic features were classified using TAIR9 Tandem Repeat Finder (version 4.04) [95] and Inverted Repeat Finder (version 3.05) [96]. Annotation of dispersed repeats was done with Repeat Masker (version 3-3-0) [97].
For analysis of locus-specific expression of smRNAs (solo LTR, AtSN1, IGN5, REG3, and REG4), the expressed normalized reads per million (rpm) were calculated for respective genomic locus and locus-specific datasets were plotted for comparisons.
Total RNA was isolated from 7-day-old seedlings using the Trizol method. For RT-qPCR, 1–4 µg of total RNA digested with DNase I (Fermentas) was reverse transcribed 1 hour either at 50°C (for oligo-dT primer) or 55°C (for specific primers) using 60–100 units SuperScript III Reverse Transcriptase (Invitrogen). Transcripts were quantified by RT-qPCR using the comparative threshold cycle method (ΔΔCt, primers listed in Table S4), using Actin2 (At3g18780) as endogenous reference. Polyacrylamide Northern Blot analyses were performed as described [25].
Genomic DNA was isolated from 7-day-old seedlings using a DNeasy kit (QIAGEN). The methylation analysis using DNA sensitive methylation enzymes was followed as described [27], [31], [77].
ChIP was performed as described [98]. One gram of 7-day-old seedlings was used for each experiment. All ChIP experiments were reproduced at least twice on each of the two or more biological replicates. The normalization was done relative to input using [99]. Anti-RNA Pol II (ab817) and anti-H3K9me2 (ab1220) were obtained from Abcam, and anti-H3K27me1 antibody from Upstate. An equal amount of chromatin not treated with antibody was used as the mock antibody control. The ChIPed DNA was purified using PCR purification kit (Fermentas) before being used for qPCR. The primer sets used for the PCR are listed in Table S4.
RIP assays were performed by adapting an existing protocol [100]. Transgenic plants expressing TAP-tagged RRP41 at physiological levels [1] were used in the experiment. Two grams of 2-week-old seedlings were collected and fixed with 1% formaldehyde. For RRP41-RNA complex purification, the chromatin solution was incubated overnight with prewashed IgG Sepharose 6 Fast Flow (GE Healthcare) at 4°C. Immunoprecipitated RNA was purified with phenol: chloroform and cDNA synthesis was performed using SuperScript III reverse transcriptase (Invitrogen) and random hexamers (Promega). The primer sets used for the PCR are listed in Table S4.
|
10.1371/journal.pcbi.1002303 | A Hierarchical Neuronal Model for Generation and Online Recognition of Birdsongs | The neuronal system underlying learning, generation and recognition of song in birds is one of the best-studied systems in the neurosciences. Here, we use these experimental findings to derive a neurobiologically plausible, dynamic, hierarchical model of birdsong generation and transform it into a functional model of birdsong recognition. The generation model consists of neuronal rate models and includes critical anatomical components like the premotor song-control nucleus HVC (proper name), the premotor nucleus RA (robust nucleus of the arcopallium), and a model of the syringeal and respiratory organs. We use Bayesian inference of this dynamical system to derive a possible mechanism for how birds can efficiently and robustly recognize the songs of their conspecifics in an online fashion. Our results indicate that the specific way birdsong is generated enables a listening bird to robustly and rapidly perceive embedded information at multiple time scales of a song. The resulting mechanism can be useful for investigating the functional roles of auditory recognition areas and providing predictions for future birdsong experiments.
| How do birds communicate via their songs? Investigating this question may not only lead to a better understanding of communication via birdsong, but many believe that the answer will also give us hints about how humans decode speech from complex sound wave modulations. In birds, the output and neuronal responses of the song generation system can be measured precisely and this has resulted in a considerable body of experimental findings. We used these findings to assemble a complete model of birdsong generation and use it as the basis for constructing a potentially neurobiologically plausible, artificial recognition system based on state-of-the-art Bayesian inference techniques. Our artificial system resembles the real birdsong system when performing recognition tasks and may be used as a functional model to explain and predict experimental findings in song recognition.
| Songbirds are able to repeat the same, often complex songs with amazing precision. When male birds sing to a female repeatedly, there is on average a 1% temporal deviation across the whole song [1], [2]. This combination of complexity and precision is remarkable. Studying the neuronal basis of birdsong generation may lead to an understanding of the mechanism underlying how sequences of song syllables are expressed as complex and temporally precise sound wave modulations. More generally, such a mechanism may also be useful for understanding how action sequences at a relatively slow time-scale (e.g. the words in a sentence) can be generated by a neuronal system while a high degree of precision is maintained in the output at a fast time-scale (e.g. the sound wave modulations necessary to form speech sounds).
Recent findings [1]–[3] have shown that the song generation mechanism in birds is hierarchical where neurons in one particular high-level structure, HVC, fire in a specific sequence with high temporal precision and drive neurons in the lower level structure RA (robust nucleus of the arcopallium).
Female birds, at which the songs are typically directed, are expert in registering variables like the speed of the song and the precision and the repertoire of the singer [4]–[8]. Unfortunately, the study of song recognition is more challenging than song generation because experimental indicators for recognition, such as the subsequent behavior of a female bird, are more difficult to measure than indicators for song generation. This has led to a long list of experimental and theoretical findings on song generation and learning while the mechanisms of song recognition remain relatively elusive.
Here, we propose that the functional mechanism of song recognition can be obtained from the song generation mechanism. The basic idea underlying this novel modeling approach is that female birds are optimal in song recognition because their mating choice critically depends on the optimal recognition of valuable features of the male which are revealed by subtle indicators in his song. Similarly, male birds should be able to distinguish the songs of their neighbors from the songs of strangers to protect their territories [9], [10]. Using a recently established Bayesian inference technique for nonlinear dynamical systems [11], we can emulate this optimal recognition: the key ingredient is a generative model (a nonlinear dynamical system) which can generate a specific song. Usually, generative models for complex sensory dynamics, such as the sound wave or spectrum of birdsong, are difficult to derive because it is hard to describe a complex multi-scale structure like birdsong using only differential equations. Fortunately, since the hierarchical birdsong generating system is so well-studied, parts of such a model already exist, in particular at the level of the HVC, RA and vocal tract dynamics [12]–[18]. We have combined these parts into a coherent whole, guided by key experimental results, to form a generative model that can play complex songs. In particular, we combined sequence-generating dynamics, attractor dynamics and a model of vocal tract dynamics [17] in a three-level, hierarchical nonlinear dynamical system. This dynamic model is based on neuronal rate models, thereby describing the biological system at a mesoscopic level. We then used Bayesian inference to derive another set of hierarchical, nonlinear differential equations (recognition system) which is, by way of construction, Bayes-optimal in recognizing this song and can be compared to the real birdsong recognition system. To do this, we exposed the agent to several tasks and found that the agent's dynamics and performance were reminiscent of song recognition in real bird brains in aspects such as sensitivity to speed changes [19] and song perturbations [20], [21]. Thus, by harnessing rich experimental and theoretical results in birdsong generation, we were able to derive a novel, functional model of birdsong recognition. We discuss the experimental evidence that the identified mechanism is indeed used for song recognition by birds. We suggest that the present model may be useful for understanding the functional and computational roles of auditory recognition areas. In addition, the identified recognition mechanism can be used as a novel machine learning tool to recognize sequential behavior from fast sensory input, e.g. in artificial speech recognition.
In this section, we will briefly summarize relevant experimental findings, motivate and describe the present model for birdsong generation and briefly give the mathematical details.
A birdsong consists of small units called notes (analogous to phonetic units in speech) which can be grouped together to form syllables [22]. A combination of identical or different syllables forms motifs. This hierarchical structure of song units is produced by two highly specialized song pathways (Figure 1, see [23] for a review). In the motor pathway, the forebrain nucleus HVC includes specific neurons called HVC(RA) that project to nucleus RA. RA neurons innervate the vocal and respiratory nuclei to produce vocal output. The anterior forebrain pathway is involved in learning new songs and producing variability for the song structure [24].
Our modeling approach is based on the following key experimental observations: During birdsong generation, HVC(RA) neurons fire sequentially at temporally precise moments where each element of this sequence fires only once during the song to control a group of RA neurons [2], [3], [25]. This suggests that bursting HVC(RA) neurons select and drive the activity of subsets of RA neurons [25]. In particular, each RA neuron can be driven by more than one HVC(RA) neuron [26], see Figure 2.
How can one model such a mechanism? There have been several approaches to model the sequential activation of HVC(RA) neurons using single neuron models [12]–[15]. Here, we follow an alternative way by capturing the neuronal mass activity using firing rate models, i.e. we consider model neurons that can be thought of as the synchronized firing activity of an ensemble of neurons. This is motivated by experimental evidence suggesting that there are about 200 co-active HVC(RA) neurons at a specific time during song generation [25]. One of the well established ways for modeling the sequential activation of neuronal ensembles is the winnerless competition using Lotka-Volterra type dynamics [27], [28]. This approach aims at modeling activity at a mesoscopic level, e.g. activity that may be expressed in local field potentials.
Another benefit of using ensemble dynamics appears at the RA level where each ensemble controls the vocal tract muscles in a specific way. Different than HVC(RA) ensembles, one or more RA ensembles can activate simultaneously (synchronize) [25] (Figure 2). We hypothesize that the complex sound wave modulations that can be observed in many birdsongs are generated by this network of RA ensembles using spatiotemporal coding (see also [26]). This coding requires the activation of different sets of RA ensembles (spatial coding) when the proper signals are received from the corresponding HVC(RA) ensembles (temporal coding). This spatiotemporal coding can be modeled with network states which are driven from one attractor to another where each of these attractors specifies the currently active RA ensembles. In other words, when HVC(RA) ensembles undergo sequential activations, the RA level is driven from one attractor to the next. Such networks with attractor dynamics (so called Hopfield networks [29]) can encode a large number of potential attractors because the forcing input from the HVC level effectively recombines subsets of RA ensembles in distinct assemblies.
Note that since the activity of each unit represents the average firing rate of an ensemble of neurons, some features of neural activity at the level of individual neurons are not considered. Here, we focus on capturing the two key features of the hierarchy, which are the sequential firing of HVC ensembles and the spatiotemporal coding at the RA level. Therefore, the choice of parameters in the computational model below are motivated by capturing the specific dynamics inferred by experiments [25].
At the lowest level, we map the dynamical RA states onto motor neurons. To do this, we compute linear combinations of oscillators at different frequencies which represent the effect of currently active RA ensembles and create dynamical control signals (Figure 3) for a model of the vocal organ, the syrinx [17]. This mathematical model of the syrinx has been used previously to model several birdsongs [30], [31].
In summary, the present three-level hierarchical model generates sequences at its top (HVC) level, which are transformed into sequences of multi-dimensional attractors at the RA level. Each of these attractors encodes a mixture of oscillations. These oscillatory dynamics enter a syrinx model as a control signal to produce a birdsong sonogram. In the following, we describe the equations used at each level in detail (see Figure 4 for an overview). The Bayesian recognition of dynamics generated by this birdsong model is described at the end of the section.
Lotka-Volterra equations are well known in population biology to describe the competition between species [32]. Rabinovich et al. (see [33] for a review) applied this idea more generally to neuronal dynamics under the name of winnerless competition, see [28] and [34] for applications. In the following, we will describe how one can apply this idea to model sequential HVC activity by a nonlinear dynamical system. In the winnerless competition setting, there are equilibrium points which are saddles of a nonlinear dynamical system. Each of these equilibrium points has a single unstable direction and all other directions are stable. One can think of these saddle points as the beads on a string where the unstable manifold of one saddle point is the stable manifold of the next saddle point and this sequence continues in a circular fashion forming a heteroclinic chain. Under some conditions [27], this sequence is stable, i.e. a solution of the system that starts from a neighborhood of the chain, stays in this neighborhood at all times while traveling through all saddle points. This stable sequential behavior is what we exploit to model the experimentally established sequential activities of HVC(RA) ensembles at the highest level. As the solution of the system moves along the string, it visits all saddle points, i.e. each HVC(RA) ensemble, one by one thereby activating each ensemble for a brief period until it is deactivated as the next ensemble becomes active.
These dynamics can be obtained from a neural mass model of mean membrane potential and action firing potential [35], reviewed in [33]. We use the equations:(1)where is the hidden-state vector (e.g., mean membrane potentials) at the third (HVC) level, and are scalars, is the sigmoid function applied component-wise and is the connectivity matrix with entries giving the strength of inhibition from state to . The second equation describes the output vector (or causal-state vector; e.g., neural firing rates) where , , is a normalizing function. We also add normally distributed noise vectors and to render the model stochastic. With an appropriately chosen connectivity matrix, one can obtain a system with saddle points forming a stable heteroclinic chain [27]. For the entries of the connectivity matrix, one chooses high inhibition from the previously active neuron to the currently active neuron and low inhibition from the current active neuron to the next neuron which will become active:(Here when and when ).
Note that, theoretically, one can generate arbitrarily long sequences of HVC activation using the above connectivity matrix. The stability region around the heteroclinic chain will persist for much longer sequences than the one modeled here. For our illustrative simulations described below, we use , i.e., there are 8 HVC(RA) neuronal ensembles but the model works robustly with more HVC(RA) ensembles as well (see Figure S1). A real bird brain has many more HVC(RA) ensembles but here we are interested in presenting a general mechanism for which a small selection of HVC and RA ensembles is sufficient. See the third level dynamics in Figure 5A for typical dynamics generated by this system.
We control the dynamics of RA ensembles by letting the kth HVC(RA) ensemble send a signal to the lower level during its activation time. See the next subsection for details of how this signal vector is computed. The total signal sent to the lower level by all HVC(RA) ensembles at any time is a linear combination of the 's: where is the output vector in Eq. (1). Note that for typical sequential dynamics at the HVC level, except for the transition times, only one entry in is active (i.e., only one entry is close to ), see Figure 5A.
Experimental findings suggest that activation of different HVC(RA) ensembles drives the activation of different combinations of RA ensembles [25]. In the present model, we capture this by forming a network of RA neuronal ensembles whose dynamics converge to one of several attractors depending on the input from the HVC level (see Figure 2). This means that the RA level receives input from the HVC level and produces output which encodes the level of activity of each RA ensemble at a given time. Since we are working with continuous systems, the notion of attractors comes up naturally as the RA ensemble activity flows from one activity pattern to another one. To achieve this smooth flow between RA attractors, we have to use a nonlinear network because otherwise the RA level would simply copy the dynamics of the HVC level. Note that, similar to the HVC level, the intrinsic neuronal dynamics of the RA are not established well experimentally. In this situation, we aim at describing underlying population dynamics which give rise to the experimentally observed key features of RA dynamics [25]. To implement these dynamics, we use a well-established type of an attractor-based network described by Hopfield [29]. Hopfield networks have been mostly used as a model of associative memory where each memory item is encoded by an attractor. When such a system receives noisy sensory input, i.e. it is started at some nearby initial state, it evolves to an attractor (the memory to be retrieved) [29], [36]. Here, we use this idea to encode the activities of RA ensembles by attractors. As the attractor of the network changes continuously due to driving HVC input, the activities of RA ensembles also changes such that some RA ensembles activate and some others deactivate. This gives us the spatiotemporal coding that drives the syrinx dynamics described in the next subsection. We use a Hopfield network with asymmetric connectivity matrices [37]–[39] given by the following equation:(2)where is the ensemble state vector with ensembles, is a diagonal positive matrix which governs the rate of change of each ensemble's state, is a synaptic connectivity matrix with entries denoting the strength of connection from ensemble to ensemble , is the activation function which we take as tanh function applied component-wise and is the direct input from the HVC level. This equation is similar to Eq. (1), i.e. both are continuous-time recurrent neural networks, but in Eq. (2) we have an additional input vector and different conditions on the connectivity matrix as described below. In addition, the use of the nonlinear activation function brings more plausibility to the network, as compared to linear dynamics, since the effect of one RA ensemble to another one does not increase linearly but saturates.
The input vector should be chosen such that RA ensembles get quickly attracted to a desired attractor. An attractor means that a subset of the RA ensembles are ‘active’ (taking the value ) while all other RA ensembles are inactive (taking the value ). The goal is to establish conditions for the network in Eq. (2) to have a globally asymptotically stable equilibrium point (a vector that makes the right hand side of Eq. (2) zero and attracts all the solutions regardless of the initial state). These conditions and the proper choice of for the desired attractor have been described in [38] and [40] (see Theorem 1 in Text S1).
Using this technique, we can employ a small number of RA ensembles to encode a larger number of desired attractors to control the lowest level, the motor output. Each HVC(RA) ensemble provides a different -vector to the RA level thereby driving the RA ensembles into a unique attractor. The application of this is that each RA level attractor will drive the motor output in a specific way thereby producing a different part of the song. We obtain the equations for the second level by combining the Hopfield network, Eq. (2), with the two output equations (state vectors) and where superscripts denote the specific level of a variable:(3)where the exact form of the connectivity matrices , and the HVC input vector are described in Text S1, is a scalar and are normally distributed noise vectors. is the normalizing function as in Eq. (1). Note that squeezes the entries of into the interval but may return values smaller than 1 since more than one entry of can be active () at a given time. The vectors and carry the output of the second level to the first level (oscillator level) as described in the next subsection.
In the present model, we use (i.e., five RA ensembles, Figure 5B). Note that there are different ways to activate RA ensembles to produce motor output. We use 7 of these 31 combinations (one occurring twice) in Figure 5 for generation of an example song (with 8 HVC(RA) ensembles at the higher level). In the figures, we used arbitrary units for both time (x-axis) and neuronal activation (y-axis) because we consider neuronal ensembles.
The avian vocal organ, the syrinx, is located at the base of the trachea (windpipe) where the trachea divides into the bronchi. A set of soft tissues within the syrinx, the labia, which are similar to human vocal folds, oscillate with the airstream propelled from the air sacs. Sound waves generated from these oscillations propagate through the trachea and beak. Therefore, these sound waves are modeled as the oscillations of the labia which are produced by the vocal control signals: the air sac pressure, , and the stiffness of the labia, . Such a mathematical model of the vocal fold oscillations was first given by Titze [41], and similar oscillations were experimentally observed in the bird syrinx [42]. A simplified version of this model (using a polynomial approximation for the nonlinear dissipation) can be given as follows [17]:(4)where is the position of the labia from the midpoint of the syrinx, denotes the air sac pressure, is the linear dissipation constant, is the stiffness of the labia and is a dissipation term to prevent the big amplitude oscillations when the labia meet each other or the walls of the syrinx [43]. The fundamental frequency of the sound wave increases or decreases proportional to . Note that there is a critical value for the pressure such that if , no phonation is produced. This region in the parameter space corresponds to the mini breaths between syllables [44]. Using this simple model, one can obtain accurate copies of some birdsongs such as canary [30], chingolo sparrow [17], white-crowned sparrow [31] and cardinal [45] by choosing appropriate vocal control signals for the syrinx ( and ) as described next.
Oscillators as in the present model have been widely used to model movement patterns in animals and humans. Central pattern generators are a well-known example of neural networks that are used to generate periodic motor commands such as locomotion [46]. We use the same principle here, and use five oscillators with different frequencies (one for each RA ensemble) to let the RA dynamics drive the vocal output (syrinx) mechanism, see Eq. (4). Note that it is experimentally not well established how the RA level controls the syrinx muscles; our approach is a natural extension of the phenomenological syrinx model described above [17]. The main point here is that the oscillator level (first level) is assumed to generate mixtures of oscillations (hidden states) where the RA level activity at the supraordinate level controls which oscillations should be produced at a given time. Each RA ensemble is assumed to control the activity of a single oscillator at the level below Therefore, the spatiotemporal coding of the RA level is transformed into the oscillatory activity of the first level which generates the final p(t) and k(t) dynamics necessary to control the syrinx.
As oscillators, we choose simple sine wave equations where the lowest frequency oscillator corresponds to the slowest-changing dynamics of the birdsong. We choose the remaining four oscillators such that their frequencies are integer multiples of this first oscillator's frequency (): , , and . Each one of these sine waves represents faster changing dynamics of the song; being the fastest. In this way, we can model effects in the birdsong which express themselves on different time-scales.
We include these five oscillators in the present model at the first level, where each of the five ensembles at the RA level controls the amplitude of one of the oscillators (through , Eq. (5) in Text S1). The observable output is obtained by taking a linear combination of these amplitude-modulated sine waves. To drive the vocal model appropriately, we produce two outputs and (the second output is simply a time-shifted copy of the first one), which are involved in producing air sac pressure p(t) and the stiffness of the labia k(t). and are described in detail in Text S1.
Laje et al. [31] chose and to form several ellipses in the parameter space where each ellipse corresponds to a different syllable. However, this parameterization may not support complicated syllables which have more fluctuations on the sonogram. Here, we extend their model to increase the complexity of the generated songs by using the linear combination of different frequency sine waves ( and described above) to parameterize these two functions and obtain a variety of ellipse-like curves in the parameter space (see Figure 3):where and are the outputs of the first level and the scalars are given in Table 1. These ellipse-like curves can be plugged into Eq. (4) to obtain synthetic birdsongs. See Figure 6 for the sonogram obtained using the first level output of the generation process shown in Figure 5C. The sonogram can be played and is reminiscent of a birdsong (Audio S1). Note that in the real system, longer HVC(RA) sequences would be required to produce a song with 6.5 seconds duration since HVC(RA) bursts last only about 6–10 ms [25]. Here, we assume that each HVC(RA) ensemble in the model is a collection of at least 80 HVC(RA) neurons that fires sequentially and controls the timing of the song for about 800 ms.
In this subsection, we will briefly describe the present recognition scheme for the generated songs. This scheme is a model of vocal communication between conspecific birds but may also serve as a functional model to explain experimental findings along the auditory processing pathway which is less understood than the song pathway. Here, we describe a potential mapping of this Bayesian inference framework to neuronal dynamics at a population level, see [47], [48]. The inference is based on hierarchical message passing and implements a predictive coding scheme for dynamics. As summarized below, all the update equations of the recognition system (to reconstruct the hidden states) consist of differential equations (as in the generation model) and therefore may be implemented by neuronal populations and their network interactions via forward, backward and lateral connections [47], [48].
How can a bird recognize a conspecific's song and decode the information contained in the song? This decoding is important as it is known that female birds select their mates according to criteria such as the complexity of the male's repertoire [7] or the precision of the vocal performance [8] and they show preference for the songs of their mates or fathers compared to the songs of strangers [4], [49], [50]. In general, this suggests that listening birds may have certain expectations (priors) about the type of the song they expect to hear. In general, we assume that listening birds have internal models for the songs they have learned before and the generative model of the heard songs should fit to this internal model.
Using this concept, we model optimal recognition using Bayesian inference for hierarchical, nonlinear dynamical systems [47].
For the sensory input, we assume that the vocal control signal , given the sound wave, can be readily extracted by the listening bird (agent) from the spectrotemporal dynamics, see Figure 3. Here, we consider the p(t) and k(t) dynamics, in the recognition step, as an abstract representation of the song spectrum and therefore a phenomenological approximation to the highly nonlinear features of the singing bird's syrinx. This means that we assume that the listening bird has access to these dynamics via some low-level recognition process. For the present implementation of the inference framework, the full inference from the soundwave (Figure 7) would currently be computationally too expensive because this would require a high temporal resolution, e.g. at 12 kHz, and long time-series. However, once an optimized (parallel) implementation of the present framework becomes available, the present model can be extended in a straightforward fashion to model recognition that receives a soundwave as sensory input by adding another level that transformed the p(t) and k(t) dynamics to soundwaves.
Given this vocal control signal, we infer the spatiotemporal RA dynamics and the sequential HVC(RA) dynamics. The proposed Bayesian inference scheme provides, under some assumptions, optimal inference to decode the RA and HVC(RA) dynamics, i.e. to recognize the hidden messages embedded into the vocal control signal.
The mathematical description is provided below and can be conceptualized as follows: At each time step t, the recognition system receives sensory input, here the current amplitudes of the p(t) and k(t) dynamics. Like the generative model, the recognition system has three levels as well. Each of these three levels consists of interacting neuronal populations, which encode predictions, i.e. expectations, about how their internal dynamics will evolve during a song. At the same time, each level receives input from the subordinate level. For the first level, this is the sensory input, which is compared with the internal prediction. The prediction error is forwarded to the second level, where again predictions are used to generate prediction errors, which are forwarded to the third level. Critically, each level adjusts its internal predictions to minimize its prediction error weighted by the prior precision of the internal prediction. At each level, the updated predictions are sent to the subordinate levels to guide their internal predictions by higher level predictions. In summary, each level minimizes its prediction error by a fusion of internal dynamics with top-down (predictions) and bottom-up (prediction error) messages. The overall result is that a listening bird fuses its dynamic and hierarchically arranged expectations about a song with the actual sensory input. Importantly, due to this dynamic fusion, the recognition is robust against deviations from its expectations by explaining away errors of the singing bird by internal precision-weighted prediction error. The derivation of the update equations to achieve Bayes-optimal online recognition solutions is non-trivial, see Friston et al. [11]. Note that this modeling approach implies that generation and recognition models are fundamentally different from each other in the sense that generation is a top-down process where recognition consists of both top-down and bottom-up processes. Although some of the computations in the generation and recognition model are the same and may provide a computational explanation for mirror neuron accounts [51], this is not a central issue in the present paper and we assume here that recognition is performed by neuronal populations different from those that generated the song. Clearly, this remains an open question that can only be settled experimentally.
For sensory input and a given model , the probability is called the model evidence or marginal likelihood of and is an important quantity for model comparison among different models. In our case, is the vocal control signal for the syrinx which we take as the input and the model (Figure 4) includes all the parameters and equations together with causal and hidden states at all levels. We take to be the set of all hidden states and causal states at all levels of hierarchy. The task for the agent is to infer the states from the sensory input under model m. We assume that the parameters (such as , and , see Figure 4) have been learned previously by the listening bird and are fixed (Table 1).
Our goal is to approximate the posterior density which will give us both the posterior mean of the dynamical states and the uncertainty about this mean. To get a good approximation for the posterior density, we follow a rather indirect way using the marginal likelihood.
The marginal likelihood of can be written as . Here, is defined in terms of the likelihood and the prior . Except for a few analytical cases, this integral is usually intractable and needs to be approximated. One way for this approximation is to introduce a free-energy term which is a lower bound for the marginal likelihood. It is not hard to show that:where is the free-energy, is the Kullback-Leibler divergence and is the recognition density. Note that is an auxiliary function that we will use to approximate the posterior density. It is easy to show that , and if and only if . This means is a lower bound for , and if we can maximize , this will minimize giving an approximation for the posterior density.
To maximize with respect to , we make the assumption of normally distributed error terms and write where consists of the mode and the variance . Then the problem turns to a maximization problem of the free energy with respect to :which gives the approximation for the posterior density . For the details of this variational process and its extension to time-dependent states, see [11].
Since we apply the variational scheme in a hierarchical setting, we write the equations in our model (see Figure 4) in a generic hierarchical form [11]. We use the same set of equations as in the generative model since we assumed the singing and listening birds have the same internal models. We denote all hidden and causal states at level by and , respectively. In particular, stands for all the and outputs of the th level. We also write and to describe the dynamics of the hidden and causal states in the th level:where denotes the normally distributed fluctuations at the th level. The present model shown in Figure 4 follows this generic form. The causal states () provide input to the subordinate level while the hidden states () are intrinsic to each level.
Note that the Gaussian fluctuations in the above hierarchical form quantify different amounts of noise at each level of the singing bird. We list the covariance matrices used in the “Ideal Communication” simulation in Table 1. Note that sensory input enters the recognition system at the first level: . The optimization process of (i.e. the estimated mode of causal and hidden states) can be implemented in a message passing scheme [11] which involves passing predictions down and passing prediction errors up from one level to another. Prediction errors can be written aswhere and denote the predictions from level above for and , respectively. In this scheme, is optimized through gradient descent on prediction errors at each level of the hierarchy. Importantly, the computations required for this gradient descent could be implemented by interacting neuronal populations at each level: Each population comprises causal and hidden state-units that encode the expected states and the error-units, with one matching error-unit for each state-unit, which encode the prediction errors. The estimated mode of the states, i.e. , is described by the activity of the state-units. The error units compare the estimated modes with predictions sent via backward and lateral connections and compute prediction errors, which are passed on via forward and lateral connections. This message passing has been shown to minimize precision-weighted prediction errors and optimize predictions at all levels efficiently (see [47], [52] for further details).
Software Note: The routines (including commented Matlab source code) implementing this dynamic inversion, which were also used for the simulations in this paper, are available as academic freeware (Statistical Parametric Mapping package (SPM8) from http://www.fil.ion.ucl.ac.uk/spm/; Dynamic Expectation Maximization (DEM) Toolbox).
To illustrate the behavior of the described generation and recognition schemes, we exposed our recognition model to four different tasks. Since the neuronal structures for song generation and song recognition are mostly different (see Discussion), we refer in the following to the levels in both the generation and recognition models as the first, second and third levels instead of ‘Oscillator’, ‘RA’ and ‘HVC’ levels, respectively.
We first show the case of ‘ideal communication’, i.e. the recognition scheme described above can appropriately infer about the states at all three levels from sensory input that describes a veridical song. In a second simulation, we show the case when the sensory input is not as expected, i.e. when, for the listening bird, there is an unexpected deviation in the song (a single syllable). We will demonstrate how the listening bird detects this deviation and what neuronal correlates are observed in presence of this deviation. In the third simulation, we show that the recognition mechanism is robust against differences in the anatomical connectivity pattern in the second layer. This robustness is a consequence of the hierarchical setup of the generative model. This is an important finding because it explains how different birds can decode the same song although their individual anatomical connectivity within some layers may differ. In our final simulation, we replicate the experimental findings of a study [53] where the authors cooled HVC and observed that the song slowed down. We also show how the listening bird (e.g., female bird in a social context) can detect the minor deviations due to a speed change of the song.
Here, we simulate the ideal situation in which both the ‘singing bird’ and the ‘listening bird’ have learned how exactly a song should sound. As before, we use eight third level ensembles that are each activated sequentially and, during this time, they control the activities of five second level ensembles (Figure 2). The third level imposes a sequence of attractors on the second level which in turn produce linear combinations of appropriate sine waves to produce the air sac pressure and labia stiffness, see Figure 4. To introduce noise (both internal state noise for the singing bird, and also transmission noise to the listening bird), we used normally distributed zero-mean noise with standard deviation of and at all levels. To show that recognition is robust against starting condition (i.e. the state of the ongoing neuronal activity within the bird brain at song onset), the initial states of the recognition are chosen differently from the true initial values used in the generation. As expected, we find that the listening bird starts tracking the sensory input very quickly and follows it robustly during the remainder of the song, see Figure 7.
Next, we show what happens if the listening bird has a different expectation than the singing bird about how a song should sound. In the generative model (singing bird), we use the same third level ensembles and the corresponding second level combinations that we used in the ‘Ideal Communication’ case (Figure 2). However, the recognition system (listening bird) knows a slightly different song where there is a deviation in a single syllable. We model this by changing the effect of the third ensemble at the third level such that it activates only the first ensemble at the second level (instead of the first and fourth as in the singing bird). This means that the motor output and the sonogram look different from the prior expectation of the listening bird but only for the third syllable, see Figure 8. The internal recognition dynamics of the listening bird register this deviation and show two effects during the third syllable, between time points and : (i) Prediction errors in the recognition are distributed throughout all three levels and are not only explained by changes at a single level (Figure 9). This makes sense since the observed deviation at the first level cannot be explained by the simple oscillatory first level dynamics. Rather, the recognition attempts to explain away the deviation at the first level by using prediction error at the second and third level as well. At the first level, this is quite successful because the recognized dynamics look very similar to the generated dynamics (see Figure 8, bottom row). However, at higher levels, there are obvious differences between the generated and recognized dynamics, i.e. the listening bird can infer a deviation via the prediction error at the second and third levels. (ii) When the deviation has finished, the recognition quickly locks back onto the ongoing song dynamics at all three levels and decodes the song veridically. In summary, this simulation shows that the dynamic recognition hierarchy uses all its levels to compensate for unexpected deviations in the song. This means that all levels of the hierarchy work together in concert to minimize the effects of deviations throughout the hierarchy. In other words, the activity of high-level auditory processing levels in songbirds in response to small deviations in the expected song may be most revealing for their function. This mechanism may be important in social context since the listening bird can recognize subtle variations in the singing bird by its activity in high-level areas and grade the singing bird's overall performance [54].
Considering the anatomical complexity of the brain, genetic and developmental variability is expected in the brains of individuals of the same species. At the macro scale, the general connectivity structure of distinct brain regions may be shared, but at the micro scale, variability is found in size, location and connections between individual neurons or neuronal ensembles [55]–[58]. Here, we simulate a difference in the connectivity structures by using different second-level connectivity matrices W (Figure 4 and Eq. (3)) in the generative model of the singing bird and the recognition system of the listening bird. In other words, the listening bird has a different internal model at the second level as would be prescribed by the generative model of the singing bird at the RA level. How can birds with individual variability in their internal models still extract the same information from a song?
The answer is that differences in the second-level connectivity matrix W can be compensated by a different driving activity I from the third level since I depends on W (see Theorem 1 in Text S1). In our simulation, we assume that these driving activities have already been learned in the corresponding birds, e.g. during juvenility. As shown in Figure 10, the states at all three levels can be recognized successfully even though the second levels in the two birds are wired differently. This means that the internal models of generation and recognition do not have to be the same but can cope with structural variations due to anatomical variability at the micro-scale. Critically, this compensation of anatomical variability at the second level relies on the hierarchical configuration and learning of the connectivity from the third level to second level.
In song generation, a critical question is which regions of the brain are involved in the timing of syllables or sub-syllable structures. A recent study tackled this question by manipulating the temperature of the HVC and RA regions in the singing bird [53]. Importantly, it was shown that song speed at all time scales slowed down but the acoustic structure stayed the same as the temperature of HVC dropped. In the sonogram, this corresponds to a temporal stretching of the song. Conversely, cooling of RA did not have any effect on the timing of the song. This suggests that HVC is involved in the control of the timing of the song [53].
We observed similar behavior in our model where we modeled the cooling by manipulating the rate (i.e. speed) constants and at the three levels. Importantly, changing the rate constant for HVC slows down the song but changing the rate constant for RA does not. In the first simulation (Figure 11, left), we ‘cooled’ HVC by changing from to . This slows down the dynamics of the HVC level and immediately slows down the RA level as well since the control signals coming from HVC now last twice as long. In other words, we find as in the cooling experiment that HVC, due to its position at the top of the hierarchy, controls directly the timing of the song. To reflect this slowing down in the output we also changed from to ( is kept constant in all simulations) to adjust the frequencies which were chosen independently from the RA level for simplicity ( where ). In the second simulation (Figure 11, right), we changed the rate constant of RA, , from to . This has no observable effect, as in the experiment [53], on the dynamics of RA ensembles since the timing of attractor activations is controlled by the timing of HVC. A change in only slows down the transition times which has no detectable effect in the output.
Speed changes may not only have an experimentally observable effect in the generated song but also in the listening bird. Interestingly, speech changes in song also occur under natural conditions, e.g. in a social context: Male birds sing slightly faster when addressing a female bird (directed song) compared to singing towards other males or when alone (undirected song) [6], [59]. Using the present model, we tested whether the listening bird can detect such small changes in the singing bird during directed song. We slowed down the song by 3%, thereby modeling an undirected song, and analyzed the prediction errors in the listening bird which expected the slightly faster, directed version. The listening bird was able to recognize the song successfully but it also reliably distinguished the subtle change in the tempo, as can be seen from the sustained prediction errors at all three levels (Figure 12).
We have described a hierarchical model for generating birdsongs and introduced an online Bayesian inversion as a recognition model. The key result is that the specific anatomical, functional and hierarchical structure of birdsong generation enables Bayesian online decoding of hidden information at a slow time-scale at the HVC and RA levels. Four simulations showed that the Bayesian recognition mechanism works efficiently in several settings and its functional behavior might be helpful to understand the mechanisms of birdsong recognition. In addition, recognition is robust to noise and can be performed online. Overall, this is a unified modeling approach which handles both generation and recognition of birdsong and may serve as a model for vocal bird communication.
Both generation and recognition models extend previous modeling work either by using novel techniques (e.g. Bayesian inference for hierarchical, stochastic, nonlinear dynamical systems) or by combining well-known nonlinear differential equation systems in a novel way (generative model). The model explains recognition of birdsong as continuous message passing scheme among auditory areas and explains the dynamic song recognition system of birds using Bayesian techniques. In the generation model, we combine a well-established syringeal model with the sequential HVC/RA model and describe a hierarchical and dynamical mechanism which transforms the spatiotemporal coding at the RA level into the rich, complex structure of the song power spectrum. Based on this generative model, we use Bayesian inference to model song recognition by a conspecific. This modeling strategy is a novel approach to employ experimental findings in birdsong generation for establishing a functional model of birdsong recognition. In fact, decoding of sensory input generated by hierarchical, nonlinear dynamical systems is usually technically challenging and often impossible [60], [61] because the sensory input may not be informative about hidden information at higher levels. However, here we found that the decoding of birdsong using hierarchical Bayesian inference based on a song generation model is feasible, robust and can be performed online. Intuitively, it may be obvious that birdsong must be generated such that conspecifics can derive information (meaning) from it. The question is how birds do this mechanistically. Here, we propose that this recognition mechanism may rest on Bayes-optimal inference given the specific hierarchical arrangement of the neuronal birdsong-generating network.
We have derived a recognition scheme using Bayesian inference. However, bird brains may have established their recognition capabilities by evolutionary processes [4], [49], [62]. What are the similarities between the proposed recognition scheme and the biological one?
Note that the present modeling does not suggest that the areas involved in generation and recognition are the same. Many computations during recognition are different from those in generation. The present recognition scheme consists of three hierarchical levels, thereby mirroring the hierarchical generation system. We found that three hierarchical levels are also appropriate for the recognition of a song. Interestingly, experimental findings point to a hierarchical arrangement of the auditory system in songbirds as three major functional levels of processing [63], [64] where it is partially unclear yet how this hierarchy maps exactly onto the auditory system. Moreover, note that these areas are mostly investigated for male (zebra finch) birds and it is quite possible that there could be different areas involved in females or in other bird species.
Experimental evidence suggests that HVC may be located at the highest level of this recognition system. In particular, HVC(X) neurons (HVC neurons that project to Area X, see Figure 1) are selectively responsive to the bird's own or a conspecific's song [65], [66]. The firing of HVC(X) neurons at temporally precise times during an auditory stimulus [65] is similar to the temporally precise activation of HVC(RA) neurons during singing. This suggests that HVC(X) neurons may be involved in the representation of the expected sequence of song dynamics. In the present model, the third level encodes both the sequence prediction but also the perceived deviation from this sequence.
The circuitry of areas subordinate to HVC during song recognition is not particularly well understood. The caudal mesopallium (CM) and caudomedial nidopallium (NCM) have been shown to be selective for particular familiar songs or sounds and are involved in auditory memory [23], [63], [64]. Similar functions are implemented by the second level of the present recognition model: The second level encodes the expectation of specific spatiotemporal patterns, i.e. it encodes auditory memory by attractors that correspond to specific vocal tract dynamics (sounds). Note that there is a clear distinction between the third and second level in the model: While the third level encodes the expected sequence of sound dynamics, the second level encodes the repertoire of song sounds (transcribed to sound waves by the vocal tract dynamics). This functional separation is also assumed to be implemented in the real bird brain [26].
In the primary auditory area, Field L, spectral-temporal receptive fields (STRF) have been proposed to explain the selective responses of neurons [67]. These selective responses may correspond to the recognition dynamics at the first level in the model which decodes the detailed spatiotemporal structure of the auditory stimulus guided by higher level predictions. It is interesting to note that we could use the present recognition model to derive, as done experimentally [67], the spectral-temporal receptive fields at the first level. Alternatively, one could use experimentally acquired STRFs to adapt the first level of the present model to establish exact equivalence of the model and the real system at the level of primary auditory areas.
There are several models that focus on the sequential activation of HVC(RA) neurons using single neuron models. Inhibition is believed to be a key element to generate rhythmic (sequential) activity in HVC [15], [16], [68]. We used winnerless competition which relies on inhibition to sequentially activate HVC(RA) ensembles. A similar generation mechanism as described here can be obtained using the synaptic chain scheme: Li and Greenside [12] proposed a conductance-based model for HVC(RA) neurons from which they obtained sequential multi-spike bursts. Later, Jin et al. [13] used an intrinsic bursting mechanism to obtain higher firing rates more consistent with the experimental data. This scheme was extended in [14] and was shown to produce robust and highly stereotyped sequential bursts. A learning mechanism was proposed in [69] showing how a sparse temporal code can emerge from a recurrent network. The models mentioned above focus on describing possible ways for the sequential activity of HVC where the downstream areas can be regarded as driven in a feed-forward fashion by HVC. A comprehensive generative model that includes HVC, RA and motor control areas was described in [18]. This study showed that the intrinsic connectivity at the RA level can substantially influence the acoustic features of syllables. This approach is similar to the present where the common research question is which parameterization (connectivity) of a recurrent neural network will generate motor control signals that result in realistic acoustic features of birdsong. However, we additionally incorporated recent findings [26] which point to a specific role of RA ensembles in encoding sound wave modulations. Furthermore, we provide evidence that the hierarchical setting of HVC and RA ensembles is the basis for robust and rapid song recognition.
Theunissen et al. [70] estimated spectral-temporal receptive fields (STRF) of nonlinear auditory neurons using natural sounds as sensory input. The STRFs describe which temporal succession of acoustical features would elicit the maximal neural response and provide useful information for modeling perception of acoustic features, e.g. in the primary auditory area, Field L [67]. A two-level model was introduced [71] where the first level encoded frequency responses identified by an STRF analysis and the second level used these features to model song selective responses of HVC neurons. In another approach, Larson et al. [72] proposed a model for auditory object recognition where the first level uses a distance metric to distinguish between different spike trains and the second level acts as a decision network. However, both of these models propagate auditory signals in a feed-forward fashion from the low to the high level while the present scheme uses dynamical and recurrent bottom-up and top-down message passing thereby providing a more comprehensive model of the neuronal dynamics observed during song recognition.
Learning models such as [73] and [74] were proposed which also include birdsong production and evaluation. These models mainly focus on the neural mechanisms of learning but they also provide mechanisms for song evaluation.
There have been also attempts for the automated recognition of birdsongs using machine learning methods, e.g. [75], [76]. However, these models are not concerned with neurobiological plausibility but rather use ad-hoc techniques as used in automated speech recognition, i.e. hidden Markov models and template-based matching of song syllables.
There are several implications for future experiments which one can derive from the present model. The first is that we observe prediction errors at all levels when there is an unexpected piece of song (Figure 9) or a song which is slower than expected (Figure 12). This suggests that there may not be a single area in the auditory pathway (such as HVC(X) or LMAN in the anterior forebrain pathway) that acts as a comparator between the stimulus and previously memorized tutor song [77] but several levels of the auditory pathway may be involved in this comparison. Comparing the neuronal recordings from a bird that listens to a normal speed song and a slower version of the same song might reveal the locations where these prediction errors are computed. Similar experiments have been done in auditory areas Field L and caudal lateral mesopallium (CLM) where some neurons responded robustly to perturbations in vocalization or playback of the bird's own song [21]. A functional model like the one presented here could predict what amount of activity should be expected in experiments given defined deviations, at different levels of the recognition hierarchy. Parallel to this idea, a recent experiment explained the activity in CLM by the level of surprise in the stimulus [20]. Our model could be used to predict the amount of surprise or prediction error at different hierarchical levels. As the present model covers much of the auditory pathway, this prediction technique may be best suited for using functional MRI on birds [78], [79] where one would model increased activation, relative to some baseline condition, as an increase in prediction error.
As noted by several authors, human speech and birdsong have in common that both are complex, hierarchical, sequenced vocalizations which are repetitions and combinations of simple units such as phonemes and syllables [2], [80], [81]. Although human speech is far more complex than birdsong, the underlying anatomical and functional features show striking similarities such as the pathways for vocal production, auditory processing and learning [22], [81]. Songbirds, similar to humans, gain their vocal abilities early in life by listening to adults, memorizing, and practicing their songs [22]. These similarities suggest that one may derive insight about human speech recognition and learning from findings in birdsong research [82].
The present results clearly point to the usefulness of a hierarchical recognition structure to decode sequences of syllables. Such hierarchical models are rarely used in automated speech recognition [83] presumably because the standard model, the hidden Markov model, is mathematically best understood only in a non-hierarchical setting. The present scheme shows that complex spectral dynamics such as birdsong may be modeled as a sequence of nonlinear dynamics, where, in the generative model, each level drives the subordinate level in a highly non-linear fashion. To invert such a hierarchical, nonlinear, dynamical system, one requires sophisticated Bayesian inference machinery [11], [84]. We described such a mechanism previously for a simple auditory sequence of sounds [85]. The novelty of the current approach is that we use a neurobiologically plausible generative model to derive a functional recognition model that has also the potential to recognize real and complex birdsong. In addition, we hypothesize that the specific arrangement of HVC and RA level (dynamic sequences driving attractor dynamics at a lower level) and its Bayesian online inversion will not only play a role in birdsong recognition models but may be successfully used for automated speech recognition as well.
The mathematical model that we used to generate birdsongs was previously shown to produce accurate copies of songs such as canary [30], chingolo sparrow [17], white-crowned sparrow [31] and cardinal [45] songs. The vocal organs of other birds, e.g. of the zebra finch, can generate highly nonlinear, more complex, acoustic dynamics than the one considered here. For modeling such songs, one would have to replace the syrinx model of Eq. (4) by a more involved syrinx model such as the one reported in [86].
For our purposes, we focused on one particular song to describe the generation and recognition framework. The recognition of different songs either by the same or different conspecifics could be modeled by using multiple sequences encoded at the third level, where we assume that the recognition will converge to the best fitting sequence. In addition, one could adapt the nonlinear syrinx model to endow a singing bird with its own low-level acoustic characteristics.
In the present model, we used rather small numbers of ensembles for visualization and computational purposes. The generative model applies to an arbitrary number of ensembles and similar type of dynamics can be obtained with larger number of ensembles at each level (see Figure S1 for generation with 100 HVC ensembles). For recognition, we performed similar experiments with larger numbers of HVC ensembles (32) and RA patterns (24) where the recognition results were as robust as with the reported smaller size models (see Figure S2 for the simulation). This indicates that the model scales to larger model sizes. However, there are two main issues that one will need to address to enable recognition using hundreds of units: (i) The computational power required for the recognition quickly increases with the number of ensembles used (with complexity due to computing a matrix exponential, see [11]). This can be resolved by parallelizing the ensemble-specific computations which would be a further step towards biological reality. Currently, we emulate these parallel computations using a single-process Matlab implementation. (ii) The complexity of the syrinx model must be matched by the ‘descriptive power’ of the RA level. In other words, if one wanted to increase the number of RA ensembles significantly, one also had to render the model at the syrinx level more complex so that the recognition can infer more RA ensembles from more complex sensory data. However, this increase in model complexity at the syrinx and RA levels would require a more sophisticated syrinx model and is beyond the scope of the present work, in which we provide a proof of concept and introduce the computational framework.
Furthermore, we tested the sensitivity of the Bayesian recognition in response to changing specific details of the generative model: (i) We used higher noise levels (standard deviation of and ) as compared to the simulations above, the recognition still robustly inferred the hidden states and causes at all levels (see Figure S3) (ii) We found that the recognition is robust against varying the initial conditions of the states in both the generative model and recognition. We tested a wide range of random initial conditions in both generation and recognition and observed that in all simulations the recognition quickly locks into the necessary dynamics. This implies that the listening bird can recognize a song reliably whatever the initial state of itself or the singing bird at the beginning of the song. (iii) We also changed the connectivity matrices at the third level (with the constraint of high inhibition from the previous neuron and low inhibition to the next neuron) and at the second level (with the constraint that global stability conditions are satisfied, see Theorem 1 in Text S1) of the generative and recognition models. The recognition was still robust with these different connectivity matrices (see Text S1 and Figure S4).
We described a model to generate artificial birdsongs and a scheme for their online recognition. We constructed a model based on key experimental findings in birdsong generation. Our results show that the specific, hierarchical mechanism how birdsong is generated enables robust and rapid decoding by a hierarchical and dynamic Bayesian inference scheme. We have interpreted this as evidence that the birdsong generation mechanism is geared toward making the song robustly decodable by conspecifics and discussed the experimental evidence that songbirds use a recognition mechanism similar to the present Bayesian inference scheme.
|
10.1371/journal.pgen.1003600 | Latent Effects of Hsp90 Mutants Revealed at Reduced Expression Levels | In natural systems, selection acts on both protein sequence and expression level, but it is unclear how selection integrates over these two dimensions. We recently developed the EMPIRIC approach to systematically determine the fitness effects of all possible point mutants for important regions of essential genes in yeast. Here, we systematically investigated the fitness effects of point mutations in a putative substrate binding loop of yeast Hsp90 (Hsp82) over a broad range of expression strengths. Negative epistasis between reduced expression strength and amino acid substitutions was common, and the endogenous expression strength frequently obscured mutant defects. By analyzing fitness effects at varied expression strengths, we were able to uncover all mutant effects on function. The majority of mutants caused partial functional defects, consistent with this region of Hsp90 contributing to a mutation sensitive and critical process. These results demonstrate that important functional regions of proteins can tolerate mutational defects without experimentally observable impacts on fitness.
| Changes in protein sequence or expression strength can both lead to adaptation in natural systems. While many studies have focused individually on either expression strength or protein sequence, in principle the fitness effects of these two protein properties are interdependent. We systematically investigated the fitness effects of both expression strength and protein sequence for the yeast Hsp90 gene (Hsp82). We analyzed the fitness effects of all possible point mutations in a putative substrate binding loop under seven different expression strengths. The fitness effects of amino acid substitutions were strongly dependent on expression strength. Many point mutations exhibited fitness defects at reduced expression strength that were hidden at the natural expression strength. Revealing these hidden mutant defects suggested that this region of Hsp90 contributes to a rate-limiting step in function, consistent with its putative role in substrate binding. This study is important because it indicates that critical regions in proteins are more prevalent than would be estimated based on experimental fitness analyses performed at natural expression strengths. As hidden fitness effects are likely to occur in other systems, these findings have broad implications for the field of experimental evolution.
| Genetic changes that alter protein sequence or expression level can lead to adaptation, suggesting these protein properties are central to evolutionary processes. Many studies have individually investigated the effects of changes to either protein sequence or expression level. For example, protein sequences have been optimized under selective pressure using in vitro evolution [1]. In addition, changes in protein sequence relative to synonymous substitutions are a hallmark of positive selection in natural populations [2], [3]. The influence of protein expression level on fitness has also been well documented [4]. For example, changes to the expression level of the Agouti protein (but not its sequence) have been shown to affect fitness in wild mice by modulating coat coloration [5]. In addition, experiments in E. coli demonstrate that expression from the lac operon is rapidly tuned for optimal growth over a wide range of lactose concentrations [6]. While most studies to date have focused individually on either expression level or protein sequence, in principle the fitness effects of these two protein properties are interdependent [7], [8]. Here, we systematically investigate selection on the sequence and expression level of yeast Hsp90 (Hsp82).
We recently developed an approach termed EMPIRIC [9], which is a genetic screen that provides fitness measurements of all possible amino acid substitutions in short regions of important genes in yeast. By sampling across the variety of different amino acid substitutions, EMPIRIC provides detailed information about the physical constraints on protein function. We previously reported a bimodal distribution of fitness effects (DFE) for an evolutionarily conserved region of the yeast Hsp90 gene[9], an essential chaperone required for the maturation of many kinases [10]–[12]. Bimodal DFEs, where most mutants have fitness effects close to either null or wild type (WT), appear common in nature as they have been observed in many other fitness studies [13]–[17].
Bi-modal DFEs are consistent with a recently proposed model where the impacts of mutations on protein stability have a dominant impact on fitness [18]. This model is founded on two concepts: positions that contribute directly to rate-limiting steps in protein function are rare; and the natively folded structure is required for function. Under these conditions, selection results in stably-folded proteins [19], [20], such that modestly destabilizing mutations can be tolerated without dramatic changes to the fraction of natively folded protein molecules and hence function. Because protein folding is cooperative there is a narrow range of stability where both the folded and unfolded state are highly populated, consistent with relatively few mutations having intermediate function. In this stability-dominated model, mutations to critical functional positions (e.g. catalytic sites in enzymes) destroy activity, but are presumed rare and so do not contribute greatly to the DFE. Of note, the prevalence of positions in proteins that directly contribute to rate-limiting steps in protein function and the fragileness of these positions to mutation have not been thoroughly investigated.
The effects of mutations on protein function can be investigated based on fitness effects; however, fitness effects need not correspond directly to functional effects. For example, many essential proteins can be dramatically reduced in net function (defined here as the product of expression level and function per molecule) without dramatic reductions of fitness [13], [21]–[25]. Heterozygotes with one null allele are often highly fit, indicating that 50% reductions in net function can be tolerated [26]. The relationship between fitness and the net function of a protein is formally an elasticity function [21]. Around the wild type net function, the elasticity function often has a slope less than one indicating that reductions in net function have dampened impacts on fitness [27], [28]. Experimental analyses of fitness effects are also constrained by experimental measurement precision, which is currently on the order of 1% [29]. In natural systems, the resolution of selection depends upon the inverse of effective population size and is on the order of 10−7 for yeast [30], [31]. Thus, the effects of mutations on function that are important in natural selection can be hidden to experimental fitness analyses. For example, the net function of lysozyme in phage T4 must be reduced about 30-fold before experimentally measurable impacts on growth are observed [13]. At the endogenous expression level in this system, large defects in per molecule function are hidden to experimental fitness analyses.
We searched for hidden fitness effects in Hsp90 by examining the Hsp90 elasticity function. We varied the expression level of the native protein sequence and monitored effects on yeast growth rate. Determining theHsp90 elasticity function enabled us to estimate mutant effects on per molecule function from fitness measurements. The elasticity function was non-linear such that at the endogenous expression level, mutant defects up to 79% in per molecule function were hidden to experimental fitness analyses. To reveal potentially hidden functional defects of mutants, we repeated EMPIRIC analyses at reduced expression strengths, which systematically varied fitness sensitivity to amino acid substitutions in Hsp90. Using this approach, we were able to construct a full distribution of mutant effects on function for a region of Hsp90. Structural analyses suggest that the region we chose to analyze is a putative substrate binding loop [32]. Our experimental fitness analyses at the wild type expression level resulted in a bimodal DFE, which is a hallmark of a scaffolding region with stability dominated effects on fitness [18]. By analyzing fitness at varied expression strengths, we found that the majority of Hsp90 point mutants had intermediate (10–90%) defects in per molecule function that were hidden to our analyses at wild type expression level. These observations indicate the region of Hsp90 we analyzed is involved in a rate-limiting step in function, and supports its putative role in binding to substrates [32].Because many mutant defects may be hidden to experimental measurement at the wild type expression level, our results suggest that rate-limiting functional sites in proteins may be more prevalent than previously appreciated, and provides a useful guide for interpreting the growing field of systematic mutant analyses [9], [33]–[40].
While our initial EMPIRIC study [9] was performed with a temperature sensitive allele of Hsp90 co-expressed with all mutants; here, we report results in an Hsp90 shutoff strain where mutants were analyzed without potential co-expression artifacts. We developed a yeast shutoff strain (DBY288) where the only chromosomal copy of Hsp90 is regulated by a strictly galactose-dependent promoter [41]. In galactose media, the DBY288 strain expressed Hsp90 at endogenous levels and grew robustly. When switched to dextrose media, the DBY288 strain stalled in growth with Hsp90 levels rapidly dropping below detection (Supplementary Figure S1). This strain enabled plasmid encoded Hsp90 variants to be maintained and amplified under non-selective conditions (galactose media). Switching to dextrose media then applied selective pressure on the plasmid encoded Hsp90 variants.
We analyzed the fitness effects of Hsp90 point mutants by performing a bulk competition in the DBY288 strain. A library of plasmids containing all possible single codon substitutions at amino acid positions 582–590 (Figure 1A) was transformed into a single batch of yeast. These experiments used a plasmid and promoter construction previously shown to match the endogenous expression level of Hsp90 [42]. Transformed yeast cells were preferentially amplified in galactose media that allowed all mutations including null alleles to propagate. The bulk culture was transferred to shutoff conditions to initiate selection on the mutant library. The beginning of strong selection on the mutant library was estimated from the growth plateau of control cells harboring a null rescue plasmid (Supplementary Figure S1). After the initiation of selection on the mutant libraries, samples were harvested over the following 36 hours and the relative abundance of each mutant quantified using focused deep sequencing. By comparing the trajectory of each mutant relative to wild type, we directly determined competitive advantage or disadvantage of each amino-acid substitution as an effective selection coefficient (s) that represents the competitive asexual growth advantage/disadvantage of each mutant in a defined environment [29]. We have previously demonstrated that the EMPIRIC approach provides highly reproducible measures of fitness effects that strongly correlates with the growth rate of individual mutants grown in monoculture [43]. Consistent with our previous work, effective selection coefficients were highly reproducible (R2 = 0.96) in a full experimental repeat (Figure S2). At the endogenous expression strength, the distribution of fitness effects for this region of Hsp90 was bi-modal (Figure 1B, Supplementary Table S1), with peaks near wild type and null. Bi-modal fitness distributions are predicted based on a model where fitness effects are dominated by the impact of mutations on protein stability [18]. Thus, our fitness analyses at wild type expression level are consistent with this region of Hsp90 serving a primarily scaffolding purpose.
To further probe the relationship between the net function of Hsp90 and fitness, we varied expression level of the WT sequence and analyzed impacts on growth rate (Figure 2). To vary expression level, we swapped both promoter and terminator (3′ untranslated) sequences. Closely following the start of strong shutoff selection (19 hours in dextrose), we observed a 2-fold range in growth rate with these constructs (Figure 2A) and a 100-fold range in expression level (Figure 2B). We quantified expression level using a Western blot assay directed against an 6×His epitope tag only present on the rescue copy of Hsp90 that we had previously optimized to yield a linear response [44]. These expression level measurements were performed after 19 hours in dextrose, where the second copy of Hsp90 driven by the galactose regulated promoter was undetectable (Supplementary Figure S1). To further investigate expression level, we developed an Hsp90-GFP fusion construct that we monitored by flow cytometry. Across all promoter constructs, the Hsp90-GFP fusion supported similar yeast growth rates to non-GFP tagged versions (Supplementary Figure S3). These findings indicate that the GFP fusion has minimal impacts on Hsp90 function. The expression levels determined by GFP and flow cytometry were in close agreement with those measured by Western blotting and the average of both measures was used to estimate expression levels (Supplementary Table S1).
Both the Western and GFP experiments demonstrate that the expression level of Hsp90 can be reduced dramatically (15-fold) without major impacts on growth rate, which is consistent with previous reports [22], [45]. The growth rate to Hsp90 expression level profile that we determined has the shape of a binding curve (Figure 2C), and can be fit to a binding equation that represents the elasticity function for Hsp90. This elasticity function defines how yeast growth rate varies with the net Hsp90 function and enabled us to calculate per molecule function of mutants from fitness measurements.
The non-linear elasticity function for Hsp90 describes the coupling of mutant effects on function and fitness. For example, when expressed at endogenous levels, an Hsp90 amino acid substitution would need to reduce per molecule function by 79% in order to result in a readily measureable growth defect of 5%. Thus the bimodal DFE that we observe for Hsp90 (Figure 1B) does not necessarily imply a bimodal distribution of mutant effects on function. In particular, the fitness analyses do not provide detailed information on mutants with up to 79% defects in function. Due to the shape of the Hsp90 elasticity curve, the bimodal DFE is consistent with either a bimodal distribution of function as predicted by the stability dominated fitness model [18], or a primarily unimodal distribution of functional effects (Figure 3). To distinguish between these possibilities we sought to reveal effects on function that could be hidden at wild type expression strength.
To reveal the latent function of Hsp90 mutants, we analyze fitness effects at reduced expression strengths (Figure 4, Supplementary Table S2). The population in all bulk competitions was managed such that the population size at constriction points was always in gross excess to library diversity (Supplementary Figure S4). Because there is selection pressure to increase expression in these experiments, we examined the expression level of the wild type Hsp90 sequence over time in shutoff conditions using Hsp90-GFP fusions (Supplementary Figure S5). Cells respond to selection by increasing expression from weak promoters over time. As predicted by the elasticity function (Figure 2), the increased expression from weak promoters results in an increase in growth rate (Supplementary Figure S6). The observed increase in growth rate closely matches predictions based on the expression increase we observed by flow cytometry and the elasticity function, indicating that the underlying model is sound. To minimize the impact of time dependent changes in expression on fitness analyses of coding sequence mutations, we performed bulk competition of Hsp90 mutants over a short time window, 12–48 hours in dextrose (Supplementary Figure S4). We performed simulations to investigate how the observed increase in expression level over time in shutoff conditions would impact competition trajectories (Supplementary Figure S7). The impact of increasing expression level has a minor impact on competition trajectories and indicates that constant expression models provide estimates of sufficient quality to interpret general features of the distribution of mutant effects on fitness and function, which is the focus of this study.
The DFEs that we observed exhibited a consistent trend as expression strength was reduced. At high expression strength, the majority of mutants had WT-like growth rates, with very few mutants of intermediate effect. As expression strength was reduced, the WT-like peak decreased and the prevalence of mutants with intermediate effects increased. In terms of epistasis, the fitness effects of amino acid substitutions displayed pervasive negative epistasis with expression strength (Supplementary Figure S8). In terms of function, these results strongly indicate that the DFE at endogenous expression strength (Figure 1B) does not mirror the underlying effects of point mutations on Hsp90 function.
We estimated mutant effects on Hsp90 function (Figure 5, Supplementary Table S3) based on fitness measurements at distinct expression strengths and the elasticity function. As described in the methods section, we employed the elasticity function to calculate per molecule function from fitness taking into account bounds on measurement and calculation precision. For example, at the endogenous expression strength, mutants with activity defects of up to 79% were obscured to fitness analyses and were demarcated as such (functional efficiency >0.21). Because a distinct range of function is revealed to selection at each expression strength (Table 1), our integrated analyses provided estimates of the functional effects of all mutants. Estimates of mutant effects on function based on fitness measurements at different expression strengths exhibit a reasonable correlation (R2 = 0.75) (Supplementary Figure S9). The strength of this correlation, despite simplifying assumptions (further discussed in the methods section), indicates that the calculated mutant effects on function are fair estimates.
The distribution of functional effects for a region of a protein provides information about the contributions of that region to biochemical activity. For example, scaffolding regions that are not directly involved in a critical or rate-limiting step in protein function should be hard to break by mutation (due to selection for stability in the wild type protein), but once broken destroy activity [19], [20]. In contrast, regions that contribute to a rate-limiting step should be easy to injure by mutation, with the severity of mutant defects mediated by the rigidness of chemical and physical requirements (e.g. catalytic sites in enzymes being ultimately rigid with any mutation destroying activity).
The distribution of functional effects (Figure 5A) for the region of Hsp90 we analyzed had one main peak with most mutations exhibiting partial defects relative to wild type. Our finding is consistent with this region of Hsp90 contributing to a critical and rate-limiting step in function. The intermediate functional defect of most mutants indicates that the chemical and physical requirements are flexible, consistent with this region of Hsp90 providing a hydrophobic docking site for binding to substrates, as was inferred from structure [32]. Taking a closer look at the aromatic amino acids at position 583 (Phe) and 585 (Trp) located on the surface of the Hsp90 structure, most amino acid substitutions are tolerated when expressed at endogenous levels, but a clear functional preference for hydrophobic amino acids is revealed at reduced expression strengths (Figure 5B). Hydrophobic interactions [46] are malleable to slight alterations in geometry and physical composition compared to other physical interactions (e.g. hydrogen bonds). Thus, it is reasonable that some substitutions that maintain hydrophobicity would be well tolerated, but that most non-conservative substitutions would result in strong defects.
Our fitness-based estimates of mutant effects on function integrate over all properties that contribute to cell growth including catalysis, binding affinity, as well as the thermodynamic stability of folding to the native state [18]–[20], [27], [47]. In terms of stability, the prevalence of intermediate functional defects that we observe is inconsistent with this region of Hsp90 serving a purely scaffolding function, which theory predicts should exhibit a bi-modal distribution [18]. Furthermore, we observed a similar distribution of functional effects for positions located on the protein surface, which should have relatively small impacts on stability [48], as those that orient towards the protein interior (Supplementary Figure S9). This finding suggests that the functional effects of mutants at solvent shielded positions are caused primarily by local structural changes that impact the organization of solvent exposed positions (e.g. as required for efficient binding to substrate). We have observed a similar surface-core relationship in ubiquitin [43], and at a lower resolution this type of surface-core association has been postulated based on the slow evolutionary divergence of sites in proteins located proximal to binding sites [49]. Of note, Hsp90 is a dimeric protein and subunit folding and association are coupled [44]. Thus, decreased expression strength could increase sensitivity to destabilizing mutations. In this case, destabilizing mutations would exhibit larger activity defects at lower expression strength. Across the dataset our functional estimates are largely independent of expression strength (Supplementary Figure S9, Panel A). Thus, the effects of mutations on dimer stability appear to have at most a minor impact on our activity estimates, consistent with the location of this region of Hsp90 far from the dimer interface [50].
To further examine the effect of mutations on stability, we simulated the stability effects of each possible point mutation based on the structure of Hsp90 [50] using Rosetta [51], which accurately predicts the experimental effects of mutations on stability. The simulated stability effects for Hsp90 correlate extremely weakly with activity (Figure 5C), consistent with our conclusion that stability is not a dominant contributor to activity for this region of Hsp90. Of note, substitutions of amino acids with similar physical and chemical properties (as estimated by BLOSUM similarity) to the wild type residue tend to be compatible with function (Figure 5D). The stronger correlation of function with amino acid similarity compared to stability suggests that the stability simulations do not fully capture all biologically relevant structures. For example, high resolution structures of Hsp90 bound to substrate are not available; but if they were available, might provide a stronger structural explanation for the observed functional effects of mutations.
To further test our model and conclusions, we experimentally investigated the biochemical properties of five non-conservative amino acid substitutions. We chose mutations that dramatically change the hydrophobic binding surface and largely destroy function (F583D and W585D), mutations that disrupt intra-molecular interactions and severely impair function (S586H disrupts a buried hydrogen bond, and A587D introduces a buried charge at a solvent shielded location), and a charge reversal mutation (E590K) on the surface that causes a moderate functional defect. The growth rate of these mutants in monoculture closely matched the fitness effects observed in the bulk competitions (Supplementary Figure S10). As discussed above, our estimates of function integrates over multiple protein properties. For example, a mutation that increases the degradation rate (with the synthesis rate unchanged) should exhibit reduced steady state levels leading to a defect in net function. All of the disruptive individual mutations that we investigated accumulated at similar steady state levels (Figure 6A), suggesting that individual mutations do not commonly disrupt Hsp90 protein levels.
We examined the biophysical properties of these non-conservative Hsp90 mutant proteins in purified form. To maximize the sensitivity of these analyses for potential alterations to structure and stability, we generated C-domain constructs. All of the mutations we analyzed are located in the C-domain and do not contact other domains in the Hsp90 structure. The circular dichroism (CD) spectra of all five mutant proteins overlay closely with WT (Figure 6B) indicating that all of the mutants fold into native conformations with similar secondary structure content to WT. We investigated the stability of each mutant protein to urea-induced unfolding (Figure 6C). Similar concentrations of urea were required to unfold all mutants and WT indicating that none of the mutants compromises folding under native conditions. These findings demonstrate that non-conservative mutations in this region of Hsp90 are generally capable of folding to stable native states, and strengthen our conclusions that the 582–590 region of Hsp90 that we analyzed is not critical for folding stability, and is instead a structurally malleable region that forms a critical hydrophobic docking site.
Our studies as well as those of others [21], [24], [25], [27], [52], [53] demonstrate that biochemical flux models and the elasticity function in particular provide a fundamental link between molecular and cellular/organismal properties. Non-linear elasticity functions of the identical form to those described here for Hsp90 have also been observed in E. coli for β-galactosidase[53], isopropylmalate dehydrogenase [24], and dihydrofolate reductase (DHFR) [25]. In E. coli, DHFR point mutations were commonly observed to impact protein degradation rates leading to fitness effects that were strongly dependent on the level of protein quality control [25]. In addition, flux models can provide a mechanistic explanation for many common fitness features including pleiotropy and epistasis [54].
This study clearly demonstrates that functional defects of mutants can be hidden to experimental fitness measurements due to a non-linear elasticity function. Uncovering these latent effects revealed that the region of Hsp90 we analyzed contributes to a rate-limiting step in Hsp90 function. These findings indicate that critical functional regions in proteins are more prevalent than considered based on fitness analyses performed without consideration of the elasticity function. The elasticity function relating net function and fitness is critical for a thorough understanding of mutant fitness effects.
For expression analysis, the yeast Hsp90 gene was cloned into the pRS414 plasmid with different promoters and 3′ untranslated region (UTR). We used constitutive promoters previously demonstrated to generate a wide variation in expression level [55] including GPD, TEF, ADH, and CYC. Constructs were generated with or without the 3′ UTR from the CYC gene, which allowed further variation in expression level [56]. In constructs lacking the CYC terminator, the 3′UTR was composed of sequence from the plasmid vector. All Hsp90 plasmids contained a 6X-His sequence (GGHHHHHHGGH) at the N-terminus to facilitate detection by Western blotting. Point mutant libraries previously generated in p417 plasmids [9] were transferred to the pRS414 promoter variant plasmids using SLIC cloning [57]. Briefly, for each promoter strength construct, we prepared a destination vector with the first and last 30 bases of Hsp90 bracketing a unique SphI restriction site. We excised the Hsp90 library from the original 417GPD plasmid using restriction enzymes that cut immediately upstream and downstream of the Hsp90 gene. We cut destination vectors with SphI. We generated ∼30 base complementary overhangs using T4 DNA polymerase in both the destination vectors and the Hsp90 library, annealed the complementary DNA, transformed into competent bacteria, grew in bulk selective (Amp) cultures and prepared plasmid. A small portion of the transformation was plated and the number of independent transformants (∼30,000) was in gross excess to the library diversity. In addition, all replication is performed in bacteria where multiple systems ensure high fidelity reducing the probability of undesired secondary mutations. The DBY288 Hsp90 shutoff strain (can1-100 ade2-1 his3-11,15 leu2-3,12 trp1-1 ura3-1 hsp82::leu2 hsc82::leu2 ho::pgals-hsp82-his3) was generated from the Ecu Hsp90 plasmid swap strain [42] by integration of Hsp90 driven by a GalS [41] promoter together with a HIS3 marker into the HO genomic locus.
DBY288 cells were transformed with pRS414 plasmids and selected on synthetic raffinose and galactose (SRGal) plates lacking tryptophan (-W). Single colonies were then grown in liquid SRGal-W on a rotator at 30°C to late-log phase (OD600∼0.8). Cells were collected by centrifugation, washed with synthetic dextrose (SD) –W media, and then grown in SD-W medium at 30°C in an orbital shaker. Culture density was maintained in log phase (OD600 between 0.1 and 0.8) by periodic dilution. Culture growth was monitored based on increases in OD600 taking into account cumulative dilution. The log of OD600 versus time was fit to a linear equation to determine growth rate. Analyses were performed on time points in dextrose where control cells lacking a rescue Hsp90 had depleted Hsp90 by Western analyses (Figure 2B & Supplementary Figure S1) and had stalled in growth (Figure 2A and Supplementary Figure S1).
To analyze expression levels of different promoter constructs, cells were grown for 19 hours in SD -W media, and 108 yeast cells were collected by centrifugation, and frozen as pellets at −80°C. Cell lysates were prepared by vortexing thawed pellets with glass beads in lysis buffer (50 mM Tris-HCl pH 7.5, 5 mM EDTA and 10 mM PMSF), followed by addition of SDS to 2%. Lysed cells were centrifuged at 18,000 g for 1 minute to remove debris, and the protein concentration of the supernatants was determined using a BCA assay (Pierce Inc.). Lysates with 15 µg of cell protein were resolved by SDS-PAGE, transferred to a PVDF membrane, and Hsp90 probed using α-HisG antibody (Invitrogen Inc.). Importantly, we have previously shown that detection of this 6×His Hsp90 construct in yeast can be detected with a broad linear range using this antibody and Western blot approach [44].
Flow cytometry was used as an alternative approach to measure the expression level of Hsp90 at the single cell level in yeast cells. A gene encoding EGFP was inserted into the unstructured tail of Hsp90 after amino acid position 684. This Hsp90-GFP fusion construct was cloned into the variable strength promoter constructs used with non-GFP tagged Hsp90. These plasmids were transformed into DBY288 yeast competent cells and grown on SRGal-W plates. A single colony of each strain was grown for two days at 30°C in SRGal-W media to near saturation. These cultures were diluted 1∶50 into SRGal-W media and grown to late log phase (∼106 cells/ml). Each strain was then further diluted 1∶50 in SD-W media for 48 hours at 30°C with dilution every 12 hrs in order to maintain cells in log phase growth. Samples of cells were collected after 19, 36, and 48 hours in dextrose. Collected cells were washed twice in wash buffer (50 mMTris, 150 mMNaCl, pH 7.6, 0.1% w/v BSA), diluted to 107cells/ml in wash buffer and analyzed on a Becton-Dickinson FACSCalibur flow cytometer equipped with a 15 mW air cooled 488 nm argon-ion laser using a 530 nm high-pass filter. Greater than 100,000 cells were analyzed for each sample. Data were processed and analyzed using FlowJo software. Debris including clumped cells was excluded by gating on the forward and side scatter (excluded less than 5% of points). To compare with bulk Western measurements, mean fluorescence was calculated using cells without GFP in order to subtract out background due to autofluorescence.
C-domain constructs of Hsp90 bearing an N-terminal 6×His tag were generated in a bacterial over-expression plasmid, expressed, purified, and analyzed by circular dichroism (CD) as previously described [44]. Briefly, CD spectra were obtained using a 1 mm path length cuvette at a protein concentration of 20 µM in 20 mM potassium phosphate at pH 7 and 25°C. Urea titrations were performed under the same conditions using samples that were equilibrated for 30 minutes. Urea concentrations were determined based on their refractive index. CD ellipticity at 222 nm was used to follow urea induced unfolding and the resulting data was fit to a two-state unfolding model as previously described [44].
The effect of point mutants on yeast growth was analyzed as previously described [36]. Time points in dextrose were selected for analysis where control cells lacking a rescue Hsp90 began to stall in growth in order to observe the rapid decrease in relative abundance of deleterious mutants (e.g. premature stop codons). The growth rate of cells harboring the WT coding sequence in bulk competitions was estimated from monoculture growth of WT constructs performed in parallel to the bulk competitions. For the GPD, TEF and TEFΔter constructs we analyzed time points in dextrose of 12, 16, 20, 24, 32, 40, and 48 hours (Supplementary Table S4). For the CYC, ADH, CYCΔter, and ADHΔter constructs where the relative decrease of deleterious mutants was less severe (due to slower growth rate of fit mutants) we analyzed time points in dextrose of 16, 20, 24, 32, 40, and 48 hours. To process these time point samples, yeast pellets were lysed with zymolyase and total DNA was extracted and purified through a silica column. The DNA encoding amino acids 582–590 was PCR amplified, and prepared for 36 base single-read Illumina sequencing. 3.4×107 high quality reads (>99% confidence across all 36 bases) were obtained and analyzed. The relative abundance of each point mutant at each time point for each promoter was tabulated. Effective selection coefficients for yeast growth were determined by linear fits to the change in mutant abundance relative to wild type for each possible codon substitution. To account for the rapid depletion of null-like mutants to noise levels, only the first three timepoints in selection were used to determine effective selection coefficients for stop codons and all other mutants with effective selection coefficients within two standard deviations of stop codons (corresponding to s = −0.28 for GPD, s = −0.37 for TEF, s = −0.4 for TEFΔter, s = −.0.35 for CYC, s = −0.46 for ADH, s = 0.44 for CYCΔter, and s = −0.43 for ADHΔter). Because these null and near-null mutants rapidly deplete from the culture it is challenging to precisely measure their relative growth effects and they were binned as “null-like” (Supplementary Table S2).Potential noise was analyzed by calculating normalized residuals (residuals/time points fit). Codon substitutions with residuals per time point greater than 0.25 or low initial mutant abundance (mutant/wt less than 0.004) were omitted (∼7% of codons). For mutants that persist in the bulk competition (s>−0.1) synonymous codons exhibit a narrow distribution (Supplemental Figure S11) indicating that the amino acid sequence is a dominant determinant of fitness. The effective selection coefficient for each amino acid substitution was estimated as the average of the effective selection coefficients of all synonymous codons. Epistasis between expression strength and amino acid substitutions was calculated as the difference in effective selection coefficient at reduced expression strengths relative to endogenous strength. For the epistasis calculations, null-like mutants were considered as true nulls. Thus, a mutant with wild type fitness at endogenous expression strength, and null-like fitness at the reduced expression strength would have an epistasis of −1.
Function per molecule was calculated based on observed selection coefficients, the elasticity function, and the expression level for each different promoter construct using the following equations.(1)(2)
Where G is growth rate, Gmax is the maximal growth rate, Em is the relative expression level that results in half maximal growth, E is the expression level relative to the endogenous level, F is the per molecule functionof a mutant relative to WT, Wmut is the growth rate of a mutant relative to WT, and s is the effective selection coefficient. Equation 1 is an extension of the elasticity equation (Figure 2), where the expression of functional molecules or net function (EF) is explicitly modeled. With the WT coding sequence (F = 1 by definition), equation 1 simplifies to the elasticity function in Figure 2. These equations can be combined and rearranged to define F as follows.(3)
Equation 3 was used to estimate mutant effects on function (Supplementary Table S3) using the observed selection coefficients (Supplementary Table S2), Em = 0.014 (Figure 2), E for each promoter construct based on experimental measurements (EGPD = 1,ETEF = 0.32,ETEFΔter = 0.094), or estimated from the observed growth rate and the elasticity function for weak promoter constructs where experimental measures of expression were noisier (ECYC = 0.028,ECYCΔter = 0.015,EADH = 0.014,EADHΔter = 0.010). Where growth rates prohibited accurate estimation of fitness (null-like mutants, or absolute growth rates within 5% of Gmax), bounds on relative per molecule function were calculated (Table 1). For each amino acid substitution, a final per molecule function estimate was generated by averaging across all promoter constructs that yielded a numerical estimate (and not a bound). For all pair-wise numerical function estimates (e.g. at two different expression strengths), we compared function effects between all constructs with adjacent expression levels (Figure S9). To facilitate biophysical comparisons, we used the Blosum62 matrix [58] to calculate the amino acid similarity to wild type for each possible point mutation, and Rosetta [51] to simulate effects on thermodynamic folding stability.
We make the simplifying assumption that expression level is independent of mutations to the coding sequence. Steady state expression level is determined by the rates of both synthesis and degradation. Because degradation occurs after protein synthesis, it should depend primarily on the protein sequence such that synonymous substitutions minimally impact degradation rates. Across our data set we noted that synonymous substitutions did not have dramatic impacts on fitness, suggesting that synthesis rates were relatively independent of mutation. Protein degradation rates vary depending on protein sequence, but all of the mutants that we analyze are single amino acid substitution, and hence minimally differ in overall sequence. In the event that a point mutant impacts degradation rate, it should be consistent across each promoter construct. Thus, mutant impacts on degradation should be rare (see Figure 6), but would be incorporated into our estimates of function.
In analyzing the effect of mutations relative to wild type, we make the simplifying assumption that function is independent of expression level. We examined the validity of this assumption by analyzing the standard deviation in function for each amino acid substitution determined at different expression levels. The average standard deviation was 0.1, indicating that this assumption is valid on a rough scale (on the order of 0.1) and is appropriate for interpreting the main features of the distribution of mutant effects on function. Of note, the mutations that we observe to improve function at reduced Hsp90 expression levels (Figure 5, Supplementary Table S3) may be an artifact of this assumption.
The elasticity function does not include a cost of expression and as such has a maximum fitness at infinite expression level. Thus, we assume that expression cost is negligible relative to expression benefit over the range of our analyses. As the expression cost of native proteins is below experimental detection in yeast [59], this assumption appears reasonable.
We infer differences in cellular growth rates from measurements of DNA abundance. This inference is valid if DNA and cellular abundance are coupled. In previous work, we demonstrated that EMPIRIC measurements of fitness based on measures of plasmid abundance correlate strongly with cellular growth rates for a large set of mutants [43], indicating that plasmid abundance and cellular abundance are coupled. In addition, the copy number of the CEN plasmids utilized in this study is regulated, as cells maintaining multiple CEN plasmids grow slowly [60]. In addition, the low copy number of CEN plasmids is dominant to the addition of high copy genetic elements [61] and genetic alterations that increase CEN abundance are rare [62]. Nonetheless, CEN plasmids are not as stable as chromosomally encoded DNA, which may lead to a small amount of noise in our measurements.
|
10.1371/journal.pcbi.1000777 | A Differentiation-Based Phylogeny of Cancer Subtypes | Histopathological classification of human tumors relies in part on the degree of differentiation of the tumor sample. To date, there is no objective systematic method to categorize tumor subtypes by maturation. In this paper, we introduce a novel computational algorithm to rank tumor subtypes according to the dissimilarity of their gene expression from that of stem cells and fully differentiated tissue, and thereby construct a phylogenetic tree of cancer. We validate our methodology with expression data of leukemia, breast cancer and liposarcoma subtypes and then apply it to a broader group of sarcomas. This ranking of tumor subtypes resulting from the application of our methodology allows the identification of genes correlated with differentiation and may help to identify novel therapeutic targets. Our algorithm represents the first phylogeny-based tool to analyze the differentiation status of human tumors.
| Gene expression profiling of malignancies is often held to demonstrate genes that are “up-regulated” or “down-regulated”, but the appropriate frame of reference against which observations should be compared has not been determined. Fully differentiated somatic cells arise from stem cells, with changes in gene expression that can be experimentally determined. If cancers arise as the result of an abruption of the differentiation process, then poorly differentiated cancers would have a gene expression more similar to stem cells than to normal differentiated tissue, and well differentiated cancers would have a gene expression more similar to fully differentiated cells than to stem cells. In this paper, we describe a novel computational algorithm that allows orientation of cancer gene expression between the poles of the gene expression of stem cells and of fully differentiated tissue. Our methodology allows the construction of a multi-branched phylogeny of human malignancies and can be used to identify genes related to differentiation as well as novel therapeutic targets.
| Cancer research has traditionally focused on the identification of oncogenes and tumor suppressor genes, but in the last decades it has become increasingly apparent that disruption of normal differentiation is an important component of tumorigenesis. Lack of cellular maturation is now recognized as a hallmark of human cancers [1], and the degree of differentiation of a tumor is important for diagnosis, prognosis, and treatment. Investigations of hematopoietic malignancies, for instance, have benefited considerably from an understanding of the differentiation hierarchy of hematopoietic cells. The identification of immunophenotypic markers and gene expression profiles correlated with maturation has enabled researchers to map the expansion of malignant cells to particular stages of hematopoietic differentiation [2]. Such characterization has proven invaluable for diagnostic and prognostic purposes, and continues to provide clues for pharmacological interventions. Furthermore, the extent of differentiation indicated by the histologic subtype of liposarcoma is the most important determinant of the clinical outcome for this cancer type [3]–[5]. Nevertheless, attempts to categorize solid tumors have proven difficult due to an incomplete understanding of differentiation pathways from stem cells into mesenchymal and epithelial tissues. The classifications undertaken so far have been based on in vitro measurements of genes expressed during the differentiation of stem cells into mature tissue; this data was then compared to expression profiles of different tumor subtypes to identify the maturation stages to which these subtypes correspond [6]. However, such approaches are not yet widely applicable since the prospective isolation of tissue-specific stem cells has been possible for only few tissue types, e.g. hematopoietic, mesenchymal, epithelial, and neural tissues ([7] and references therein). Similarly, in vitro methods of differentiation are available for only a few histologies [8]. Furthermore, the necessity of an array of growth factors for in vitro differentiation raises questions about the similarity of the in vitro model to in vivo processes. Often only a fraction of cells undergoes differentiation under in vitro conditions, and currently available methods do not allow isolation of those cells during the differentiation process from the bulk of unchanged cells.
An objective categorization of cancers according to maturity requires a methodology that does not depend on expression data obtained from in vitro models of differentiation. In this paper, we develop a novel computational algorithm that assigns a degree of dissimilarity from stem cells to human cancer subtypes. Our methodology utilizes gene expression data of tumor subtypes to construct a phylogenetic tree based on genes differentially expressed among the subtypes, as well as gene expression data of stem cells and fully differentiated cells. The resulting phylogeny provides information about the maturation status of tumor subtypes and the relationship between them. The results of our algorithm are conceptually similar to the mapping of cellular expansion occurring during hematopoietic malignancies to the differentiation hierarchy of hematopoiesis. Our methodology allows classification of cancer subtypes according to their maturation status, to identify genes whose expression correlates with differentiation, and to discover candidate genes which are promising therapeutic targets. Our methodology is part of an increasing literature of mathematical and statistical investigations of cancer [9]–[14].
Our algorithm uses gene expression data of tumor samples that have been pathologically classified into subtypes. The expression data is normalized and then analyzed for differentially expressed genes, i.e. those genes whose expression in samples from one tumor subtype differs from the expression in samples from at least one other subtype. We use these genes to compute the distances between all pairs of subtypes; the resulting distance matrix is then used to construct a phylogenetic tree. This construction is repeated several thousand times using different subsets of genes (of varying size) to estimate the statistical significance of the branches of the tree (Fig. 1). We perform a systematic analysis of several methods and parameters used in our algorithm (see Methods for details). We find that combining ANOVA and Benjamini-Hochberg with a p-value of 0.01 gives good and robust results, while the Weighted Least Squares (WLS) tree reconstruction method works best when combined with the Pearson correlation matrix. Other combinations of methods give similar results and therefore should be tested in order to have an accurate understanding of a given dataset.
The phylogenetic tree resulting from this analysis contains information about the relation among subtypes as well as between subtypes and the root of the tree. The branching points represent the ‘common ancestors’ of the subtypes that are situated at the leaves of those branches. If the tree is rooted with expression data of a primitive cell type such as embryonic or tissue-specific stem cells, then the subtypes that are located more closely to the root correspond to types that are more similar to stem cells while the subtypes that are located farthest away from the root represent the most dissimilar types. The order of the branching points along the differentiation course can be interpreted as the ranking in dissimilarity of each of the subtypes to stem cells. The differences between stem cells and tumor subtypes are in part caused by different differentiation status and in part by the abnormal cancer phenotype. In some situations, the order of the subtypes dictated by the tree is not unique, resulting from a non-fully balanced tree. For instance, more than one subtype can be mapped to exactly the same point in the ordering according to dissimilarity from stem cells. Furthermore, the two subtypes farthest away from the root share the same common ancestor and therefore cannot be distinguished in their level of dissimilarity. To resolve this conflict, expression data of a fully differentiated cell type can be included, which unambiguously defines the last branching point in the ranking.
We validate our methodology with three datasets: (i) a dataset containing gene expression data of acute myeloid leukemia (AML) samples which are categorized according to the French-American-British (FAB) classification into classes that mirror maturation status [2]; (ii) a dataset containing gene expression of breast cancer samples classified according to estrogen receptor status and Elston histological grade [15]–[17]; and (iii) a dataset containing gene expression data of liposarcoma subtypes which have been analyzed for their differentiation status by comparing them to an in vitro differentiation time course [6].
Acute myeloid leukemia (AML) is a clonal disease characterized by the accumulation of myeloid progenitor cells in blood and bone marrow [18]. AML results from changes in transcription factor regulation that lead to a disruption of normal cellular differentiation. AML is classified into seven distinct subtypes depending on the morphology and differentiation status of tumor cells: dedifferentiated, myeloblastic, myeloblastic with maturation, promyelocytic, myelomonocytic, monocytic, and erythroleukemic AML. According to the FAB classification, these subtypes are denoted by M0, M1, …, and M6, respectively. Since AML is the result of alterations of the differentiation process, we validate our approach with a dataset of gene expression of AML patients.
Our leukemia dataset contains gene expression data of 362 AML patients and of 7 patients with unclassified Myelodysplastic Syndrome (MDS) (see Methods for details of dataset compilation) (Table 1). To root the AML tree, we use expression data of human embryonic stem cells (hESC); additionally, we include expression data of CD34+ hematopoietic cells from both peripheral blood (CD34 PB) and bone marrow (CD34 BM), human mesenchymal precursor cells (hESC MPC), as well as fully differentiated mononuclear cells from peripheral blood (PB) and bone marrow (BM). The surface glycophosphoprotein CD34 is expressed on undifferentiated hematopoietic stem and progenitor cells [19] and is widely used as a marker for less differentiated hematopoietic cells. We include these two subgroups as a further test of our methodology since their differentiation status is known. We use ANOVA to identify those probe sets that are significantly differentially expressed in at least one subtype as compared to all other AML subtypes. The analysis identifies 11,105 probe sets that are differentially expressed among AML subtypes if a false discovery rate (FDR) [20] of 0.01 is used. Use of this cutoff would lead us to expect 111 false positives. If we use the Holm correction method instead [21], which controls the family-wise error rate, then the number of differentially expressed probe sets decreases to 4,051 (with 0.01 expected false positives). The inclusion of less significantly differentially expressed genes is a potential source of noise; however, high cutoffs for significance discard genes that could be interesting for further analysis. The tradeoff between these two effects must be examined carefully to choose an appropriate cutoff. We decided to use a standard cutoff FDR of 0.01 because the tree topology remains stable for large gene sets, and also a larger number of potentially interesting genes are included which can be further filtered with other techniques.
The consensus phylogenetic tree based on this data is shown in Fig. 2. The order of the branching points of the subtypes coincides with the differentiation stages specified by the FAB classification: dedifferentiated AML (the M0 subtype) is located close to the stem cells while myelomonocytic (M4) and monocytic (M5) AML are located in the most distant leaves of the tree. The inner branching of the tree is also in accordance with the differentiation status suggested by the FAB classification (Table 1). The tree topology specifying the correct order of myeloblastic and promyelocytic maturation (M2 and M3), however, only has a moderate bootstrap value because the two subtypes are very similar in maturity. The branch leading to the erythroleukemic subtype (M6) is relatively unstable. This could be attributed to the small number of samples in this subtype or to a possible misclassification or erroneous diagnosis. Therefore, the position of this subtype in the tree is less certain than that of other subtypes; this uncertainty decreases the bootstrap values of the other branches at which this subtype can be located. All other branches in the tree are very stable under bootstrapping. Of central importance for the interpretation of the results is how well the tree captures the observed relationships in the data. A good measure of this fit is the average percent standard deviation of the distances between subtypes in the data compared to the ones in the tree. The Least Squares algorithm minimizes this score. For the Pearson correlation distance, the mean observed average percent deviation is 12.05%, which is a reasonable fit for this distance measure [22]; hence our algorithm produces a phylogeny which accurately recapitulates the relationships seen in the data.
We also apply our algorithm to a breast cancer dataset in order to study the performance of our method using cancers of epithelial origin. The samples in our dataset were characterized by immunochemistry methods according to their estrogen receptor status (ER+ and ER−) and Elston histologic grade (G1, G2, and G3). We compile a total of 483 unique samples, among which we find all combinations of ER status and grade (Table 2). The raw data is analyzed as described in the methods section. We root the tree with human mesenchymal stem cells and also include samples of normal breast [23]. Results are shown in Fig. 3. We find 17,966 probes differentially expressed between the subgroups when using ANOVA with Benjamini-Hochberg correction and a cutoff value of 0.01. A negative ER status has been shown to correlate with poor prognosis [24]. Consistent with this observation, our algorithm places ER-negative subgroups closer to stem cells, reflecting the more stem-like properties of these aggressive tumors, while ER+ tumors are placed closer to the normal breast tissue samples. Tumor grades are ordered similarly, placing tumors of higher grade closer to stem cells. Most trees reconstructed with the different sets of genes have the same topology (bootstrap values close to 100%), reflecting a very robust phylogeny. We conclude that our methodology is also able to accurately rank tumors of epithelial origin according to maturity.
Next we construct a phylogeny of liposarcoma subtypes. Liposarcoma is the most common type of soft tissue sarcoma accounting for about 20% of all tissue sarcomas [25]. In 2008, 10,390 new cases of sarcoma were reported in the US [26]. Surgery is the standard care for localized tumors but leads to worse prognoses in cases of locally advanced or disseminated disease [27]. Liposarcomas are classified into three biological types encompassing five subtypes: (i) well-differentiated/dedifferentiated, (ii) myxoid or round cell, and (iii) pleomorphic liposarcoma, based on morphological features and cytogenetic aberrations [28]. Although the subtype is the main determinant of clinical outcome [3], [4], [29]–[31], liposarcomas of similar morphology can differ in response to treatment and in prognosis [27]. Microscopically well-differentiated liposarcoma is composed of relatively mature adipocytic proliferation showing significant variation in cell size and at least focal nuclear atypia. Histologically dedifferentiated liposarcoma is represented by the transition from well-differentiated liposarcoma to non-lipogenic sarcoma. Both well-differentiated and dedifferentiated liposarcomas contain characteristic ring or giant marker chromosomes with 12q14-15 amplification. Myxoid liposarcomas contain uniform round to oval shaped primitive non-lipogenic mesenchymal cells and a variable number of small signet-ring lipoblasts in a prominent myxoid stroma. Round cell tumors are characterized by solid sheets of primitive round cells with no intervening myxoid stroma. Pleomorphic liposacoma is a pleomorphic high grade sarcoma containing a variable number of pleomorphic lipoblasts.
Recently, progress has been made towards a classification of liposarcoma subtypes utilizing gene expression data. In 2007, a 142-gene predictor was identified which correctly distinguishes between liposarcoma subtypes and generates a set of differentiation-related genes that may contain candidate therapeutic targets [27]. In 2008, Matushansky et al. showed that the main liposarcoma subtypes can be ranked according to their differentiation status by comparing gene expression data of the tumor subtypes with the genes expressed during normal in vitro adipogenic differentiation [6]. The ranking generated by the latter approach is useful for validating our methodology.
Our liposarcoma dataset includes 180 surgical samples that have been pathologically classified as 61 dedifferentiated, 52 well differentiated, 26 pleomorphic, 18 round cell, and 23 myxoid liposarcomas [27], [30]. Samples that were likely misclassified were filtered in previous studies, which is a pre-processing step critical for the outcome of the algorithm. For an FDR of the ANOVA filter of 0.01 after correction with the Benjamini-Hochberg method, we find 13,429 probe sets that are differentially expressed among the liposarcoma subtypes. Those sets are then used to construct an unrooted phylogenetic tree. To root the tree, we use expression data of mesenchymal stem cells and fully differentiated adipocytes. The resulting consensus tree is shown in Fig. 4a. The tree topology is stable with bootstrap values larger than 85%. Based on the consensus tree, the subtypes can be ordered by increasing dissimilarity from stem cells as dedifferentiated, pleomorphic, myxoid/round-cell, and well-differentiated liposarcoma (Fig. 4a). This order coincides with experimental results based on the gene expression observed during in vitro differentiation published earlier (Fig. 4b) [6]. By setting the p-value threshold of the Holm correction to 0.01, we obtain 7,290 differentially expressed probe sets; these probe sets generate a tree topology that is identical to the case described above with bootstrap values larger than 91.5% (data not shown). When rooting with embryonic stem cells, the branching between embryonic stem cells and the rest of the tree is less stable since the expression of embryonic stem cells differs considerably from all other samples (data not shown). To increase the stability of the tree, it is preferable to root with an outgroup that is relatively closely related to the investigated samples (in this case, mesenchymal stem cells; see also the section “Systematic analysis of methods and parameters”) [32]. Again we test how well the tree fits the distance matrix and observe a mean average percent standard deviation of 11.3%, which has been reported to be a good fit for the Pearson correlation distance [22]. Therefore, our methodology is also able to rank liposarcoma subtypes in the correct order according to their dissimilarity to stem cells.
Since our methodology correctly ranks leukemia, breast cancer, and liposarcoma samples according to their differentiation status, we now investigate a larger number of sarcoma subtypes to identify their relationship in maturity as well as candidate targets for therapeutic intervention. The sarcoma dataset includes the 180 liposarcomas discussed above as well as 36 myxofibrosarcomas, 5 pleomorphic malignant fibrous histiocytomas (MFH), 7 lipomas, and 23 leiomyosarcomas (Table 3) [27], [30]. We use expression data of both mesenchymal stem cells and embryonic stem cells to root the tree. The consensus tree is shown in Fig. 5. Our methodology determines that leiomyosarcoma is closest in its differentiation status to stem cells, followed by MFH and myxofibrosarcoma, and finally the liposarcoma subtypes (ranked as determined above) and the benign subtype lipoma. The algorithm also clusters the subtypes according to tissue of origin, predicting that leiomyosarcoma branches before all other subtypes, and that MFH and myxofibrosarcoma have a common ancestor; so do all liposarcoma subtypes and lipoma. Note that although pleomorphic liposarcomas and MFH/myxofibrosarcomas are very similar subtypes at the level of their genetic copy number aberrations [30], our algorithm places them in different branches of the tree. This effect is a result of the phenotype-based nature of our method and is in accordance with the different tissues of origin of these subtypes. The tree has a very stable topology with bootstrap values larger than 0.90 except for the MFH subtype, which exhibits a lower bootstrap value of 0.60; this value is likely due to the small number of samples (5) available for this subtype. Note that with the current dataset, we cannot distinguish between the case in which the subtype located most closely to stem cells, leiomyosarcoma, is situated on the adipocytic differentiation path and the case in which leiomyosarcoma is alternatively located on a branch leading to fully differentiated tissue of another type. To resolve this ambiguity, gene expression data of fully differentiated tissue of all the types giving rise to sarcomas is needed.
We are interested in identifying genes that are related to adipogenesis, i.e. those genes that correlate with adipocyte differentiation. To identify such genes, we cluster our list of differentially expressed genes into a chosen number of groups depending on their expression pattern in sarcoma subtypes. When the subtypes are arranged according to their distance from stem cells (as indicated by the tree in Fig. 4a), the expression of some genes continuously increases from the less differentiated to the more differentiated subtypes, while the expression of other genes decreases or exhibits more complicated patterns (Fig. 6). We hypothesize that genes whose expression continuously increases or decreases are possibly related to gain of the features of differentiation and loss of stem cell-associated functions, even though this association with maturation may not be causative. To test this hypothesis, we compare the genes whose expression increases or decreases along the order of subtypes to previously published lists of adipocytic differentiation-specific genes [6], [31]. In these two studies, mesenchymal stem cells were differentiated in vitro into normal fat cells, and the expression profiles of cells were measured at multiple time points during the differentiation process. An investigation of genes whose expression levels changed statistically significantly along the differentiation time course led to the identification of 67 and 69 genes, respectively [6], [31]. These genes are thought to be related to adipocytic differentiation.
We rank the genes whose expression increases or decreases along the liposarcoma subtypes (see Fig. 6 for example clusters) according to the fold change between their expression in hMSC and in normal fat. Among the 11,105 probe sets obtained by the ANOVA filtering with FDR of 0.01 after Benjamini Hochberg correction, the top 25 genes in this ranking are listed in Table 4. About 64% of these genes coincide with the published lists [6], [31]. These results suggest that our methodology is able to identify differentiation-related genes from the large number of differentially expressed genes. Additionally to the previously identified genes, our method identified other genes that have not been associated with adipocytic differentiation (Table 4). For instance, the protein phosphatase inhibitor 1 (PPP1R1A) is thought to be important in the control of glycogen metabolism and is primarily expressed in liver cells; the tyrosine kinase NTRK2 is part of a signaling pathway leading to neuronal differentiation, and the metabolism related enzyme system ACACB is exclusively expressed in adipocyte tissue.
We compare the results obtained from phylogenetic tree reconstruction algorithms with other methods of data clustering and organization such as a simple greedy algorithm (in which subtypes are linearly ordered by their distance from hESC), self-organizing maps (SOMs), and minimum spanning trees (MSTs) (see the Methods section for details of the algorithms). When applying the greedy algorithm to our AML dataset, we find similar results to those produced by the tree reconstruction analysis (Fig. 7a). Although the correspondence between the results of this method and the reconstructed phylogenetic tree is very good, the former only contain information of a linear organization, as opposed to the richer information that can be extracted from the tree topology and branch lengths. An example of a self-organizing map (SOM) algorithm applied to the AML dataset is shown in Fig. 7b. Subtypes that are known to be similar are mapped close together on the grid – e.g. human embryonic stem cells (hESC), mesenchymal stem cells (MSC), and samples with markers of poor differentiation (BMCD34 and CD34PB). Unfortunately, the overall organization of a SOM strongly depends on the shape and size of the grid, making it difficult to interpret the results in a robust and useful way for our purposes. Finally, we calculate a minimum spanning tree (MST) for the AML dataset (Fig. 7c). This algorithm accurately reproduces the reconstructed tree found with our original method, with the exception of mesenchymal stem cells being placed at the edge of the tree (instead of embryonic stem cells).
We compare the different methodologies implemented in our algorithm for each step of the analysis in order to identify those methods and parameters that perform well in the analysis of our datasets. We apply our algorithm to all datasets using all combinations of the following methods and parameters: for finding differentially expressed genes: ANOVA, Kruskal-Wallis (KW) and Welch approximation (Welch); two methodologies for p-value correction: Benjamini-Hochberg (BH) and Holm; two p-value cutoffs: 0.01 and 0.05; five tree reconstruction and clustering algorithms: Weighted Least Squares (WLS), Minimum Evolution (ME), Neighbor-Joining (NJ), FastME, and Average Linkage (UPGMA); and two distance measures: Pearson correlation and Euclidean distance. The results of these analyses are shown in Figs. S1, S2, S3, S4. The topologies found among the different combinations of parameters show that WLS, Pearson correlation, and BH with a cutoff value of 0.01 perform accurately in accordance with the AML (Fig. S1), breast cancer (Fig. S2), and liposarcoma datasets (Fig. S3).
Note that two main assumptions of the UPGMA algorithm are not fulfilled by cancer subtype data, namely: all species originate from a common ancestor and they all have evolved at the same pace. This issue explains why this method fails to reconstruct the right tree topologies; for example, in all sarcoma UPGMA topologies (trees 1 and 4 of Fig. S4), some liposarcoma subtypes branch together with leiomyosarcoma, which is thought to arise from smooth muscle tissue.
It has been shown in previous studies that, in general, WLS performs better than NJ when trees have long external or internal branches (e.g. [33]). Note also that the use of Euclidean distance leads to less robust results than the use of Pearson correlation when trees with long branches are considered. For example, when the Euclidean distance method is applied to the liposarcoma data, the dedifferentiated and pleomorphic subtypes cluster together with the well-differentiated subtype and normal fat (Topology 3 of Fig. S3). The effect of long branches on the Euclidean distance method becomes even more pronounced when analyzing the sarcoma data (Fig. S4); in this case, the least common topologies are observed only when the Euclidean distance method is used. If distant subgroups (i.e. hMSC and hMSC MPC) are removed from the analysis, then most parameter combinations including the Euclidean distance method favor topology 5. This topology was previously only observed with the Pearson correlation distance (see Table in Fig. S4, left).
We do not observe a significant influence of the choice of the method on the identification of differentially expressed genes. More important for our data is the choice of the p-value cutoff. For the sarcoma data, conservative p-value cutoffs favor topology 3 while parameter combinations with Benjamini-Hochberg adjusted p-values seem to favor topology 5 (Fig. S4). The results of our study suggest that BH with a cutoff of 0.01 is a good compromise, but we recommend investigating the effects of using different cutoff values.
In general, all tree reconstruction methods are very fast, especially since the number of different tumor subtypes in our analysis is typically limited. So it is possible to test many parameters in a reasonable time and we recommend doing so.
We have presented a rational methodology to investigate the dissimilarity between cancer subtypes and stem cells. Our approach uses gene expression data of tumor samples which have been classified into histological subtypes as well as expression data of an ‘evolutionary outgroup’ such as embryonic stem cells, tissue-specific stem cells, and/or fully differentiated normal cells. The data of tumor subtypes is used to identify the genes that are differentially expressed among the subtypes, and those genes, together with data of the outgroup, allows construction of a phylogeny of cancers. Our algorithm estimates the statistical significance of the tree branches by bootstrapping, a repeated tree construction using a varying number of randomly chosen genes. The distance between the branching points of the tumor subtypes and the stem cells specifies their dissimilarity, which is caused in part by differences in maturity, and ranks the subtypes according to increasing differentiation. This ranking is then used to identify genes whose expression continuously changes depending on the degree of maturation.
Our methodology is validated by being able to correctly reproduce experimental results concerning the relationship in differentiation status of liposarcoma, breast cancer and AML subtypes [2], [6], [18] and concerning genes related to adipocytic differentiation [6], [31]. Our method is useful for identifying genes that are overexpressed in some tumor subtypes (Fig. 6c). For instance, genes whose expression is increased in a particular tumor type but not in normal tissue-specific stem cells and differentiated cells may represent candidates for targeted therapy, possibly with lessened side effects. Interestingly, some of the genes found to be differentially expressed in only one or a few liposarcoma subtypes can be targeted by currently available drugs. It will be an important next step to test those genes for a causal role in tumorigenesis.
In recent years, bioinformatic tools have been widely used to analyze the vast amount of data produced experimentally. In analyses of microarray data, simple algorithms for phylogenetic tree reconstruction, such as Average linkage (UPGMA) [34], produce rooted bifurcating trees and are routinely applied to visualize similarities in gene expression. The most prominent example for this type of analysis are heatmaps, a graphical representation of the clustered expression matrix where colors represent the measured gene intensities; a dendrogram is often added which shows the bifurcating tree best describing the differences in gene expression [35]. Another important application of such algorithms is the clustering of tumor samples for improving or discovering subtype classifications (e.g. [36]). Other more sophisticated tree reconstruction algorithms are only rarely applied to expression data [22], [37]–[42]. The ‘molecular clock’ assumption of UPGMA (specifying that changes occur at a constant rate, [43]) renders this algorithm inappropriate for our investigation. Other algorithms such as Maximum Parsimony, Neighbor-Joining (NJ) [44], or Least-Squares [45] enable us to root the tree and to estimate the differentiation status of tumor subtypes by a simple comparison of the lengths between the root of the tree and the branching points of the leaves. We do not use character-based methods such as Maximum Parsimony due to the necessity of artificially discretizing the continuous values of gene expression intensities.
The estimation of evolutionary distances between tumors from gene expression data is hindered by the fact that small differences in the biology of tumors may cause large differences in gene expression. Examples of such situations are given by genes which trigger the expression of cascades of other genes [40] and mutational events affecting the expression of several genes [46]. In a recent paper [46], Park et al. proposed the use of correction methods that estimate mutational distances from the observed expression distances. This approach represents an interesting new avenue to further explore in future work.
The phylogeny of tumor subtypes identified by our methodology cannot be used to reconstruct the evolutionary history of a single tumor sample. The fact that dedifferentiated liposarcomas, for example, branch earlier than well-differentiated liposarcomas is not to be taken as evidence that one subtype evolved into the other. Rather, it specifies the dissimilarity of the bulk of tumor cells between cancer subtypes from stem cells at the time of observation. Similarly, our methodology cannot be used to identify the cell of origin of a tumor type. Both the position of a subtype in a differentiation-based phylogeny and the similarity of a subtype to an in vitro differentiation time course provide information about the bulk of tumor cells only; to determine whether these cells are produced from tumor stem cells which arose from cells of similar, earlier or more complete differentiation stages is outside the scope of this approach. Furthermore, the ability of a phylogenetic tree to reconstruct evolutionary trajectories when applied to genetic data rests on the assumption that the genetic material records the evolutionary history of the system. In the case of phenotypic information such as gene expression data, this assumption does not hold, and hence any information about the origin of the investigated cancer subtypes cannot be obtained.
The generality of our approach and the extensive availability of high-quality input datasets (e.g. GEO) makes this methodology a unique tool to investigate differentiation-related genes and the relationship in maturity of cancer subtypes. The use of data from patient samples reduces the problems encountered with in vitro studies regarding the reproducibility of the results in other systems and their significance to in vivo situations.
We use gene expression data of sarcoma samples from Singer et al. [27] and Barrentina et al [30]. The gene expression was measured on Affymetrix U133a oligonucleotide arrays. The classification in [27] was performed using unsupervised hierarchical clustering and an SVM-based supervised classification method. To root the tree, we use expression data of 17 normal fat samples from the same study as well as expression data of 3 human embryonic stem cell lines (hESCs) and 3 hESC derived mesenchymal precursor lines (downloaded from NCBI Geo [47] accession number GSE7332 [48]). We use gene expression data of AML [47] patient samples available within GEO (accession numbers GSE1159, GSE9476 [49], GSE1729 [50], and GSE12417 [51]). The breast cancer dataset is also compiled from Microarray data published in GEO with dataset numbers GSE7390 [16], GSE2990 [15], GSE3494 [17], and GSE9574 [23]. A problem of micrarray meta-analyses is that the different dataset sources may introduce a bias. We therefore applied hierachical clustering to the compiled breast cancer dataset and did not observe a clustering according to the sources.
The R code and the compiled AML dataset are available from the authors upon request. A user-friendly GUI that supports most of the methods described in this paper is available as Plugin for MAYDAY [70].
|
10.1371/journal.pntd.0005065 | Immune Modulation as an Effective Adjunct Post-exposure Therapeutic for B. pseudomallei | Melioidosis is caused by the facultative intracellular bacterium Burkholderia pseudomallei and is potentially fatal. Despite a growing global burden and high fatality rate, little is known about the disease. Recent studies demonstrate that cyclooxygenase-2 (COX-2) inhibition is an effective post-exposure therapeutic for pulmonary melioidosis, which works by inhibiting the production of prostaglandin E2 (PGE2). This treatment, while effective, was conducted using an experimental COX-2 inhibitor that is not approved for human or animal use. Therefore, an alternative COX-2 inhibitor needs to be identified for further studies. Tolfenamic acid (TA) is a non-steroidal anti-inflammatory drug (NSAID) COX-2 inhibitor marketed outside of the United States for the treatment of migraines. While this drug was developed for COX-2 inhibition, it has been found to modulate other aspects of inflammation as well. In this study, we used RAW 264.7 cells infected with B pseudomallei to analyze the effect of TA on cell survival, PGE2 production and regulation of COX-2 and nuclear factor- kappaB (NF-ĸB) protein expression. To evaluate the effectiveness of post-exposure treatment with TA, results were compared to Ceftazidime (CZ) treatments alone and the co-treatment of TA with a sub-therapeutic treatment of CZ determined in a study of BALB/c mice. Results revealed an increase in cell viability in vitro with TA and were able to reduce both COX-2 expression and PGE2 production while also decreasing NF-ĸB activation during infection. Co-treatment of orally administered TA and a sub-therapeutic treatment of CZ significantly increased survival outcome and cleared the bacterial load within organ tissue. Additionally, we demonstrated that post-exposure TA treatment with sub-therapeutic CZ is effective to treat melioidosis in BALB/c mice.
| Burkholderia pseudomallei is the causative agent of melioidosis, a fatal tropical disease endemic in parts of Southwest Asia and Northern Australia. While it was once believed to be isolated to these areas, recent research indicates the global burden on melioidosis is growing. Furthermore, treatment of melioidosis is difficult because of the high occurrence of disease relapse and increasing antibacterial resistance. Recent research suggests that immunomodulation via COX-2 inhibition to subsequently reduce with PGE2 production is an effective therapeutic strategy for melioidosis. The current study was built on this immunomodulatory principle by using an orally administered COX-2 inhibitor and evaluating its effects on the COX-2 and NF-ĸB pathways. We also investigated whether the conjunctive therapies of immunomodulation and antibiotics increased efficacy of the treatment. We confirmed immunomodulation is effective as a post-exposure therapeutic in BALB/c mice. More importantly, we found that conjunctive post-exposure treatment via immunomodulation increased antibacterial treatment efficacy. Conjunctive therapy may prove efficacious for other infectious diseases resembling melioidosis. Hence, further research is needed to identify the long-term effects of the described treatment(s) across multiple animal models.
| Melioidosis is a tropical and often fatal disease caused by the aerobic, Gram-negative facultative intracellular bacterium Burkholderia pseudomallei [1]. Traditional B. pseudomallei infection is associated with environmental exposures during the monsoon season in the tropics. It has also been identified as a Tier 1 select agent due to its high mortality rate, ability to cause respiratory infection and its drug resistance. B. pseudomallei is most common in Southeast Asia and Northern Australia were it is found naturally as a soil-dwelling bacteria [2] [3], Evidence suggests the incidence of melioidosis is underreported and the global burden is increasing with an estimated 169,000 cases per year and 89,000 deaths across 34 countries annually [4]. Therefore, the need to establish multiple therapeutic strategies is paramount.
B. pseudomallei is naturally resistant to many antibiotics, such as penicillin, many cephalosporins and aminoglycosides, but can be susceptible to such antibiotics as doxycycline, ceftazidime and chloramphenicol. The typical treatment regime for melioidosis lasts 20 weeks with both intravenous and oral phases of antibiotic administration. Due to these intense treatment requirements and antibacterial resistant isolates, relapse of the disease is common [2]. Recent evidence suggests that modulating the immune modulation by inhibiting cyclooxygenase-2 (COX-2) to reduce prostaglandin E2 (PGE2) expression is an effective post-exposure therapeutic. Another consideration of the current work is use of the COX-2 inhibition standard, NS-398, since is not approved for human use [5]. A COX-2 inhibitor with similar effects as NS-398 in a form administered easily to patients would be one step closer to developing a successful immune modulation regimen to treat melioidosis. This has profound implications as effective immune modulation treatment can reduce the selective pressure for bacteria to evolve to become drug resistant. Immune modulation can augment treatment of disease by either enhancing the effectiveness of a given antibiotic or by reducing the antibacterial dose required for treatment. Such interventions are promising developments toward the ultimate goal of eliminating an infectious disease by optimizing the host innate immune response [6].
Tolfenamic acid (TA) belongs to the fenamate class of NSAIDS and can be administered orally and intravenously to various animal species. Although not approved for human use in the United States, oral administration is used elsewhere in the world for treatment of migraines [7] [8]. While TA is primary known for its ability to inhibit COX-2, TA has also shown to be effective at modulating other key players in inflammation such as NF-ĸB. Evidence suggests a reduction in cytoplasmic NF-ĸB p65 activation in colon cancer cells and LPS stimulated RAW 264.7 cells as well [9] [10]. In veterinary medicine, TA has been shown to be a potent inhibitor of NF-ĸB p65 canine-derived tumor cells [11]. With its wide range of inflammatory modulation implications, TA could prove to be a valuable augmentation to current treatment therapeutics for melioidosis. Additionally, because of the broad impact of TA on many inflammatory mediators, its use may further elucidate how B. pseudomallei causes mortality.
In the present work, COX-2 inhibition was used to reduce the inflammatory response caused by B. pseudomallei. We examined the role of NF-ĸB, COX-2 and PGE2 during acute pulmonary infection with B. pseudomallei. For the first specific aim, we investigated the characteristics of the immune response and the potential of TA treatment to modulate the immune response and survival outcome. This work was done in-vitro by infecting RAW 264.7 cells with the B. pseudomallei 1026b ΔpurM strain Bp82. These results served as the basis of a second specific aim to expand our BALB/c in-vivo study protocol and monitor the bacterial dissemination and organ system burden of B. pseudomallei 1026b in order to confirm the relationship between bacterial burden and dissemination. From there, the study focused on how treatment with TA and known effective antibiotics, alone or in combination, affect survival outcome over time.
Dimethyl sulfoxide (DMSO) was purchased from ATCC (Manassas, VA). Dulbecco’s Modified Eagles Medium (DMEM) and trypsin used for cell culture were purchased from GE Healthcare Sciences (Hyclone) (Logan, UT) and fetal bovine serum plus (FBS +) was purchased from Atlas Biologicals (Fort Collins, CO). Tolfenamic acid was purchased from Cayman Chemical (Ann Arbor, MI). NF-ĸB monoclonal antibodies where purchased from Santa Cruz Biotechnology (Paso Robles, CA) and the secondary alexa fluor antibodies, 647 and HRP secondary antibodies used for all applications were purchased from Cell Signaling Technologies (Danves, MA). Luria-Bertani (LB) agar, cation adjusted Mueller-Hinton broth (ca-MHB) and COX-2 monoclonal antibodies were purchased from BD Sciences (Franklin Lakes, NJ). Dibutylhydroxytoluene (BHT), bovine serum albumin (BSA), Ceftazidime (CZ), crystal violet (CV) and sodium dodecyl sulfate (SDS) were purchased from Sigma-Aldrich (St. Louis, MO). Formalin and triton X-100 was purchased from ThermoFisher-Scientific (Waltham, MA). Ketamine for animal studies was purchased from Aurora Veterinary Supply (Aurora, CO).
RAW 264.7 cells were purchased from ATCC and maintenance was performed to the company’s specifications. Cells were grown in DMEM with 4 mM L-glutamine, 4500 mg/L glucose, 5 mM sodium pyruvate, 1500 mg/L sodium bicarbonate, and 10% FBS. Cells were grown in an incubator at 37°C with 5% CO2.
Bp82, a ΔpurM B. pseudomallei 1026b mutant incapable of adenine and thymine biosynthesis, [12] was used as a 1026b BSL-2 surrogate organism for in-vitro experiments. 1026b [13] was used for all BSL-3 animal studies. Both strains were prepared by growing 1 colony of the respective bacterial cultures in 50 ml of LB broth for 48 hours at 37°C. Bacterial stock was then frozen back in ~1.0 ml in LB broth and 10% glycol.
RAW 264.7 cells were plated in a 96-well plate with 100,000 cells per well and incubated for 24 hours. Wells were washed with sterile PBS, pretreated with either 100μM TA or 0.01% DMSO for 30 minutes, and infected with Bp82 at an MOI of 5. Each time point also included an untreated/infected control and an uninfected/untreated control. At each time point, 100μL of 10% formalin in methanol was added to each treatment well and placed on a rocker for 15 minutes with 20 tilts per minute. The formalin solution was removed and 100μL of a 0.5% solution of CV in 25% methanol/75% ddH2O was added to each well and placed on the rocker for 15 minutes. After staining, the 96 well plates were gently washed under tap water until the water ran clear from each well. The plate was air dried overnight until there was no visible liquid in any well. Finally, the CV was re-suspended from the cells by adding 100μL of 1% SDS solution to each well and placed on the rocker for 30 minutes at room temperature. The plate was then read on a plate reader at 570nm and 600nm. This procedure was adapted and optimized from Kursheed et al [14] and Castro-Garza et al [15]. Each treatment had a technical replicate of 10 and the experiment was run in biological triplicate for statistical significance.
To determine if TA possessed any natural antibiotic properties, a MIC was established. The procedure outlined in [16] was used. Briefly, Bp82 stock was incubated in LB broth overnight (18 hours) at 37°C, passed at a 1:100 dilution, and incubated for an additional 6 hours. This stock was then diluted to reach an optical density of 0.1 at 600 nm in caMHB. This diluted culture was further diluted 1:100 to achieve approximately 1x 106 CFU/ml inoculum. In 50μl of caMHB, 1:2 serial dilutions of TA and CZ were prepared in a 96-well plate with the highest concentration being 128μg/ml to the lowest concentration of 0.0625μg/ml. Each treatment was run in triplicate and inspected by different personnel for verification. 50μl of inoculum was added to each well of the 96-well plate and plate incubated overnight at 37°C for 18 hours. The MIC concentration was determined as the lowest concentration of treatment that resulted in no bacterial growth.
RAW 264.7 cells were cultured in six-well plates at a density of 2x106 cells per well for 24 hours. Cells were then pre-treated with 100μM TA or 0.01% DMSO for 30 minutes. The untreated and uninfected control were pretreated with fresh media. After pre-treatment, cells were infected at an MOI of 5, and centrifuged for 5 minutes at 500 rpm to allow the infection to reach the monolayer. At 90 minutes or 6 hours post infection (NF-ĸB/COX-2) cells were washed twice with sterile PBS (pH7.4) and removed utilizing 0.025% trypsin and neutralized with complete DMEM. Cells were fixed in 3.7% paraformaldehyde in PBS for 10 minutes and washed with PBS. Cell membranes were permeabilized with 0.01% triton X-100 in PBS for 10 minutes and the primary antibody stain was added at a dilution of 1:100. Samples were incubated at 42°C for 20 minutes, washed with PBS and re-suspended in 0.01% triton X-100 in PBS and the secondary antibody at a dilution of 1:100. Samples were washed and prepared for analysis in PBS. Flow cytometry was conducted on a Beckman Coulter CyAn ADP Flow Cytometer operating Summit v4.3 software for data collection. All further data analysis was done with FlowJo software. Samples were run in biological triplicate with two technical duplicates.
RAW 264.7 cells were cultured as described in 2.5. At each time point after infection, the supernatant was removed and 1% BHT solution was added to avoid the free-radical peroxidation as explained in [17]. Samples were then analyzed or stored at -20°C for later analysis. PGE2 analysis and quantification was conducted via ELISA using a PGE2 analysis kit (R&D Systems) in accordance with the manufacturer's instructions, except samples were not diluted as indicated. This experiment was conducted three independent experiments run in duplicate.
Ethics Statement: Animal experiments were performed in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Colorado State University Institutional Animal Care and Use Committee, protocol number 15-6138A. Six to eight week old, female BALB/c mice (Jackson Laboratories) were maintained under pathogen-free conditions and allowed free access to sterile food and water with 12 hour light/dark cycles. 45 mice were separated into 9 groups (n = 5/group) depending on treatment and survival regime.
For bacterial challenge, mice were anesthetized with ketamine/xylazine (100/10 mg/kg). The 1026b bacterial inoculum contained ~2×103 CFU suspended in 20 μl sterile saline and was delivered dropwise via pipet. Bacterial CFU were confirmed by plating the inoculum on LB agar. Treatments were initiated 3 hours post infection and repeated every 24 hours for two consecutive days as done by Asakrah et al [5]. Mice were euthanized when morbidity characteristics of hunched posture, loss of response to stimuli and loss of >20% body weight were reached. After euthanasia, the lung and spleen were removed and homogenized in 1 ml 0.9% sterile saline. Serial dilutions of tissue homogenates were plated on LB agar and bacterial CFU were counted after 2 days of incubation at 37°C.
For the effects of treatment on organ bacterial burden and dissemination, 30 mice were divided into 6 treatment groups (n = 5/group): TA (50 mg/kg) suspended in corn oil given via oral gavage, CZ (200 mg/kg and 25 mg/kg) dissolved in sterile PBS (pH 7.4) given subcutaneously, or co-treated with TA (50 mg/kg) at a sub-therapeutic dose of CZ (25 mg/kg), untreated, or vehicle control (corn oil) via oral gavage. Once the untreated infected control group showed signs of morbidity (typically around 60 hours post infection), mice were euthanized, and the lung and spleen from 3 mice (selected randomly) from each group were homogenized and plated for bacterial load determination.
For the survival study, mice were divided into 3 treatment groups (n = 5/group): 25 mg/kg CZ administered subcutaneously, 50 mg/kg TA administered via oral gavage in corn oil, and a co-treatment of 25 mg/kg CZ and 50 mg/kg TA. After the treatment, mice were monitored daily for signs of mortality for up the 37 days post-infection.
Statistical analyses was performed using Prism 6.0 software (Graphpad). Log-rank Mantel-Cox analysis was conducted for survival curves. All other data were analyzed using a one-way or two-way ANOVA followed by the Bonferroni post-test to determine statistical differences between groups or a two-tailed t-test for experiments with less than three groups. p<0.05 was considered statistically significant.
To determine if treatment with TA resulted in increased cell viability in-vitro, pretreated RAW 264.7 cells were infected with Bp82 at an MOI of 5 and cell viability was monitored over 4, 6, and 8 hours. Fig 1 reveals that Bp82 infection resulted in 44% reduction in cell viability after 4 hours, 66% after 6 hours, and 75% after 8 hours. RAW 264.7 cell cultures pretreated with 100μM TA revealed a reduction in cytotoxicity induced by Bp82 by 8% at 4 hours, by 42% at 6 hours and 30% at 8 hours when compared to the vehicle control group. Experimentation after 8 hours was not conducted as greater that 75% cytotoxicity was shown in later time points.
100 μM TA is a significant treatment, as treatments of this magnitude have shown to limit cell growth and cause cell death in colon cancer cells via activation of apoptosis pathways [9]. Therefore, we needed to confirm TA did not possess an inherent chemical nature that affected the bacterial growth and proliferation of Bp82. The MIC of TA exceeds 489 μM (Table 1), which exceeds the treatment doses used in this study. This suggests that the ability of TA in enhancing cell viability is a result of the immune modulatory properties of TA. Additionally, the MIC of DMSO was confirmed to be greater than 10%, confirming that the 0.01% vehicle concentration used in this study does not affect bacterial growth and proliferation. The MIC of CZ was used as an experimental control. The MIC of CZ fell within previously published results [16].
A combination of flow cytometry and western blotting were used to confirm Bp82 activation of the NF-ĸB pathway. Infection results in a 2.5 fold increase in p65 mean channel fluorescence (MCF) in RAW 264.7 cells (Fig 2) and pretreatment with TA reduced the MCF by 20% when compared to the DMSO control. This suggests that TA is able to significantly reduce p65 levels in Bp82 infected RAW 264.7 cells. S3 Fig confirms the effects of TA on p65 via western blot. This is consistent with previously published work as TA was shown to reduced p65 levels in LPS activated RAW 264.7 cells [10].
Infection with Bp82 resulted in a ~1.5-fold increase in COX-2 MCF from uninfected groups as presented in Fig 3. We also determined that pretreatment of RAW 264.7 cells pre-treated with 100 μM TA and infected with Bp82 resulted in a ~30% reduction in COX-2 MCF versus the vehicle control group. S5 Fig offers visual representation of this reduction in COX-2 via immunofluorescence.
It has been shown that PGE2 plays an important role during infection and its suppression is a possible therapeutic strategy [18]. Additionally, it has been shown that B. thailandensis infection results in increased production of PGE2, suggesting that PGE2 is necessary for the bacteria’s intracellular survival [5]. Our next goal was to confirm that Bp82 upregulates the production of PGE2 in RAW 264.7 cells and pretreatment with 100 μM TA reduces PGE2 production. Indeed, Bp82 infection resulted in 3-fold increase in supernatant PGE2 concentration after 4 hours of infection, and over a 4 fold increase in supernatant concentration at 6 and 8 hours. TA effectively reduced PGE2 levels by over 4-fold at all time points when compared to the vehicle control (Fig 4). Additionally, it would appear that pretreatment with TA reduced the PGE2 concentration to levels lower than the uninfected control, however, this was not significant.
The only treatment successful at significantly reducing the bacterial burden in both the lung and spleen was 200 mg/kg CZ (Fig 5). This confirms 200 mg/kg CZ administered subcutaneously is effective at reducing bacterial burden as shown in previous studies [19]. To assess if immune-modulation can potentiate the efficacy of CZ against an acute B. pseudomallei infection in the murine model, TA was co-administered with a sub-therapeutic dose of CZ. This data also indicates that 25 mg/kg CZ was ineffective at reducing the bacterial burden in the lung and spleen.
In TA-treated mice there was a 40% increase in survival time for 40% of the group (Fig 6). Co-treated TA/CZ group had 100% survival until 37 days post infection (pre-determined study endpoint). After 37 days, bacterial burden was assessed in the lung and spleen and compared to burden assessed at day 2.5 (Fig 7). Notably, the lung showed a reduction in bacterial burden of ~5x107 CFU/ml and the spleen a slight reduction of ~1 x 103 CFU/ml between 2.5 and 37 days post infection.
Combating the problem of antibacterial resistance is a global responsibility. This is particularly relevant for B. pseudomallei due to its ability to efflux many antibiotics [20]. Novel conjunctive therapies should be investigated to reduce the current antibacterial treatment regimes. Immunomodulation utilizing COX-2 inhibition during B. pseudomallei infection may prove an effective strategy to increase antibacterial effectiveness and treat the disease.
We corroborated previous studies by Asakrah et al [5] on cytotoxicity and the important role of COX-2 and PGE2 during melioidosis. These findings were expanded upon by investigating the effects of TA treatment on the NF-ĸB pathway during infection. We were able to mimic similar treatment outcomes using TA, which is approved for human use in certain areas around the globe [7]. Most importantly, we showed that the conjunctive treatment of immunomodulation with sub-therapeutic antibiotics significantly increased survival outcome and decreased organ bacterial burdens.
PGE2 has been linked to the regulation of a number homeostatic biological functions. Of greatest relevance here is that its involvement in initiating the classic signs of inflammation including redness and swelling due to PGE2-mediated arterial dilatation and increased vascular permeability and pain from PGE2 acting on sensory neurons [21] Additionally, all of the aforementioned mechanisms can lead to tissue damage. This compromised tissue provides an opportune environment for bacterial proliferation. Limiting inflammatory damage during bacterial infections has been shown to be an effective therapeutic strategy [6]. Our in-vitro studies revealed that limiting COX-2 induction and PGE2 production using TA translated to a significant increase is cell viability in mouse macrophage-like RAW 264.7 cells. This further confirms the results published by Asakrah et al [5]. Furthermore, pretreatment with TA was not only able to inhibit COX-2, as shown by the reduction of PGE2 production, but treatment also limited COX-2 induction. This may be linked to effects of TA on NF-ĸB and the interaction between the two inflammatory pathways.
NF-ĸB may prove an important inflammatory target during B. pseudomallei infection as it is believed to activate during infection and lead to the translation of many pro-inflammatory cytokines [3]. These cytokines amplify the systemic inflammatory response, often times by inducing COX-2 and resulting in continued production of PGE2 and other prostaglandins. PGE2 can have a positive feedback loop with NF-ĸB by increasing its transactivation and enhancing the production of pro-inflammatory cytokines [22]. Previous studies indicate TA is effective at reducing NF-ĸB activation in stimulated RAW 264.7 cells [10], but it is unclear whether this is linked to the PGE2/NF-ĸB relationship. Here, we show that treatment with TA reduces Nf-ĸB p65 during B. pseudomallei infection.
Immunomodulation may be necessary for comprehensive treatment of bacterial infections, particularly to combat bacterial resistance [6]. Our in-vivo studies reveal that immune modulation with orally administered TA is effective at increasing survival outcome in BALB/c mice. This is consistent with the findings of Asakrah et al [5]; although those studies were conducted using NS-398, an experimental COX-2 inhibitor given intraperitoneally. TA treatment alone did not show the same efficacy as NS-398 by Asakrah et al [5] but this may be due to the present dose being too low. More research is needed to determine if a larger dose of TA affords greater protection from a lethal B. pseudomallei pulmonary challenge.
Co-treatment of TA and CZ significantly increased survival outcome, and contributed to a nearly complete bacterial clearance after 37 days post infection. This suggests that the immunomodulation activity of TA allows for the exploitation and enhancement of the therapeutic benefit of CZ.
To the best of our knowledge, this is the first study indicating that immune modulation with orally-administered TA enhances the therapeutic benefit of antibiotic treatment during pulmonary melioidosis. The increased survival outcome resulting from the reduction of PGE2 production during bacterial infections through COX-2 inhibition and reduction in NF-ĸB activation has profound implications. PGE2 is produced during a number of lethal bacterial infections (i.e Francisella tularesis [23]) and opportunistic bacterial infections (Pseudomonas aeruginosa and Staphylococcus aereus) for which antibacterial treatment is complex due to antibacterial resistance [24] [25]. Proving the efficacy of immunotherapy using commercially available, orally administered TA in combination with sub-therapeutic antibiotic treatment during melioidosis in BALB/c mice warrants further investigation in other animal models of melioidosis. Such experiments may further reveal the effectiveness of TA and other NSAIDS across a wide range of bacterial infections. While this may have a profound impact on treatment of human bacterial infections, additional animal studies are needed to ensure efficacy of this treatment strategy before human clinical trials.
|
10.1371/journal.pgen.1004469 | A Novel MMP12 Locus Is Associated with Large Artery Atherosclerotic Stroke Using a Genome-Wide Age-at-Onset Informed Approach | Genome-wide association studies (GWAS) have begun to identify the common genetic component to ischaemic stroke (IS). However, IS has considerable phenotypic heterogeneity. Where clinical covariates explain a large fraction of disease risk, covariate informed designs can increase power to detect associations. As prevalence rates in IS are markedly affected by age, and younger onset cases may have higher genetic predisposition, we investigated whether an age-at-onset informed approach could detect novel associations with IS and its subtypes; cardioembolic (CE), large artery atherosclerosis (LAA) and small vessel disease (SVD) in 6,778 cases of European ancestry and 12,095 ancestry-matched controls. Regression analysis to identify SNP associations was performed on posterior liabilities after conditioning on age-at-onset and affection status. We sought further evidence of an association with LAA in 1,881 cases and 50,817 controls, and examined mRNA expression levels of the nearby genes in atherosclerotic carotid artery plaques. Secondly, we performed permutation analyses to evaluate the extent to which age-at-onset informed analysis improves significance for novel loci. We identified a novel association with an MMP12 locus in LAA (rs660599; p = 2.5×10−7), with independent replication in a second population (p = 0.0048, OR(95% CI) = 1.18(1.05–1.32); meta-analysis p = 2.6×10−8). The nearby gene, MMP12, was significantly overexpressed in carotid plaques compared to atherosclerosis-free control arteries (p = 1.2×10−15; fold change = 335.6). Permutation analyses demonstrated improved significance for associations when accounting for age-at-onset in all four stroke phenotypes (p<0.001). Our results show that a covariate-informed design, by adjusting for age-at-onset of stroke, can detect variants not identified by conventional GWAS.
| Ischaemic stroke places an enormous burden on global healthcare. However, the disease processes that lead to stroke are not fully understood. Genome-wide association studies have recently established that common genetic variants can increase risk of ischaemic stroke and its subtypes. In this study, we aimed to identify novel genetic associations with ischaemic stroke and its subtypes by addressing the fact that younger onset cases may have a stronger genetic component, and using this information in our analyses. We identify a novel genetic variant on chromosome 11 (rs660599), which is associated with increased risk of large artery stroke. We also show that mRNA expression of the nearest gene (MMP12) is higher in arteries with the disease process underlying large artery stroke (atherosclerosis). Finally, we evaluate our novel analysis approach, and show that our method is likely to identify further associations with ischaemic stroke.
| Genome-wide association studies (GWAS) in ischaemic stroke have begun to identify the common genetic variants that confer risk of the disease. However, there is considerable heterogeneity present in stroke phenotypes: GWAS analyses have primarily looked at the three main subtypes; cardioembolic (CE), large artery atherosclerosis (LAA) and small vessel disease stroke (SVD). Within these subtype analyses, numbers of cases are smaller, but the expectation is that the effects of SNPs identified within the subtypes will be considerably larger. Indeed, all validated GWAS SNPs for ischaemic stroke to date have been stroke subtype-specific [1], [2], [3], [4], [5], indicating the importance of subtyping of cases.
Clinical risk factors are important in stroke; as many as 77% of first-ever stroke patients are hypertensive [6], and other factors such as diabetes mellitus and elevated serum cholesterol confer a considerable proportion of disease risk [7]. These risk factors increase in prevalence in older age groups, suggesting older stroke patients may have a reduced stroke-specific genetic contribution. Indeed, IS is uncommon in individuals below middle age, but increases greatly in prevalence beyond the age of 65 [8], with a lifetime risk of 1 in 5 for women and 1 in 6 for men [9].
Under the assumptions of the liability threshold model, the low prevalence of IS in younger age ranges suggests that individuals who do suffer strokes in this age group are likely to have an increased genetic predisposition. This is supported by family history data; with stronger family history seen in younger onset cases [10], [11], [12], and twin studies [13], which suggest that early onset cases may have higher heritability. We recently showed stronger effects for all stroke-associated SNPs in younger age groups, found evidence genome-wide that a significant number of SNPs show stronger association p-values when the oldest cases are removed, and showed increased pseudoheritability estimates for younger onset cases in certain stroke subtypes, thereby supporting this hypothesis [14]. However, the question of how best to integrate this information into GWAS analyses of ischaemic stroke remains unanswered. Previous GWAS have analysed younger subsets of ischaemic stroke cases [1], [15], but this approach may not be optimal for existing GWAS datasets if the increase in odds ratios for SNPs in younger cases are not sufficient to justify discarding a large proportion of the ascertained cases. All previous young onset analyses have been restricted to all ischaemic stroke cases versus controls; this may be particularly relevant given that all known loci for ischaemic stroke to date are for stroke subtypes [16].
A recent publication [17], outlined a novel method of informing genetic association analyses on important clinical covariates. Using the liability threshold model in conjunction with estimates of disease prevalence for individuals with specific clinical covariates, the method estimates posterior disease liabilities for each individual in a GWAS, and uses these liabilities in regression analyses to test for association with genome-wide SNPs. This approach avoids issues due to multiple testing across age-at-onset thresholds, and provides a simple solution that is rooted is previous epidemiological research. In the present study, we extend the clinical covariate informed analysis approach to imputed genotypes, informing our analyses on the age-at-onset to identify novel variants associated with IS. We perform a genome-wide analysis with four stroke phenotypes (IS, CE, LAA, SVD), and then determine the utility of the approach in ischaemic stroke GWAS, testing whether SNPs increase in significance.
We performed age-at-onset informed association analysis for a total of 6,778 ischaemic stroke cases and 12,095 controls across four ischaemic stroke phenotypes; all IS and the three major subtypes: CE, LAA, and SVD (Table 1); with 1,637, 1,316, and 1,108 cases in the CE, LAA and SVD analyses respectively. With the exception of the young Milanese cohort, the age-at-onset distributions were similar in all cohorts (Table S3).
We identified a group of twenty SNPs proximal to MMP3 and MMP12 on chromosome 11 in the LAA subtype that met our criteria for replication. The strongest associated of these was rs662558 (p = 1.4×10−7), a SNP that is in 1000 Genomes, but not HapMap II. Therefore, to enable replication in existing METASTROKE datasets, which were imputed to HapMap II, we selected the most strongly associated SNP from the HapMap II panel, which was in perfect LD with the lead SNP in our discovery meta-analysis (rs660599: uninformed, p = 1.6×10−6; informed, p = 2.5×10−7; Figure 1) [16]. We found no evidence of between-study heterogeneity at either SNP (Cochran's Q p = 0.22 and p = 0.19 for rs662558 and rs660599, respectively). The evidence of an age-at-onset effect at rs660599 was p = 0.011 (from permutations). We calculated age-at-onset quartiles for all large artery stroke cases from the discovery cohorts, and used these to evaluate this region at different age-at-onset thresholds. The median age-at-onset was 71 years, and the interquartile range was between 61 and 78 years. Post-hoc analyses of rs660599 in the discovery cohorts using logistic regression (full details in Text S2) showed considerably stronger associations in younger age-at-onset quantiles (Q1; OR(95% CI) = 1.83 (1.46–2.30), Q1–Q2; 1.56 (1.33–1.83), Q1–Q3; 1.30 (1.14–1.49), Q1–Q4; 1.30 (1.15–1.46)). No other regions met our criteria for replication.
The associated locus was evaluated in a further 1,881 large artery stroke cases and ancestry matched controls in 9 cohorts from METASTROKE (Table 2). We found evidence for replication of the SNP (rs660599) in all large artery stroke cases of European Ancestry (p = 0.0048, OR(95% CI) = 1.18(1.05–1.32)). Combining this result with the discovery p-value gave a genome-wide significant p-value of 2.6×10−8 (Table 3). Secondly, we used the Han and Eskin random effects meta-analysis approach to evaluate the association [18] after including a further 355 cases and 1,390 controls of Pakistani ancestry. The evidence for replication in this sample was p = 0.0063, giving an overall p-value of 3.4×10−8. Age-at-onset information was available across all age-at-onset quantiles for a subset of the replication studies (1,240 cases, 9,238 controls; ASGC, HVH, ISGS/SWISS, MGH-GASROS, Utrecht). We evaluated the SNP (rs660599) in these studies at different age-at-onset quantiles using logistic regression, meta-analysing as previously. We again found the strongest effects in the youngest age quantile, consistent with a stronger effect in younger onset cases (Q1; OR(95% CI) = 1.27(1.02–1.57), Q1–Q2; 1.18(1.00–1.39), Q1–Q3; 1.22(1.05–1.40), Q1–Q4; 1.22(1.07–1.41)).
mRNA expression of the two proximal genes, MMP3 and MMP12 was analysed from 29 carotid, 15 abdominal aorta, 24 femoral plaques, and 28 atherosclerosis free left internal thoracic artery controls. MMP12 expression was upregulated in carotid plaques compared with left internal thoracic artery controls (P = 1.2×10−15; fold change [FC] = 335.6). It was also upregulated in femoral plaques (P = 3.2×10−14; FC = 306.0) and abdominal plaques (P = 5.0×10−11; FC = 399.3) compared with controls. Conversely, MMP3 was not significantly overexpressed in carotid, femoral or abdominal plaques versus controls (p>0.05).
Eight SNPs were identified that were perfect proxies (r2 = 1) with the associated SNP (rs660599) in the region. Seven of the SNPs were in an intergenic region between MMP3 and MMP12, while one fell within an intron of MMP12. We investigated the evidence that any of these SNPs are functional variants using RegulomeDB [19]. Of the eight SNPs, we found strong evidence that one of these SNPs (rs586701) affects binding. The SNP overlaps both CHIP-seq and DNA-seq peaks from ENCODE analyses, indicating that there is open chromatin in the region, and therefore that the SNP is likely to be functional. There is also evidence from a separate CHIP-seq analysis that the SNP affects protein binding [20], and evidence from multiple sources that the SNP overlaps a predicted motif [21], [22], [23]. Histone modifications were observed in CHIP-seq experiments from ENCODE in a number of cells types, including Human umbilical vein endothelial (Huvec) cells. Two other SNPs (rs17368582, rs2276109) in moderate LD with the associated SNP (r2 = 0.64) have been previously shown to directly influence MMP12 expression by affecting the affinity of an AP-1 binding site in the MMP12 promoter region [24], [25]. Using RegulomeDB, we found further evidence from ENCODE that one of these SNPs (rs2276109) is indeed functional, giving evidence that the associated locus in this analysis is likely to affect MMP12 expression through altered transcription. Detailed results for all analysed SNPs are given in Table S1. Additionally, we investigated if these SNPs (rs17368582, rs2276109, rs586701) were associated with MMP12 expression in tissues from the GTEx project [26]. However, we could not confirm an association with MMP12 expression in any relevant tissues (p>0.4 in whole blood, tibial artery, aortic artery).
Finally, we evaluated the overall utility of the age-at-onset informed approach in permutation analyses for SNPs that met p-value thresholds in the case control discovery data set. We generated 1000 permutations of age-at-onset within each centre, and performed age-at-onset informed analysis and subsequent meta-analysis for these SNPs, in the relevant stroke subtype.
We compared the sum of the meta-analysis Z scores from all SNPs with p<0.05 in the observed age at onset informed meta-analysis with those from permutations. At this p-value selection threshold, we found strong evidence (p<0.001) for genome-wide age-at-onset effects in each of the stroke phenotypes, with consistently increased summed Z scores in the observed age-at-onset informed meta-analysis compared to the permutations (Figure 2, red points, right hand axis). These results suggest that many of the risk variants for each stroke subphenotype have a higher frequency in younger onset cases. As the p-value selection threshold decreased, the summed Z score statistic became less significant in each stroke type, possibly reflecting lower overall power when fewer SNPs are included, even as these SNPs may have larger average effects. Further details are seen from the median proportion of SNPs more significant in the age-at-onset informed analysis than in the permutations (Figure 2, blue points, left hand axis). For CE and LAA stroke, the proportions increased with more stringent p-value thresholds (from 52.1% to 56.3% for p<0.05 and p<0.00005 thresholds in CE, and from 51.4% to 56.0% for p<0.05 and p<0.00005 thresholds in LAA). Interestingly, in the all ischaemic stroke analysis the median proportion of SNPs more significant in the observed results than permutations dropped from 55.1% for SNPs with p<0.05 to 49.2% for only SNPs with p<0.00005. This result may indicate a reduced proportion of true associations at stricter p-value thresholds for all ischaemic stroke compared to the subtypes, which is consistent with the observation that all common variants associated with stroke are for stroke subtypes, rather than for the phenotype of all ischaemic stroke [16].
The previously reported GWAS associations from a recent ischaemic stroke meta-analysis (9p21, HDAC9, PITX2, ZFHX3) were all found to be more significant using the age-at-onset informed approach than the uninformed analysis (Figure 3). The increase in significance ranged from over half an order of magnitude (7.9×10−9 to 1.5×10−9 for rs879324 in ZFHX3, CE), to under half an order of magnitude (5.7×10−9 to 2.5×10−9 for rs2107595 in HDAC9, LVD). To ensure these analysis methods were comparable, we calculated genomic inflation factors and plotted QQ-plots. These were similar in the standard and the age-at-onset informed approach (Table S4, Figure S1, S2). For these four associated SNPs, we further used the permuted data sets to assess the observation of increased significance in the age-at-onset informed analysis. We compared the observed meta-analysis p-value to those from the permutations, generating an empirical p-value by dividing the number of permutations more significant than the observed results by the number of permutations. In LAA stroke, we observed a significant age-at-onset effect (p = 0.018, 0.011 and 0.002 for the HDAC9, MMP12 and 9p21-associated SNPs in Figure 3, respectively). Similarly, for CE, we observed a significant age-at-onset effect for rs879324 (ZFHX3, p = 0.026), and a near-significant effect in rs6843082 (PITX2, p = 0.081). This result provides further evidence that risk variants associated with ischaemic stroke subtypes have a stronger role in younger onset cases, and suggests that the age-at-onset informed approach will produce improved significance when the magnitude of genetic effects are stronger in younger onset cases.
We used a large GWAS dataset to evaluate the utility of an age-at-onset informed analysis approach to ischaemic stroke, and to identify novel variants associated with ischaemic stroke phenotypes. We identified a novel MMP12 locus that is associated with large artery atherosclerotic stroke, and verified that the age-at-onset informed approach produces improved significance for loci associated with each of the stroke phenotypes studied, as well as demonstrating that it increased the significance of four previous GWAS associations with ischemic stroke, all without systematic inflation of the test statistic. Importantly, the novel associated SNP would not have been identified using a standard logistic regression framework.
We identified a group of SNPs proximal to Matrix Metalloproteinase 12 (MMP12) that showed increased significance when using the age-at-onset informed approach. The increase in significance from the equivalent uninformed analysis was of almost an order of magnitude (from p = 1.6×10−6 to p = 2.5×10−7 for rs660599). We took a single SNP from this region forward for replication in an independent dataset, finding further evidence that the region is associated with large artery stroke. Two SNPs (rs17368582, rs2276109) in this LD-block have previously been shown to directly influence MMP12 expression by affecting the affinity of an AP-1 binding site in the MMP12 promoter region [24], [25], and another variant in this block (rs17361668) is associated with increased fibrinogen levels, leading to an increased risk of developing advanced carotid atherosclerotic lesions, and an increased risk of myocardial infarction. We identified a second functional candidate (rs586701), which falls within both CHIP-seq and DNA-seq peaks from ENCODE, and is in complete LD with the associated SNP in our analysis.
We investigated mRNA expression of MMP12 and MMP3 in carotid atherosclerotic plaques in individuals from the Tampere Vascular Study. MMP12 was overexpressed in diseased tissue compared to healthy controls, while no significant difference was found for the other nearby gene, MMP3. MMP12 is a member of the Matrix Metalloproteinase (MMP) family of proteases, which are capable of degrading extracellular matrix proteins, and have a prominent role in atherosclerosis. They are thought to promote macrophage invasion [27], [28], [29], promote angiogenesis [30], and show increased activity in atheromatous plaques [31]. MMP12 deletions are associated with smaller, more stable lesions in the brachiocephalic artery of rabbits [32], and reduced elastin degradation in the aortic arch [33], indicating that MMP12 may have a role in destabilising plaques. Studies in humans have found MMP12 is localized to the core of advanced plaques, in macrophages with decreased arginase-I expression [34], that MMP12 localizes selectively to macrophages at the borders of the lipid core [35], and that MMP12 is significantly overexpressed in ruptured plaques when compared with thick or thin cap plaques, or with plaques with pathological intimal thickening [36]. This indicates that MMP12 is likely be involved in late-stage plaque instability: our study suggests that genetic variation impacts on this process.
Secondly, we performed extensive permutation analyses to assess the utility of the age-at-onset informed approach genomewide. In each phenotype studied we found evidence that SNPs were more strongly associated using the approach than would be expected by chance, indicating that multiple risk variants are likely to be more common in younger onset cases. The significance was strongest when more SNPs were included in the analysis, which likely reflects the cumulative impact of age-at-onset effects on many SNPs. An alternative explanation might be that the increased significance for lower p-value thresholds is the result of the cumulative effects of subtle confounding. However, this is unlikely because any subtle biases will also be present in the permutations, and should therefore not affect the significance of the results. This result supports observations from family history and prospective cohort studies, which have observed stronger effects in younger onset cases [6], [11]. Furthermore, all known associations with stroke were more significant using the age-at-onset informed approach. The increase in significance was around half an order of magnitude (e.g from p = 7.9×10−9 to 1.5×10−9 for ZFHX3, Figure 2), and was significant in all but one locus, as assessed by permutation. Taken together, these results indicate that age-at-onset is an important measure to stratify stroke cases, and show that, as expected by theory [17], integrating this information into association studies is likely to increase power to identify novel loci when the relative contribution of genetic is dependent on age-at-onset.
Our study has limitations. We used imputed data from the Immunochip platform, meaning we only had access to ∼40% of the genome across all centres. Secondly, cases were drawn from a number of international centres, meaning that despite efforts to standardize phenotyping, we cannot rule out differences in screening and clinical ascertainment.
Of complex diseases, IS has a particularly large degree of heterogeneity, exemplified by the fact that all validated associations identified to date have been within subtypes defined by clinical and radiological information. Further heterogeneity by risk factor and clinical covariate profiles is likely to exist, but the optimal method of incorporating this information into analyses remains an unanswered question. Our results indicate that a covariate-informed design, conditioning on age-at-onset of stroke, can unearth further associated variants. We provide evidence for this by identifying an association with a novel MMP12 locus in large artery stroke, supported by increased mRNA expression of the implicated gene in carotid plaques. GWAS in ischaemic stroke have begun to identify the genetic component of the disease, but these results are not yet clinically useful. Our study suggests that a more refined approach to analysis of genetic data, incorporating covariate information, is an important step in this process, and will help to ensure success in future GWAS.
All studies were approved by their local ethics committees; all patients gave informed consent.
The initial dataset consisted of 6,778 ischaemic stroke cases of European ancestry and 12,095 ancestry-matched controls from the Wellcome Trust Case-Control Consortium II project in ischaemic stroke [1], as well as a cohort from Milan, Italy [16]. These included 2,858 cases and 5,716 matched controls genotyped using the Immunochip platform; and 3,940 cases genotyped using either the Illumina 610 k or 660 k platforms matched with 6,379 controls genotyped on the Illumina Human 1.2M Duo (UK), Illumina Human 550 k (German) and Illumina 610 k platforms (Italian) (Table 1). The Immunochip cases were described in the previous WTCCC2 ischaemic study, where they formed the replication effort [1], as well as in a recent paper [37]. Genotyping of the five Immunochip case cohorts on the commercially available Immunochip array (Illumina, San Diego, CA, USA) was performed at the Sanger Centre, Hinxton, Cambridge UK. Swedish controls were provided and genotyped by the Swedish SLE network, Uppsala, Sweden. Belgian control samples were provided through the efforts of the International Multiple Sclerosis Genetics Consortium (IMSGC). German controls were derived from the PopGen biobank, [38]. UK controls were derived from the 1958 Birth cohort. Any of the 1958 Birth controls overlapping with those from the WTCCC2 datasets, as assessed by IBD estimates, were removed prior to analysis. Standard quality control procedures were undertaken on all centres, before centre-wise imputation to the 1000 Genomes phase 1 integrated variant set (March 2012), using IMPUTE v2.2.0 [39], [40]. SNPs with poor imputation quality (info<0.3) or low minor allele frequency (MAF<0.01) were discarded.
Ischemic stroke was defined as a typical clinical syndrome with radiological confirmation; ascertained cases were classified into individual stroke subtypes using the Trial of Org 10172 in acute stroke (TOAST) criteria in all centres [41]. Age-at-onset was defined as age at first hospital admission for stroke; where this information was unavailable, age at blood draw was used (7.3% of cases). The age-at-onset and gender distributions of the populations are given in Table S3. Age-at-onset quantiles were calculated from all the cases from the discovery datasets in the four stroke phenotypes (all IS and the three stroke subtypes: CE, LAA, SVD) and these were used to evaluate associated loci at different age-at-onset thresholds.
The prevalence of ischaemic stroke by age was obtained from a recent publication [9]; gender-specific estimates were averaged, and prevalences within each of the stroke subtypes were assumed to be approximately 20% of the overall total, similar to proportions seen in population-based studies [42]. We modeled phenotype data using a continuous unobserved quantitative trait called the disease liability, which we used to approximate the effect of age-at-onset on the liability scale, based on estimates of ischaemic stroke prevalence by age from epidemiological data (full details in Text S2). We developed two models for our analysis; one based on the prevalence rates for all ischaemic stroke cases, and secondly for the three stroke subtypes. We used these models to calculate posterior mean liabilities after conditioning on age-at-onset for the four stroke phenotypes separately. Controls were modeled in the same way, but were assumed to take the posterior mean from the lower (unaffected) portion of the distribution in the liability threshold model. Where age data was missing, individuals were assigned the median age value. Full descriptions of the models used and the formulae used to calculate posterior mean liabilities are given in Text S2. Regression was then performed on posterior liabilities by multiplying the number of samples by the squared correlation between the expected genotype dosage and posterior mean liabilities for each of the discovery cohorts in the four ischaemic stroke phenotypes (CE, LAA, SVD, IS), following a previous approach [17]. Ancestry-informative principal components were included where appropriate (6 of 8 centres), using the EIGENSTRAT procedure [43]. All analysis was performed using the R statistical software.
The results from each centre were meta-analysed for each of the four phenotypes using Stouffer's Z-score weighted approach, as implemented in METAL [44]. Genomic control was used to correct for any residual inflation due to population stratification [45]. Between-study heterogeneity was assessed using Cochran's Q statistic. We considered only SNPs present in at least 75% of the cases, and with no evidence of heterogeneity (Cochran's Q p-value>0.001). All SNPs analysed were either genotyped or imputed in both the Immunochip and the genome-wide datasets. After meta-analysis, the resulting p-values were compared with the equivalent values from an unconditioned analysis. For SNPs more significant in the age-at-onset informed analysis and with p<5×10−6, we determined the evidence of a true age-at-onset effect by generating 1000 permutations of age-at-onset and rerunning the age-at-onset informed analysis, meta-analysing as previously. We calculated an empirical p-value by dividing the number of permuted observations showing greater significance in the meta-analysis than the observed results by the number of permutations. Any novel SNP with a meta-analysis p<5×10−6 and evidence of an age-at-onset effect at p<0.05 were taken forward for replication. We set the experiment-wide significance threshold at p<5×10−8.
Replication of an associated variant was performed in a further 10 cohorts from METASTROKE. Nine of the centres used a cross-sectional design, while one was a large prospective, population based cohort (ARIC). Nine of the centres were of European ancestry, while one consisted of individuals of Pakistani ancestry (RACE) (Table 2). All centres used a case-control methodology; centres with a cross sectional design used logistic regression to model the association of genotype dosages from imputation with the dichotomous outcome of ischaemic stroke and prospective cohorts used Cox proportional-hazards models to evaluate time to first stroke, fitting an additive model relating genotype dose to the stroke outcome. European ancestry replication centres were meta-analysed using a fixed effects inverse-variance weighted method. To assess the evidence for association of the SNP for replication samples of all ancestries, we performed a trans-ethnic meta-analysis using a random-effects model to control for any resulting heterogeneity [18]. To evaluate the overall evidence for association, the results of the discovery and replication analyses were combined using Fisher's Method.
Expression of the two genes proximal to the associated variant was tested in atherosclerotic plaques from the Tampere Vascular study [27], [46], [47], [48], [49]. Carotid, femoral, and aortic atherosclerotic plaques constituting the intima and inner media were prospectively obtained between 2005 and 2009 from patients fulfilling the following inclusion criteria: (1) carotid endarterectomy attributable to asymptomatic or symptomatic >70% carotid stenosis, or (2) femoral or (3) aortic endarterectomy with aortoiliac or aortobifemoral bypass attributable to symptomatic peripheral arterial disease. Whole thickness left internal thoracic artery samples obtained during coronary artery bypass surgery and identified as being microscopically atherosclerosis free were used as controls. The patients were consecutively recruited and stratified according to indication for surgery. All open vascular surgical procedures were performed at the Division of Vascular Surgery and Heart Center, Tampere University Hospital.
Fresh tissue samples were immediately soaked in RNALater solution (Ambion Inc) and homogenized using an Ultra-Turrax T80 homogenizer (IKA). RNA was extracted with the Trizol reagent (Invitrogen) and miRNEasy Mini-Kit (Qiagen) with the RNase-Free DNase Set (Qiagen) according to manufacturer instructions. The RNA isolation protocol was validated by analyzing the integrity of the RNA with the RNA 6000 Nano Chip Kit (Agilent). The expression levels were analyzed with an Illumina HumanHT-12 v3 Expression BeadChip (Illumina). In brief, 300–500 ng of RNA was reverse transcribed in cRNA and biotin-UTP labeled using the IlluminaTotalPrep RNA Amplification Kit (Ambion), and 1500 ng of cRNA was then hybridized to the Illumina HumanHT-12 v3 Expression BeadChip.
The BeadChips were scanned with the Illumina iScan system. After background subtraction, raw intensity data were exported using the Illumina Genome Studio software. Further data processing was conducted by means of R language and appropriate Bioconductor modules. Data were log2-transformed, and robust multichip average and robust spline normalization (rma_rsn) were used. Accuracy of the expression array was validated with qRT-PCR [50]. mRNA Expression levels in the tissues were determined; a fold change statistic was estimated between the two tissues, and significance was calculated using a t test.
Recent evidence indicates that a significant proportion of GWAS SNPs fall within regions that are likely to affect binding of nearby proteins, such as transcription factor binding sites [51], [52]. We used the RegulomeDB database to access regulatory information from ENCODE and other existing publications [19], investigating the evidence that the SNPs in the associated locus have a regulatory function. First, the linkage-disequilibrium (LD) patterns amongst the most strongly associated SNPs were determined. We then used PLINK to determine the LD structure of the associated region, using LD-patterns from the 85 Utah residents from the 1000 Genomes project [53], [54]. All SNPs with r2>0.6 were identified within a 2,000 kb window from the index SNP. All of the SNPs identified were then investigated using RegulomeDB to determine the evidence that any of the SNPs have a regulatory function.
Permutation analysis was performed to evaluate the age-at-onset informed approach, to show that including age at onset information directly led to the increased significance, due solely to inclusion of age-at-onset information at tested SNPs. First, we identified a set of SNPs enriched for true association in the case control analysis of ischaemic stroke and subtypes. An expanded set of discovery and METASTROKE studies were analysed using standard case control methods and subsequent meta-analysis (see Table S2). SNPs with p<0.05 and no evidence of heterogeneity (p>0.0001) were extracted and pruned for LD (300 kb window, r2<0.25), leaving a set of almost independent SNPs for further analysis. Each retained SNP represented the most significant association in each LD block, as determined by the “clump” procedure in PLINK, based on LD patterns from the CEU individuals from 1000 Genomes. The number of SNPs used in each analysis is given in Table S5. These SNP subsets were derived for ischaemic stroke, and for each stroke subset and then used in the age-at-onset informed analysis. Analysis was performed as previously for each stroke subtype using the age-at-onset informed method within studies and meta-analysis across studies (giving observed results, as obtained above). We then performed a permutation study to obtain the expected distribution of p-values at these SNPs. Age at onset for cases was permuted within stroke subtypes within each study, and then the data were re-analysed, for 1000 permutations. Two summary statistics were constructed: (1) within permutations, we compared p-values from analysis of permuted age at onset with p-values from the observed data, and tabulated the proportion of SNPs with increased significance in the observed data set than in the permuted data set; across permutations, we calculated the median proportion of SNPs with increased significance in the observed data; (2) Within permutations, we converted each SNP p-value to a Z score and summed the absolute value of the Z score across SNPs (sumZ). An empirical p-value for the age-informed analysis was calculated from the proportion of simulated data sets where sumZ exceeded the value in the observed analysis. This analysis was performed at SNP subsets defined from four SNP p-value thresholds in the discovery and METASTROKE studies: p<0.05, p<0.005, p<0.0005, and p<0.00005.
Finally, we assessed the evidence of an age-at-onset effect at the four stroke loci identified in the METASTROKE ischaemic stroke collaboration (9p21, HDAC9, PITX2, ZFHX3) [16]. For each SNP, we generated an empirical p-value from the proportion of permutations showing stronger association than in the observed age-at-onset informed analysis.
|
10.1371/journal.pbio.1002028 | Evolution of RNA-Protein Interactions: Non-Specific Binding Led to RNA Splicing Activity of Fungal Mitochondrial Tyrosyl-tRNA Synthetases | The Neurospora crassa mitochondrial tyrosyl-tRNA synthetase (mtTyrRS; CYT-18 protein) evolved a new function as a group I intron splicing factor by acquiring the ability to bind group I intron RNAs and stabilize their catalytically active RNA structure. Previous studies showed: (i) CYT-18 binds group I introns by using both its N-terminal catalytic domain and flexibly attached C-terminal anticodon-binding domain (CTD); and (ii) the catalytic domain binds group I introns specifically via multiple structural adaptations that occurred during or after the divergence of Peziomycotina and Saccharomycotina. However, the function of the CTD and how it contributed to the evolution of splicing activity have been unclear. Here, small angle X-ray scattering analysis of CYT-18 shows that both CTDs of the homodimeric protein extend outward from the catalytic domain, but move inward to bind opposite ends of a group I intron RNA. Biochemical assays show that the isolated CTD of CYT-18 binds RNAs non-specifically, possibly contributing to its interaction with the structurally different ends of the intron RNA. Finally, we find that the yeast mtTyrRS, which diverged from Pezizomycotina fungal mtTyrRSs prior to the evolution of splicing activity, binds group I intron and other RNAs non-specifically via its CTD, but lacks further adaptations needed for group I intron splicing. Our results suggest a scenario of constructive neutral (i.e., pre-adaptive) evolution in which an initial non-specific interaction between the CTD of an ancestral fungal mtTyrRS and a self-splicing group I intron was “fixed” by an intron RNA mutation that resulted in protein-dependent splicing. Once fixed, this interaction could be elaborated by further adaptive mutations in both the catalytic domain and CTD that enabled specific binding of group I introns. Our results highlight a role for non-specific RNA binding in the evolution of RNA-binding proteins.
| The acquisition of new modes of post-transcriptional gene regulation played an important role in the evolution of eukaryotes and was achieved by an increase in the number of RNA-binding proteins with new functions. RNA-binding proteins bind directly to double- or single-stranded RNA and regulate many cellular processes. Here, we address how proteins evolve new RNA-binding functions by using as a model system a fungal mitochondrial tyrosyl-tRNA synthetase that evolved to acquire a novel function in splicing group I introns. Group I introns are RNA enzymes (or “ribozymes”) that catalyze their own removal from transcripts, but can become dependent upon proteins to stabilize their active structure. We show that the C-terminal domain of the synthetase is flexibly attached and has high non-specific RNA-binding activity that likely pre-dated the evolution of splicing activity. Our findings suggest an evolutionary scenario in which an initial non-specific interaction between an ancestral synthetase and a self-splicing group I intron was fixed by an intron RNA mutation, thereby making it dependent upon the protein for structural stabilization. The interaction then evolved by the acquisition of adaptive mutations throughout the protein and RNA that increased both the splicing efficiency and its protein-dependence. Our results suggest a general mechanism by which non-specific binding interactions can lead to the evolution of new RNA-binding functions and provide novel insights into splicing and synthetase mechanisms.
| RNA-binding proteins play critical roles in post-transcriptional regulation of gene expression in all domains of life [1]. However, the complexity of this regulation is far greater in eukaryotes than in prokaryotes, reflecting both the larger number of RNAs requiring regulation and the evolution of new RNA processing and regulatory mechanisms. The latter include extensive RNA splicing and alternative splicing to produce different protein isoforms; an increased importance of RNA localization in larger and more complex eukaryotic cells; nonsense-mediated decay to prevent translation of intron-containing RNAs; and combinatorial regulation of mRNA translation and stability by RNA-binding proteins and miRNAs acting in ribonucleoprotein complexes [2]–[5]. These new modes of post-transcriptional regulation necessitated and were enabled by corresponding increases in the number and diversity of RNA-binding proteins and the evolution of new RNA-binding functions [3],[6]. Thus far, however, the molecular mechanisms underlying the evolution of new RNA-binding functions have remained unclear.
Cellular proteins that adapted to splice autocatalytic group I and group II introns provide powerful model systems for investigating how proteins evolve new RNA-binding functions. Group I and group II introns are found in prokaryotes and in the mitochondrial (mt) and chloroplast DNAs of some eukaryotes, with group I introns also found in the nuclear rRNA genes of certain fungi and protozoa [7],[8]. Both types of introns are ribozymes that catalyze their own splicing as well as mobile genetic elements that can be horizontally transferred to different hosts where they propagate by inserting into new genomic sites [9],[10]. Although some group I and II introns self-splice in vitro, most have acquired mutations that impair formation of a catalytically active RNA structure, necessitating the recruitment of cellular proteins to promote RNA folding for efficient splicing in vivo [11],[12]. These group I and group II intron splicing factors include both host-encoded proteins, such as aminoacyl-tRNA synthetases (aaRSs) and translation factors, and intron-encoded proteins, such as DNA endonuclease and reverse transcriptases, that evolved secondarily to function in RNA splicing [11]. Such co-option of pre-existing proteins to function in splicing is pertinent to the evolution of splicing mechanisms in higher organisms, as emphasized by recent findings that a key spliceosomal protein, Prp8, was derived from a group II intron-like reverse transcriptase [13],[14].
One of the most extensively studied examples of a cellular protein that evolved to function in RNA splicing is the Neurospora crassa mtTyrRS (CYT-18 protein), which acts as a splicing factor for mt group I introns [15]–[18]. Biochemical and structural studies showed that CYT-18 functions in splicing by recognizing and stabilizing the conserved phosphodiester backbone structure of group I intron RNAs [19]–[22]. This splicing function has been found only for those mtTyrRS of fungi belonging to the subphylum Pezizomycotina and can be traced to a series of structural adaptations of the protein that were acquired during or after the divergence of Pezizomycotina from Saccharomycotina [23].
CYT-18 and other mtTyrRSs are class 1 aaRSs that are closely related to bacterial TyrRSs [24]. They consist of an N-terminal catalytic domain, which binds the acceptor stem of tRNATyr, followed by an intermediate α-helical domain and a C-terminal anticodon-binding domain (CTD), which bind the anticodon and variable arms (Figure 1A and 1B; the catalytic and intermediate α-helical domains together are denoted the N-terminal domains or NTDs). Like its bacterial counterparts, CYT-18 functions as a homodimer, with each dimer binding either one molecule of tRNATyr or group I intron RNA [25]–[28]. CYT-18 binds group I introns by using both its N-terminal catalytic domain and CTD, but only some introns require the CTD for RNA splicing [29]–[31].
Both the N-terminal catalytic domain and CTD of Pezizomycotinia mtTyrRSs have distinctive structural adaptations that are absent in non-splicing mtTyrRSs, including the closely related Saccharomycotina mtTyrRSs [23]. These structural adaptations include a small N-terminal α-helical extension (denoted H0) and a series of small insertions (Ins 1–5), whose presence correlates with RNA splicing activity (Figure 1B) [32],[33]. The Pezizomycotina mtTyrRS also have a non-essential C-terminal tail of variable length (C-tail; 13–152 amino acids) appended to the CTD ([29] and this work).
Structural studies, including a co-crystal structure of a splicing-active CYT-18 protein lacking the CTD (here denoted CYT-18 NTDs) bound to a group I intron RNA (the bacteriophage Twort orf142-I2 ribozyme), provided insight into group I intron binding by the N-terminal catalytic domain [22],[33]. These studies showed that CYT-18 binds group I introns asymmetrically across the two subunits of the homodimer by using a newly evolved group I intron-binding surface on the side of the catalytic domain opposite that which binds tRNATyr. This new RNA-binding surface includes the N-terminal extension H0, Ins 1, and Ins 2 and provides an extended scaffold for the conserved phosphodiester backbone structure of the group I intron catalytic core.
The CYT-18 constructs used for crystallography lacked the flexibly attached CTD, which has been problematic for X-ray crystallography of TyrRSs [33]–[36]. Recently, we determined an NMR structure of the isolated CTD of the splicing-active Aspergillus nidulans mtTyrRS, which is closely related to CYT-18 [37]. The structure showed that the mtTyrRS CTD resembles those of bacterial TyrRSs in having a fold similar to that of bacterial ribosomal protein S4, but with novel structural features. The latter include three Pezizomycotina-specific insertions (Ins 3–5), with Ins 3 corresponding to an expansion of the flexible linker between the NTDs and CTD. Modeling of the NMR structure onto the CYT-18 NTDs+Twort co-crystal structure using distance constraints from directed hydroxyl-radical cleavage assays suggested that the two CTDs of the homodimeric protein bind opposite ends of a group I intron RNA. This model requires that the CTD of one subunit of the CYT-18 homodimer undergo a large shift on its flexible linker to interact with either tRNATyr or the group I intron RNA bound on opposite sides of the catalytic domain [37]. Thus far, however, there has been no structural data for a CYT-18 protein that contains both the NTDs and CTD, and the role of the CTD in promoting group I intron splicing has remained unclear.
CYT-18 has been used as a model for the theory of constructive neutral evolution (referred to here as “pre-adaptive evolution”). This theory holds that complex multi-protein or RNP complexes arise by a “ratchet-like process” in which a pre-existing neutral or mildly deleterious interaction is “fixed” by a mutation in one partner that makes it dependent upon the other to perform a biological function. Once fixed, this dependence can be further elaborated by adaptive changes in both partners, which increase reaction efficiency and co-dependence [38]–[40]. In the case of CYT-18, this hypothesis suggests that an ancestral non-splicing fungal mtTyrRS had a pre-existing ability to bind group I introns, which became fixed when the intron RNA acquired mutations that impaired self-splicing, resulting in dependence upon the bound protein for structural stabilization [11]. After the interaction was fixed, further adaptive mutations in both the RNA and protein increased both the efficiency of RNA splicing and its protein-dependence. Early studies suggesting that CYT-18 recognized tRNA-like structural features of group I intron RNAs were cited as a prime example of a pre-adaptive interaction leading to the evolution of a new RNA-splicing function [39],[41]. However, subsequent findings that CYT-18's N-terminal catalytic domain binds group I introns specifically by using a separate non-tRNA-binding surface [22],[33] made the nature of the initial non-adaptive interaction unclear.
Here, we used small angle X-ray scattering (SAXS) and biofchemical assays to investigate the solution structures of full-length CYT-18 protein and its CTDs and their mode of interaction with group I intron RNAs. The SAXS analysis shows that the CTDs of both subunits of the CYT-18 homodimer extend outward from the NTDs, but move inward to bind opposite ends of the group I intron RNA. Surprisingly, we find that the CTD of CYT-18 has a high non-specific RNA binding affinity, which may contribute to its interaction with group I intron RNAs, and that the Saccharomyces cerevisiae (yeast) mtTyrRS, which diverged prior to the evolution of splicing activity, can also bind intron RNAs non-specifically via its CTD. Finally, experiments with chimeric proteins show that the yeast CTD can replace CYT-18's to promote aminoacylation but not group I intron splicing. Our results suggest a scenario of pre-adaptive evolution in which the initial non-adaptive interaction between an ancestral mtTyrRS and group I intron RNA was non-specific binding by the CTD and highlight a role for non-specific binding in the evolution of new RNA-binding functions.
First, we used SAXS to investigate the conformational changes of CYT-18 and the position of its CTDs in the absence and presence of a group I intron RNA. Scattering data were collected for three CYT-18 constructs: CYT-18*, a wild-type protein truncated to delete most of the non-essential C-tail in order to simplify modeling and analysis; CYT-18 NTDs, which contains the N-terminal catalytic and α-helical domains, but lacks both the CTD and C-tail; and CTD, the isolated C-terminal anticodon-binding domain (Figure 1C). CYT-18* is fully active in tyrosyl-adenylation, which measures the number of TyrRS active sites, and it functions similarly to full-length CYT-18 both in aminoacylation of Escherichia coli tRNATyr, a standard substrate for this protein, and in splicing the N. crassa mt large subunit rRNA (Nc mt LSU) intron, which requires a functional CTD (Figure S1). The CYT-18 NTDs construct is also fully active in tyrosyl-adenylation, but cannot aminoacylate tRNATyr as expected because of the lack of the CTD (Figure S1A and S1B) [30].
Figure 2 shows SAXS curves for all three proteins, and Table 1 summarizes size parameters calculated from the SAXS curves, including the protein molecular weight; the maximum dimension of the particle (Dmax); and the radius of gyration (Rg), which is the root mean square distance to the center of mass of a particle and provides an estimate of the overall particle size [42]. For all three proteins, Kratky plots of the SAXS data show a bell shape curve with a distinct peak, indicative of a folded globular protein (Figure S2).
Focusing first on the CYT-18 NTDs protein, the scattering curve overlays well (χ = 1.9) with a theoretical scattering curve calculated from the previous CYT-18 NTDs crystal structure [33] by using the program CRYSOL (Figure 2A, top curve) [43]. The SAXS curve gave an estimated molecular weight of 84.4 kDa and Rg and Dmax values of 35.6 and 123 Å, respectively, in good agreement with the molecular weight calculated from protein sequence (89.6 kDa) and with Rg and Dmax values calculated from the crystal structure using CRYSOL (35.2 and 125 Å, respectively) (Table 1). The distance distribution function P(r) for the CYT-18 NTDs displays a single peak with a tail (Figure 2B), a pattern indicative of a protein having an elongated structure [44]. Ab initio models of the CYT-18 NTDs protein were built from the SAXS data by simulated annealing of either dummy atoms by DAMMIN or a chain-like ensemble of dummy residues by GASBOR (Figures 2C and S3, respectively) [45],[46]. The DAMMIN and GASBOR models show good fits to the experimental SAXS curve (χ = 1.8 for both models) and are similar in shape to each other and to the high-resolution structure as shown by the superposition of the crystal structure within the SAXS model envelopes. The final DAMMIN and GASBOR models are the result of analyzing multiple solutions and either averaging the models (DAMMIN) or picking the most representative one (GASBOR). The normalized spatial discrepancy (NSD) value, which describes the similarity between the different models produced by the programs, is low for both the DAMMIN and GASBOR models (0.63±0.03 and 1.10±0.02, respectively), indicating that the multiple solutions built by the programs are similar to each other (Table 2). Taken together, these results indicate that the conformation adopted by CYT-18 NTDs in solution is similar to that in the crystal structure [33].
The SAXS data for the isolated CTD overlays well with a theoretical scattering curve calculated from a homology model of CYT-18's CTD constructed from the NMR structure of the A. nidulan CTD using I-TASSER (χ = 2.8) (Figure 2A, middle curve) [47]. The molecular weight of 13.3 kDa estimated from the scattering data (Table 1) indicates that the CTD is monomeric in solution. The Rg and Dmax values for the CTD calculated from the SAXS data (17.7 and 62 Å, respectively) are in good agreement with those for the I-TASSER model (17.0 and 55.9 Å, respectively) (Table 1). The DAMMIN and GASBOR models of the CYT-18 CTD (χ = 1.7 and 2.2, respectively) also superpose well with the homology model (Figures 2D and Figure S3, respectively). Thus, the SAXS analysis indicates that the CYT-18 CTD folds independently of the remainder of the protein and that the I-TASSER model provides a good representation of the structure of the CYT-18 CTD in solution. The latter finding validates the use of the I-TASSER model in building high-resolution structures of CYT-18* from the SAXS data (see below).
CYT-18* is the first CYT-18 protein to be investigated structurally that contains both the NTDs and CTD. The molecular weight for this protein estimated from the SAXS data is 119 kDa, confirming that CYT-18* is a dimer in solution (Table 1). The Rg and Dmax values from the CYT-18* scattering data are 46.9 and 170 Å, respectively, both larger than that for the CYT-18 NTDs, as expected. Ab initio models of the CYT-18* homodimer indicate an open conformation with both CTDs extending outward from the NTDs (χ = 1.5 and 2.1 for the DAMMIN and GASBOR models, respectively) (Figures 2E and S3). A rigid-body model of CYT-18* was also built by CORAL, which constructs models that fit the SAXS data by combining high-resolution models of individual components, in this case the crystal structure of the CYT-18 NTDs and the I-TASSER model of the CTD (see above), with different conformations of flexible dummy residue linkers [48]. The CORAL model overlays well with the scattering curve (χ = 1.8) (Figure 2A, bottom curve) and superposes well into the SAXS envelopes of the ab initio models (Figures 2E and S3). These findings indicate that in the absence of intron RNA, CYT-18* adopts an S-shaped configuration with the two CTDs of the homodimer extending outward in opposite orientations.
The co-crystal structure of the CYT-18 NTDs bound to Twort intron RNA indicated that the RNA binds asymmetrically across the dimer interface and that the structure of the NTDs does not change substantially upon binding the intron RNA [22]. To elucidate interacting regions and conformational changes of the CTDs upon binding to the intron RNA, we obtained SAXS data for complexes of both the CYT-18 NTDs and CYT-18* bound to the same Twort group I intron RNA. The experimental scattering curve of the CYT-18 NTDs+Twort RNA complex overlays reasonably well with the scattering curve calculated from the co-crystal structure (χ = 4.4) (Figure 3A, top curve), and gave Rg and Dmax values (39.2 and 137 Å, respectively) in agreement with those calculated from the co-crystal structure (39.1 and 134 Å, respectively) (Table 1). Likewise, a rigid-body model of the CYT-18 NTDs+Twort RNA complex built using CORAL shows a good fit to the experimental SAXS curve (χ = 2.2) (Figure 3B; top curve) and is similar in shape to the co-crystal structure (Figures 3A and 3B, compare insets above the top curves). These findings indicate that the CYT-18 NTDs+Twort RNA co-crystal structure is similar to the structure of the complex in solution and can be used as a component for structural modeling of the CYT-18*+Twort RNA complex from the SAXS data.
Finally, the scattering data for the Twort RNA complex with CYT-18*, which contains both the NTDs and CTD, gave Rg and Dmax values of 41.9 Å and 146 Å, respectively (Figure 3A, bottom curve; Table 1). The relatively small difference between these values and those for the CYT-18 NTDs+Twort complex (Rg and Dmax values of 39.2 Å and 137 Å, respectively; Table 1) suggests that CYT-18* forms a compact complex with the RNA in which the CTDs make a smaller than expected contribution to the overall particle dimensions. This conclusion was supported by ensemble optimization analysis using the program EOM, which generates a large random pool of conformations and picks an optimized ensemble that best fits the scattering data (see Materials and Methods). This optimized ensemble pool displayed a smaller, tighter range of Rg and Dmax values than a random pool of protein-RNA conformations consistent with a compact rigid complex (Figure S4).
A rigid-body model of the CYT-18*+Twort RNA complex built using CORAL indicates that both CTDs are positioned near the Twort intron (Figure 3C). This CORAL model can be compared to a previous model of CYT-18*+Twort built using biochemical data (Figure 3D) [37]. While both the CORAL and biochemical models show that both CTDs are located near the intron RNA, the CORAL model better fits to the scattering data than does the biochemical model (χ = 3.2 and 8.8, respectively). In both models, the CTD of one subunit of the CYT-18 homodimer is close to and may interact with P2, P6–P6a, and P8 of the intron RNA, while the CTD of the other subunit is close to and may interact with P4–P5, and P9 of the intron RNA (Figure 3C and 3D). Considered together, the SAXS analyses indicate that upon binding a group I intron RNA, CYT-18* forms a compact complex in which both CTDs of the CYT-18 homodimer clamp down to interact with opposite ends of the group I intron RNA.
To investigate how the RNA-binding properties of CYT-18's CTD enable it to interact with the two structurally distinct ends of a group I intron RNA, we analyzed the interaction of the CYT-18 NTDs and the isolated CTD with various RNAs by equilibrium-binding assays at 25°C and 37°C (Figures 4 and S5, respectively). The RNAs compared were three group I introns (the N. crassa mt large ribosomal subunit-ΔORF intron (Nc mt LSU); the N. crassa NADH dehydrogenase subunit 1-ΔORF intron (Nc ND1m); and the Twort intron RNA); a group II intron RNA (Lactococcus lactis Ll.LtrB-ΔORF; Ll.LtrB); and poly(U)30, which presumably lacks higher-order structure.
The binding curves for the CYT-18 NTDs to the Nc mt LSU, Nc ND1m, and Twort group I intron RNAs were best fit by hyperbolic functions with Kds ranging from 200 to 590 nM at 25°C (Figure 4A–4C) and 440 to 740 nM at 37°C (Figure S5A–S5C). The Kd values for the Nc ND1m intron are substantially higher than that calculated from previous koff measurements, which assumed that the kon of the construct lacking the CTD is the same as that for the wild-type protein (71±24 pM) [30]. This difference suggests that the CTD might make a major contribution to kon by mediating the initial interaction with intron RNA substrates. At both temperatures, the strongest binding group I intron RNA was the Nc ND1m intron and the weakest was the Nc mt LSU intron, consistent with previous findings that the CTD is required for tight binding and splicing of the Nc mt LSU, but not the Nc ND1m intron [30]. At 25°C, the CYT-18 NTDs bound the Ll.LtrB group II intron RNA with a K1/2 = 240 nM, within the range of Kds for group I intron RNAs, but the binding curve was sigmoidal, with n = 1.6, suggesting cooperative and possibly non-specific binding (Figure 4D), whereas at 37°C, binding of the group II intron RNA was weaker (Kd = 440 nM) and the binding curve was hyperbolic (Figure S5D). The CYT-18 NTDs did not bind appreciably to poly(U)30 at either 25°C or 37°C (Figures 4E and S5E).
Surprisingly, the isolated CYT-18 CTD bound group I and II intron RNAs and poly(U)30 more strongly than did the CYT-18 NTDs and with similar affinities for all five RNAs tested (Figure 4). Indeed, the binding curves for the isolated CTD to these radically different RNAs were remarkably similar to each other, each being sigmoidal with K1/2s = 57–64 nM at 25°C and 51–77 nM at 37°C. These sigmoidal binding curves (i.e., Hill coefficients (n)>1) suggest cooperative binding of two or more CTDs to each RNA. The ability of the isolated CTD to bind similarly to group I and II intron RNAs, as well as unstructured poly(U)30 indicates that it is a non-specific RNA binding domain.
The finding that the CYT-18 CTD is a non-specific RNA-binding domain led us to wonder whether an ancestral Pezizomycotina mtTyrRS might have initially bound group I intron RNAs non-specifically. To address this question, we turned to the S. cerevisiae (Sc) mtTyrRS, which is closely related to CYT-18 but branched from the Pezizomycotina mtTyrRSs prior to the evolution of splicing activity [23]. We compared two constructs that were expressed in E. coli: recombinant wild-type Sc mtTyrRS and a derivative lacking the CTD (Sc NTDs). We confirmed that both proteins are fully active in tyrosyl-adenylation, indicating correct folding of the catalytic domain (Figure S1A). Aminoacylation assays showed that the Sc mtTyrRS has higher activity with E. coli tRNATyr than does CYT-18 (Figure S1B), possibly reflecting that the E. coli tRNATyr and the Sc mt tRNATyr are more similar to each other than to the Nc mt tRNATyr. All three tyrosyl-tRNAs share the same N73 identity element (A73) and anticodon, but differ in the N1-N72 identity element at the end of the acceptor stem (G-C in E. coli tRNATyr and Sc mt tRNATyr, but A-U in Nc mt tRNATyr) and the length of the variable arm (13–14 nt in E. coli tRNATyr and Sc mt tRNATyr, but unusually long at 16 nt in Nc mt tRNATyr) (Figure S6) [49]–[51].
Equilibrium binding assays showed that the Sc mtTyrRS, although incapable of splicing group I intron RNAs [23], can bind both the Nc mt LSU group I intron RNA and Ll.LtrB group II intron RNA with Kds = 430 and 440 nM, respectively (Figure 5A and 5B), within the range of Kds for specific binding of CYT-18 NTDs to group I intron RNAs (see above). However, the similar affinity of the Sc mtTyrRS for the group I and group II intron RNAs suggests that this binding is non-specific. Strikingly, the ability of the Sc mtTyrRS to bind group I and II intron RNAs was entirely dependent upon its CTD, with the Sc NTDs protein showing no detectable binding of either intron RNA over the concentration range tested (Figure 5A and 5B).
To investigate if the Sc mtTyrRS CTD is a non-specific RNA-binding domain like CYT-18's CTD, we expressed and purified this domain separately including a small segment of the upstream linker region (denoted Sc CTD). We then assayed equilibrium binding of the Sc CTD to three group I introns (Nc mt LSU, Nc ND1m and Twort), a group II intron (Ll.LtrB), and poly(U)30 at 25°C. These assays showed that the Sc CTD is capable of binding all the RNAs tested with Kd or K1/2 values ranging from 110 nM to 1 µM (Figure 5C). The Sc CTD had the highest affinity for the Ll.LtrB group II intron RNA (K1/2 = 110 nM), which it bound in a cooperative manner (n = 1.8), and the lowest affinity for poly(U)30 (K1/2 = 1 µM; n = 1.6). Notably, for all RNAs tested, the non-specific binding by the isolated Sc mtTyrRS CTD is 1.5- to 15-fold weaker than the binding of CYT-18's CTD to the same RNA (K1/2s at 25°C = 57–64 nM) (Figure 4). Together, these findings indicate that the CTD of the Sc mtTyrRS is also a non-specific RNA-binding domain and that the Sc mtTyrRS binds both group I and group II intron RNAs non-specifically via its CTD.
Finally, we investigated whether the Sc CTD could replace the CYT-18 CTD to promote group I intron splicing by making CYT-18/Sc mtTyrRS chimeric proteins. Two chimeric constructs were made differing in whether they contain the flexible linker region from CYT-18 or the Sc mtTyrRS (Figure 6). Chimera 1 consists of the CYT-18 catalytic and α-helical domains fused to the Sc CTD via the Sc mtTyrRS linker region, whereas chimera 2 contains the same CYT-18 NTDs fused to the Sc CTD via the CYT-18 linker, which includes Ins 3. Both chimeric proteins showed tyrosyl-adenylation activity similar to wild-type CYT-18, but displayed differences in aminoacylation activity (Figure 7A and 7B). Chimera 1, which contains both the Sc mtTyrRS linker and CTD, is similar to the yeast Sc mtTyrRS in having higher aminoacylation activity with E. coli tRNATyr than does CYT-18, likely due to its better recognition of the bacterial tRNA (see above). By contrast, chimera 2, which contains the CYT-18 linker followed by the yeast mtTyrRS CTD, has substantially lower aminoacylation activity than CYT-18, indicating that the CYT-18 linker impairs charging of E. coli tRNATyr. The findings for chimera 1 indicate that higher TyrRS activity with E. coli tRNATyr correlates with presence of the yeast CTD and linker, regions of the TyrRS that recognize the tRNA variable arm and anticodon stem, and not with the catalytic domain, which recognizes the acceptor stem [26],[28],[52]. The Sc mtTyrRS linker may contribute to the recognition of E. coli tRNATyr, either by contacting the tRNA directly or by facilitating binding of the CTD to the variable arm and/or anticodon stem.
To determine whether the Sc mtTyrRS CTD could function in splicing, we compared the ability of the chimeric proteins to splice the Nc mt LSU and ND1m group I introns, which do and do not require the CTD for splicing, respectively [30]. Group I introns splice via two sequential transesterification reactions initiated by the addition of guanosine nucleotide to the 5′ end of the intron, resulting in ligated exons and excised linear intron RNA with a non-coded G residue at its 5′ end [53]. The ability of the chimeric proteins to splice the Nc mt LSU and ND1m group I introns was assayed by using 200 nM 32P-labeled precursor RNA containing the introns, 100 nM protein, and unlabeled GTP at three different temperatures (25°C, 30°C, and 37°C) at either 100 mM KCl (the standard condition for CYT-18) or a lower salt concentration, 25 mM KCl (Figures 7C–7F, S7A, and S7B). The assays showed that the chimeric proteins could splice the Nc ND1m intron, which does not require the CTD, but could not splice the Nc mt LSU intron, which requires the CTD, under all conditions examined. The inability of the chimeric proteins to splice the Nc mt LSU intron was additionally confirmed by splicing assays done with higher protein concentration (500 nM; protein excess conditions) to compensate for potentially weaker binding of the intron RNA by the Sc CTD (Figure S7C), and by using a more sensitive assay in which [α-32P]GTP is incubated with unlabeled precursor RNA to label the 5′ end of the intron RNA during the first step of splicing (Figures 7G and S7D).
Notably, although the chimeric proteins were capable of splicing the ND1m intron (Figures 7D and S7B), they did so at a slower rate than full-length CYT-18, with the rate decreasing further at higher salt conditions (Figure 7E and 7F). The slower rate of ND1m intron splicing by the chimeric proteins is similar to that found previously for the CYT-18 NTDs alone [30], suggesting that it reflects a lack of contributing but nonessential interactions with the CTD. Interestingly, chimera 2 has higher splicing activity with the ND1m intron than does chimera 1, the reverse of what was found for TyrRS activity (Figure 7B). This finding suggests that the longer CYT-18-linker region, which is present in chimera 2 but not chimera 1, contributes to splicing activity. This contribution could involve either specific or non-specific interaction of Ins 3 with the group I intron RNA or increased conformational flexibility of the CTD due to expansion of the linker. Considered together, the findings for chimera 1 and chimera 2 indicate that although both the Sc mtTyrRS and CYT-18 CTDs bind group I intron RNAs non-specifically, the Sc CTD lacks further adaptations required for group I intron splicing activity.
Our results provide insight into the function of CYT-18's CTD and its contribution to the evolution of group I intron splicing activity, highlighting a role for non-specific binding interactions in the evolution of new RNA-binding functions. First, the SAXS analysis indicates that the CTDs of both subunits of the CYT-18 homodimer have a preferred orientation in solution extending outward in opposite directions from the NTDs, but move inward to bind opposite ends of a group I intron RNA. The CORAL model of CYT-18* bound to Twort intron RNA based on the SAXS data suggests that the CTD of one subunit binds the intron near P2, P6–P6a, and P8, while the CTD of the other subunit likely interacts with P4–P5 and P9 (Figure 3C). These interaction sites agree with a previous biochemical model based on directed hydroxyl-radical cleavage assays in which Fe-EPD with a cleavage radius of 25 Å was conjugated at two sites in the CTD (G493C and C494) [37]. These assays found cleavage sites in P6–P6a, P3–P8, and P5 in the Nc ND1m intron and P2, P4, and P6–P6a in the Nc mt LSU intron [37],[54]. The additional cleavages in the Nc mt LSU intron P2 helix, which is considerably longer than P2 of the Twort or Nc ND1 introns, are consistent with its proximity to P8. The putative interaction sites between the CTDs and intron RNA in our CORAL model are also consistent with genetic experiments showing that CTD binding can suppress intron RNA mutations that impair long-range tertiary interactions P5-L9 and P2-L8 on opposite ends of the intron RNA [31].
The relatively fixed orientation of the CTDs in the free CYT-18 protein agrees with previous 15N-1H- two-dimensional NMR analysis showing that the CTDs of the full-length A. nidulans mt and Geobacillus stearothermophilus TyrRSs do not tumble independently in solution [37]. Nevertheless, the linker must be sufficiently flexible to allow the CTDs to bind to group I introns or tRNATyr on opposite sides of the catalytic domain, and the SAXS analysis provides the first direct evidence for this conformational flexibility by showing the two CTDs of the homodimer swing downward from their starting position in the free protein to interact with different regions of a group I intron RNA.
We were surprised to find that the CTDs of both CYT-18 and the non-splicing yeast mtTyrRS are non-specific RNA-binding domains. The isolated CTDs of both proteins bind structured group I and group II intron RNAs, or the simple homopolymer, poly(U)30 with similar affinities, with this non-specific binding 1.5- to 15-fold stronger for the CYT-18 CTD than the yeast mtTyrRS CTD (see Results). The non-specific RNA-binding activity of the TyrRS CTDs may contribute to its function in aminoacylation by augmenting its specific-binding interactions with tRNATyr, which include recognition of the variable arm and anticodon bases [28],[55]. Likewise, the high non-specific binding activity of the CYT-18 CTD does not preclude and may bolster specific-binding interactions of this domain with group I intron RNAs. The latter could result either from further adaptive evolution of the CTD or simply from positioning of the CTD on the intron RNA via specific binding of the NTDs.
Although non-specific RNA binding was unexpected for an aaRS domain involved in tRNA recognition, yeast and higher eukaryotic aaRSs have been shown previously to have appended non-specific RNA-binding domains that are not present in their bacterial counterparts and contribute to aminoacylation efficiency. Thus, the yeast glutaminyl-tRNA synthetase (GlnRS) has an N-terminal non-specific RNA binding domain, which when fused to a bacterial GlnRS enabled it to functionally replace the yeast enzyme in vivo, as did fusion of the yeast Arc1 protein, a non-specific RNA-binding protein that ordinarily helps mediate tRNA/aaRS interactions in trans [56],[57]. Similarly, some higher eukaryotic aaRSs have tandem repeats of a small non-specific RNA-binding motif that enhances tRNA binding [58]. These non-specific RNA-binding domains are thought to act by adding sufficient binding energy to compensate for relatively weak specific binding interactions of aaRSs with tRNA substrates, similar to the augmentation of specific binding of tRNA and intron RNA substrates suggested above for the TyrRS CTD.
Notably, the ribosomal protein S4-like fold, which forms the core of bacterial and mitochondrial TyrRS CTDs, has been identified previously as an ancient RNA-binding domain. This domain is found in all three kingdoms of life in a variety of proteins that bind structurally different RNAs, including two families of pseudouridine synthetases, a family of predicted RNA methylases, an RNA-modification enzyme with both pseudouridine synthetase and cytidine deaminase activity, threonyl-tRNA synthetases, and a heat-shock protein [59]–[61]. The S4-like fold consists of two α-helices arranged as a helical hairpin packed against three or four β-sheets. Connecting two of the β-sheets is a characteristic L-shaped loop, which together with the two α-helices is termed the αL motif. This motif generally contains clusters of basic and polar residues that are capable of interacting with various nucleic acid substrates in the different S4-like fold containing proteins. In TyrRSs, the αL motif interacts in a region between the variable and anticodon arms [61],[62]. We suggest that the inherently high non-specific RNA-binding affinity of the S4-like fold was the key factor enabling it to evolve interactions with different RNA substrates in the course of evolution. Indeed, the fungal mtTyrRSs provide a dramatic example of a case in which the S4-like fold of a single enzyme may bolster specific-binding interactions with three different regions of two different RNA substrates, a mt tRNATyr and a group I intron RNA.
Although we suggest that the non-specific binding of the CTD played a key role in initial interaction with group I intron RNAs, the CTDs of present-day fungal mtTyrRS appear to have evolved specific interactions with group I intron RNAs. Thus chimeric proteins containing the CYT-18 NTDs linked to the yeast CTD can efficiently aminoacylate E. coli tRNATyr, as well as splice the Nc ND1 intron, which requires only the NTDs [30]. However, the chimeric proteins splice the Nc ND1m intron less efficiently than full-length CYT-18 at a rate expected for loss of contributing CTD interactions, and they are unable to splice the Nc mt LSU intron, which requires the CTD [30]. Additional adaptations of the CYT-18 CTD required to promote splicing may include RNA-binding contacts by Ins 3–5, which are found in the CTDs of splicing-competent Pezizomycotina mtTyrRS, but not in the Sc mtTyrRS [37]. Both the previous biochemical model of CYT-18*+Twort [37] and the new CORAL model based on the SAXS data (Figure 3C) place Ins 4 and 5 in position to bind group I intron RNAs.
Since the discovery of the splicing function of CYT-18 [17], there have been numerous additional examples of aaRSs that have acquired new functions unrelated to translation, in most cases via addition of non-catalytic domains [63],[64]. The acquisition of these new domains and functions is thought to reflect that aaRS are ancient essential enzymes whose presence early in evolution of the cell provided a robust scaffold for the addition of new structural elements [65]. In archael and eukaryotic TyrRSs, the N-terminal catalytic domain is followed by a different anticodon-binding domain, known as the C-W/Y domain, which is homologous to the anticodon-binding domain of TrpRSs [66]. Two additional structural elements were acquired during the evolution of higher eukaryotes and function in receptor-mediated signaling pathways associated with angiogenesis: the ELR motif in the catalytic domain and a C-terminal EMAP II-like domain, which has non-specific RNA-binding properties [63],[67]–[69]. The ELR motif is on the intron-binding side of the catalytic domain [70] and incorporated in the same α-helix as Ins1 in the fungal mtTyrRS, suggesting that this region may be a particularly robust location for insertion of new functional elements.
Finally, our results provide evidence that non-specific binding can play a key and perhaps widespread role in pre-adaptive interactions that lead to the evolution of new RNA-binding functions of proteins. For the group I intron splicing activity of fungal mtTyrRSs, our findings suggest a scenario outlined in Figure 8 in which an initial non-specific interaction between the CTD of an ancestral mtTyrRS and a group I intron RNA was fixed by an intron RNA mutation that made formation of active ribozyme structure dependent upon interaction with the protein. After the interaction was fixed, the mtTyrRS and group I intron were forced to co-evolve, with further adaptive mutations in the protein leading to specific binding of both the catalytic domain and CTD to the intron RNA. These specific-binding interactions extended the intron RNA-binding surface, both increasing the efficiency of splicing and permitting additional mutations in the intron RNA that made it more dependent upon the protein for structural stabilization. RNA-editing enzymes such as APOBEC1, which evolved from enzymes that acted on mononucleotide substrates, may be additional examples of constructive neutral evolution in which a relatively non-specific pre-adaptive interaction with an RNA substrate was fixed by a deleterious mutation, in this case one that could be corrected by RNA editing, and then elaborated by further adaptive mutations [71],[72]. Indeed, a similar evolutionary pathway may have been used more generally for other RNA-modification enzymes, including the ones mentioned above that contain an S4-like non-specific RNA-binding domain.
Beyond the initial pre-adaptive phase, the extensive structural data for the interaction of fungal mtTyrRSs with group I intron RNAs provide strong evidence for a ratchet-like process in which multiple adaptive mutations, including six different Peziomycotina-specific insertions, led to the evolution of an efficient splicing apparatus for group I introns. It is highly unlikely that the multiple adaptive mutations in the protein leading to an extensive group I intron-binding surface occurred in one step. The surprising finding that the structural adaptations of the mtTyrRS catalytic domain utilized a non-tRNA-binding surface could reflect that the tRNA-binding site in the catalytic domain could not be easily modified to function in group I intron splicing without inhibiting mtTyrRS activity, which is essential in an obligate aerobe. Additionally, the non-tRNA-binding side of the catalytic domain may have had a pre-existing auxiliary RNA-binding function, as found for some aaRSs [73],[74]. By contrast to the catalytic domain, the regions of the CTD needed for splicing activity overlap tRNA-binding regions requiring co-evolution with both the intron RNA and mt tRNATyr. Indeed, the unusually long variable arm of Pezizomycotina mt tRNATyrs (see Results) (Figure S6) may be an example of a feature that co-evolved with the CTD to allow it to better accommodate group I intron RNAs [37]. We also note that although the initial interaction of an ancestral fungal mtTyrRS likely involved a single group I intron RNA, perhaps the mt LSU intron, which is dependent upon the mtTyrRS for splicing in all Pezizomycotina fungi examined [23], the fungal mtTyrRSs ultimately evolved to function in splicing multiple group I introns by recognizing the conserved phosphodiester backbone structure of the catalytic core. This binding mode has the evolutionary advantages of enabling the fungal mtTyrRSs to coordinate the splicing of multiple group I introns as well as the ability to accommodate new group I introns that invade genomes as mobile genetic elements.
Recombinant plasmids used for protein expression in E. coli are derivatives of the phage T7 promoter-driven expression vectors pET3a, pET11a, or pET11d (EMD Millipore). pEX560, which expresses a wild-type CYT-18 protein (amino acids 33–669), contains the cyt-18 ORF (nucleotides 97–2,010) cloned downstream of the T7 promoter in pET3a [29]. pCYT18/ΔC-tail, which expresses CYT-18* (C-terminal truncation of the non-essential C-tail; amino acids 584–669), was derived from pEX560 by introducing three stop codons (TAATAGTAG) after Leu583 by site-directed mutagenesis (QuikChange; Agilent Technologies). pHISTEV602 expresses the CYT-18 NTDs (C-terminal truncation of both the CTD and C-tail; amino acids 424–669), with an N-terminal tobacco etch virus (TEV) protease-cleavable 6× His-tag. It was constructed by PCR of pEX560 using primers that amplify nucleotides 97–1,251 of the CYT-18 ORF and append NcoI and BamHI sites, and then cloning the resulting PCR product between the NcoI and BamHI sites of pET11d. pCYT18-CTD, which expresses the CYT-18 CTD (amino acids 448–583) with an N-terminal TEV-cleavable 6× HIS-tag, was constructed by PCR of pEX560 using primers that amplify nucleotides 1,342–1,749 of the CYT-18 ORF and append NdeI and BamHI sites, and then cloning the resulting PCR product between the NdeI and BamHI sites of pET11a. All CYT-18 expression constructs lack the mt targeting sequence (amino acids 1–32). Wild-type CYT-18 and CYT-18* have an extra N-terminal methionine, while CYT-18 NTDs and CTD have an extra N-terminal glycine resulting from TEV-protease cleavage of the N-terminal 6× His-tag.
pHISTEVScTyrRS, which expresses the full-length mature S. cerevisiae mtTyrRS with an N-terminal TEV-cleavable 6× HIS-tag, contains Sc mtTyrRS codons 38–492 (lacking the mt target sequence; amino acids 1–37) cloned between the Nco1 and BamHI sites of pET11d [23]. pHISTEVSc/ΔCTD expresses Sc mtTyrRS lacking the CTD (denoted Sc NTDs) and was derived from pHISTEVScTyrRS by using site-directed mutagenesis to add three stop codons (TAATAATAA) after Asp400. pMAL-ScCTD, which expresses the Sc mtTyrRS CTD (denoted Sc CTD), contains Sc CTD codons 414–492 cloned between the BamHI and HindIII sites of pMAL-c2t [75], a derivative of plasmid pMAL-c2x (New England Biolabs) that expresses the protein with an N-terminal maltose-binding protein tag followed by a TEV-protease site.
Chimeric proteins containing the N-terminal catalytic domain of CYT-18 and the CTD of the Sc mtTyrRS were made by overlap PCR. Chimera 1 contains the CYT-18 NTDs (amino acids 33–417) fused to the Sc mtTyrRS flexible linker and CTD (amino acids 397–492). Chimera 2 contains the CYT-18 NTDs and linker including Ins 3 (amino acids 33–451) fused to the Sc mtTyrRS CTD (amino acids 416–492). The chimeric protein ORFs were cloned between the BamHI and HindIII sites of pMAL-c2t (see above), enabling the expression of fusion proteins with an N-terminal TEV-protease cleavable maltose-binding protein tag.
Recombinant plasmids used for in vitro transcription contain group I or II introns cloned downstream of a phage T3 or T7 promoter. pBD5a contains the N. crassa mt large subunit rRNA-ΔORF (Nc mt LSU) intron cloned downstream of a T3 promoter in pBS(+) [19]. Transcription of pBD5a linearized with BanI yields a 503-nt RNA containing a 65-nt 5′ exon, the 388-nt mt LSU intron, and a 50-nt 3′-exon. pND1m contains the N. crassa NADH dehydrogenase subunit 1-ΔORF (Nc ND1m) intron cloned downstream of a T7 promoter in pUC18 [18]. Transcription of pND1m linearized with NdeI yields a 209-nt RNA containing a 6-nt 5′ exon, the 196-nt ND1 intron, and a 7-nt 3′ exon. pTWORT-P2 contains a ribozyme derivative of a group I intron of the Staphylococcus aureus bacteriophage Twort orf142 gene (intron nucleotides 9-250) cloned downstream of a T7 promoter in pUC19 [76]. Transcription of pTWORT-P2 linearized with EarI yields a 242-nt transcript of the Twort ribozyme. pSSltrBΔA contains a derivative of the L. lactis Ll.LtrB-ΔORF intron with a deletion of the branch-point nucleotide to prevent splicing during binding assays cloned downstream of a T7 promoter in pUC19 [77]. Transcription of a DNA template made by PCR of the pSSltrBΔA plasmid (forward primer 5′-ATGAATTCTAATACGACTCACTATAGGGTTATAATTATCCTTACACATCCATAAC and reverse primer 5′-CGCTGCAGAATTGATATCAAAAATGATATG) yields an 807-nt RNA containing a 28-nt 5′ exon, the 749-nt intron, and a 30-nt 3′ exon.
Proteins were expressed from the recombinant plasmids indicated above in E. coli HMS174(DE3) (CYT-18, CYT-18*, and CYT-18 NTDs); BL21(DE3) (CYT-18 CTD, chimera 1, and chimera 2); or Rosetta 2(DE3) (EMD Millipore) (Sc mtTyrRS, Sc NTDs, and Sc CTD). Overnight cultures of fresh transformants were inoculated into LB media, and the proteins expressed via auto-induction [78]. Cells expressing CYT-18 and CYT-18* were grown at 35°C overnight with shaking at 260 rpm. Cells expressing all other proteins were grown at 37°C for 4 h then shifted to 25°C overnight with shaking at 260 rpm.
Wild-type CYT-18 and CYT-18* were purified as described [22],[27]. Briefly, cells were lysed by incubation with lysozyme at 1 mg/ml for 30 min followed by polyethyleneimine precipitation to remove nucleic acids, and ammonium sulfate precipitation [27]. The ammonium sulfate pellet was dissolved in 500 mM KCl, 25 mM Tris-HCl (pH 7.5) and then dialyzed overnight in 25 mM KCl, 25 mM Tris-HCl (pH 7.5). The protein was purified from the dialysate by using a HiTrap SP XL cation exchange column (GE Healthcare Life Sciences), followed by a size-exclusion column (HiLoad 16/60 Superdex 200; GE Healthcare Life Sciences) [22].
The 6× HIS-tagged proteins CYT-18 NTDs, CYT-18 CTD, Sc mtTyrRS, and Sc NTDs were purified similarly, except that the ammonium sulfate pellet was dissolved in 500 mM KCl, 25 mM Tris-HCl (pH 7.5), and 30 mM imidazole, and the proteins were purified by nickel-affinity chromatography using a HisTrap HP column (GE Healthcare Life Sciences) [23], followed by TEV protease-cleavage of the 6× HIS-tag in dialysis buffer (500 mM KCl, 25 mM Tris-HCl [pH 7.5], 5 mM DTT) to remove imidazole. The proteins were then further purified by an additional round of nickel-affinity chromatography, followed by size-exclusion chromatography (HiLoad 16/60 Superdex 200; GE Healthcare Life Sciences).
The maltose-binding protein (MalE) fusions MalE-ScCTD, MalE-chimera 1, and MalE-chimera 2 were purified by polyethyleneimine precipitation of nucleic acids, as described above for CYT-18, and then loaded onto an amylose affinity column (New England Biolabs) in buffer containing 25 mM Tris-HCl (pH 7.5), 500 mM KCl, 1 mM DTT, 1 mM EDTA, and 10% glycerol followed by elution with 10 mM maltose in the same buffer. The proteins were further purified using a heparin-sepharose column (HiTrap heparin HP column; GE Healthcare Life Sciences) in 300 mM KCl, 25 mM Tris-HCl (pH 7.5), 1 mM DTT, and 1 mM EDTA and eluted with a salt gradient of 300 mM to 1.5 M KCl in the same buffer. The final purification step was size-exclusion chromatography (HiLoad 16/60 Superdex 200; GE Healthcare Life Sciences) in 25 mM Tris-HCl (pH 7.5), 200 mM KCl, and 10% glycerol.
Proteins used for SAXS were stored in buffer containing 100 mM KCl, 5 mM MgCl2, 10 mM Tris-HCl (pH 7.5), 5% glycerol at −80°C. Proteins used for biochemical assays were dialyzed into 100 mM KCl, 25 mM Tris-HCl (pH 7.5), and 50% glycerol and stored at −80°C. Protein yields ranged from 14 to 44 mg/l (monomer concentrations), and all proteins were >99% pure as judged by SDS-polyacrylamide gels stained with Coomassie blue. Protein concentrations were determined by measuring A280 under denaturing conditions (6 M guanidine hydrochloride). Concentrations of wild-type CYT-18, the Sc mtTyrRS, and C-terminal truncations of these proteins refer to the homodimer, while CTD concentrations refer to the monomer.
Intron-containing RNA substrates for SAXS and biochemical assays were transcribed from the linearized recombinant plasmids indicated above. The Twort intron for SAXS was synthesized by large-scale in vitro transcription reactions (10–30 ml) with T7 polymerase at 37°C in reaction buffer containing 40 mM Tris-HCl (pH 8.1), 1 mM spermidine, 10 mM DTT, 8 mM NTPs, and 15 mM MgCl2. Transcription reactions were incubated at 37°C for 8 h and terminated by adding 50 mM EDTA followed by extraction with phenol-chloroform-isoamyl alcohol (25∶24∶1; phenol-CIA). The RNA was then purified through a 5-ml HiTrap desalting column (GE Healthcare Life Sciences) and a size exclusion column (HiLoad 16/60 Superdex 200; GE Healthcare Life Sciences). T7 RNA polymerase for the large-scale transcriptions was expressed with an N-terminal 6× HIS-tag from pRC9 and purified as described [79].
32P-labeled Nc mt LSU, Nc ND1m, and Twort RNAs for equilibrium-binding assays were synthesized by using a MAXIscript transcription kit (Life Technologies), with the concentration of unlabeled UTP changed from that recommended in the manufacturer's protocol (0.5 mM) to 10 µM UTP to obtain higher specific activity transcripts (300 Ci/mmol). The Ll.LtrB group II intron was synthesized by using a mutant T7 polymerase that can read through a T7 polymerase transcription termination site within the intron [80] in reaction medium containing 40 mM Tris-HCl (pH 7.9), 6 mM MgCl2, 10 mM DTT, 2 mM spermidine, 1 mM GTP, 1 mM CTP, 1 mM ATP, 250 nM UTP, and 1 µM [α-32P]UTP (3,000 Ci mmol−1; Perkin Elmer). After transcription and DNase treatments (MAXIscript transcription kit; Life Technologies), transcripts were purified by extraction with phenol-CIA, followed by gel filtration through two consecutive 1-ml Sephadex G-50 columns (Sigma-Aldrich).
Intron-containing RNA substrates for splicing reactions were transcribed from linearized DNA template using a MEGAscript transcription kit (Life Technologies) with 1 µCi [α-32P]UTP (3,000 Ci mmol−1; PerkinElmer) added for standard 32P-labeled substrates and 3 µCi [α-32P]UTP (3,000 Ci mmol−1) added for higher specific activity subtrates (Figures 7C, S7A, and S7C). The Nc mt LSU intron substrate was synthesized by in vitro transcription of pBD5a (BanI digested) using a MEGAscript T3 kit, while the Nc ND1m intron substrate was synthesized by in vitro transcription of pND1m (NdeI digested) using a MEGAscript T7 kit. The intron RNAs were purified as described above.
The poly(U)30 oligonucleotide used for binding assays was synthesized and HPLC-purified by Integrated DNA Technologies. The oligonucleotide was dissolved in 10 mM HEPES (pH 7.5), 1 mM EDTA and stored at a concentration of 25 µM. For equilibrium-binding assays, 25 pmoles of the oligonucleotide was 5′-end labeled with [γ-32P]ATP (3,000 Ci mmol−1; PerkinElmer) using T4 kinase (New England Biolabs) and then purified by phenol-CIA extraction followed by desalting through a Sephadex G-25 column.
Proteins and RNAs used for SAXS analysis were prepared as described above. RNA-protein complexes were formed by mixing protein dimer and RNA at a 1∶1 molar ratio in 1.2 ml of 100 mM KCl, 5 mM MgCl2, 10 mM Tris-HCl (pH 7.5), and 5% glycerol. After incubation at room temperature for 15 min, RNP complexes were purified by size-exclusion chromatography (Hi-Load 16/60 Superdex 200 column; GE Healthcare Life Sciences) in the same buffer. RNP complexes and proteins for SAXS were concentrated by using Amicon Ultra-4 centrifugal filter units (EMD Millipore) and frozen for storage at −80°C. The size-exclusion chromatography column buffer was used as a solvent blank for SAXS.
SAXS data were collected on beamline 12-ID-C at the Advanced Photon Source (Argonne, Illinois). Each sample had 20 1-s exposures taken at a sample-to-detector distance of 2.0 m, covering a momentum transfer range of 0.007<q<0.35 Å−1. Samples were continuously passed through the beam using a flow-cell to minimize radiation damage. The 20 consecutive exposures were compared and showed no change in scattering intensity, indicating no radiation damage. Radially averaged scattering data were buffer subtracted and analyzed by using ATSAS [42] and IGOR-Pro (WaveMetrics). Scattering curves were displayed as the scattering intensity (I(q)) as a function of momentum transfer q = (4πsinθ)/λ, where λ is the wavelength of the incident X-ray beam and θ is half the angle between the incident and scattering radiation. SAXS data were obtained for least three different concentrations of each protein and checked for aggregation and interparticle interference by examination of the Guinier region [81]. Guinier plots (log(I(q)) versus log(q)) were checked for linearity in the Guinier region, a diagnostic of sample quality. For globular proteins, the Guinier approximation is valid for qRg<1.3. The q range used for SAXS analysis was 0.015<q<0.3 Å−1 for CYT-18 protein constructs and 0.02<q<0.3 Å−1 for CYT-18+Twort complexes. The I(0) (extrapolated forward scattering at zero angle) and Rg (radius of gyration) were evaluated using the Guinier approximation for scattering intensity (I(q)) according to the equation:I(0) and Rg were also computed from the scattering curve by using the indirect Fourier transform program AUTOGNOM, which additionally provides an estimate of the maximum particle dimension (Dmax) from the distance distribution function P(r) [42]. The Rg values determined by using the Guinier approximation were consistent with those determined by AUTOGNOM. Molecular weights were calculated by comparing the extrapolated forward scattering at zero angle, I(0), with that of a protein standard, bovine serum albumin (BSA), by using the equation:where MMp and MMst are the molecular weights of the protein sample and protein standard, respectively, cp and cst are their concentrations in g/l, and I(0)p and I(0)st are the forward scattering intensities of the protein and standard, respectively. Agreement with the calculated molecular weights of the samples indicates sample quality and monodispersity [81]. Experimental scattering curves were compared with theoretical scattering curves calculated by the program CRYSOL (for qmax<0.3) from the crystal structures of those macromolecules with known atomic structures (CYT-18 NTDs, CYT-18 NTDs+Twort, CYT-18 CTD homology model) [43].
Ab initio shape reconstructions were done by using DAMMIN (for qmax<8/Rg) and GASBOR, which use simulated annealing methods to build low resolution protein models from dummy atoms or residues, respectively [45],[46]. The program DAMMIN uses dummy atoms packed into a sphere with the beads determined to be either protein or solvent. The final DAMMIN model was obtained by using the DAMAVER program suite to align ten models from independent DAMMIN runs and produce an average model. The latter was further refined by using DAMMIN to produce the final model [82]. GASBOR represents the protein as a chain-like ensemble of dummy residues equal to the number of residues in the protein. The final GASBOR model was chosen as the one with the lowest NSD value after running DAMSEL to compare ten models from independent GASBOR runs [82]. No symmetry was specified for the building of CYT-18* or CYT-18 CTD ab initio models, while P2 symmetry was specified for the CYT-18 NTDs models based on prior knowledge from the CYT-18 NTDs crystal structure. DAMMIN and GASBOR produced similar models of CYT-18* with or without P2 symmetry enforced.
Rigid-body models of CYT-18* by itself and of CYT-18* and the CYT-18 NTDs bound to Twort RNA were built by using the program CORAL [48]. This program employs a simulated annealing method to place high resolution models of individual components in orientations that minimize the discrepancy between the calculated SAXS profile and the experimental SAXS data, with distances between the structured components constrained by randomized dummy residue linkers chosen from a generated library of non-clashing loop structures. To build models of CYT-18*, a homology model of the CYT-18 CTD (amino acids 448–583) was generated by I-TASSER [47], using the A. nidulans CTD NMR structure (PDB:2KTL) as a template for modeling [37]. The confidence (C-score) and TM-scores of the CYT-18 CTD homology model, which are indicators of model quality, are high at 0.98 and 0.85, respectively. This CYT-18 CTD model and available high-resolution crystal structures for CYT-18 NTDs+Twort RNA (PDB:2RKJ) were used for rigid-body modeling. The final CORAL models were chosen from among ten independently derived models based on the best fit to the experimental scattering data as indicated by a low χ value [48].
Ensemble optimization analysis to characterize the flexibility of CYT-18*+Twort system was conducted by using the program, EOM [83],[84]. This program generates a random pool of 10,000 structures and creates an optimized ensemble from this pool, such that the average scattering pattern of the ensemble fits the experimental SAXS data. Comparison of the shape of the Rg and Dmax distributions of the optimized ensemble with those of the random pool provides information about the size and flexibility of the structure, with a broad peak resembling that of the random pool suggesting a flexible, extended structure and a peak narrower than the random pool suggesting a more rigid structure.
Tyrosyl-adenylation assays were done by incubating 100 nM protein in a 50-µl reaction containing 5 mM ATP, 100 mM KCl, 10 mM MgCl2, 144 mM Tris-HCl (pH 7.5), 2 mM DTT, 0.1 mg/ml BSA (New England Biolabs), 0.1 unit of yeast inorganic phosphatase (New England Biolabs), and 5 µCi of L-[3,5-3H]-tyrosine (53 Ci mmol−1; Amersham Biosciences Corp.) [30]. Reactions were initiated by adding protein and incubated at 30°C for 10 min. Reactions were terminated by adding 1 ml of reaction medium and immediately filtering through a nitrocellulose membrane to trap protein bound tyrosyl-adenylate. Radioactivity was measured by Beckman Coulter LS 6500 scintillation counter using Ready Protein scintillation cocktail (Beckman).
Aminoacylation assays were done as described previously with protein concentrations normalized to tyrosyl-adenylation activity [23]. Reactions of 120 µl contained 100 nM protein and 6 µM E. coli tRNATyr (Sigma-Aldrich) in 100 mM KCl, 15 mM MgCl2, 50 mM Tris-HCl (pH 7.5), 5 mM ATP, and 10 mM L-tyrosine (a 1∶10 mixture of L-[3,5-3H]-tyrosine and unlabeled L-tyrosine). Reactions were initiated by adding protein and incubated at 30°C. For time courses, 20-µl portions were removed after times ranging from 2 to 60 min, and the reaction was terminated by precipitation with 0.8 ml of a solution containing 10% trichloroacetic acid and 20 mM sodium pyrophosphate. Reactions were filtered through Whatman 3 MM filter paper to collect the precipitates, and the filters were washed three times with 1 ml of a solution containing 5% trichloroacetic acid and 20 mM sodium pyrophosphate followed by 2 ml of 95% ethanol. The filters were then dried and quantified by using a Beckman Coulter LS 6500 scintillation counter as above.
32P-labeled RNAs (5 pM; 300 Ci/mmol) were incubated with increasing concentrations of protein in a 50-µl reaction containing 100 mM KCl, 5 mM MgCl2, 20 mM Tris-HCl (pH 7.5), 5 mM DTT, 0.1 mg/ml BSA, and 10% glycerol at either 25°C (Figures 4 and 5) or 37°C (Figure S4). Binding reactions were initiated by adding 10 µl protein and terminated after 30 min by filtering 10 µl of the reaction through a nitrocellulose membrane (Amersham Hybond ECL nitrocellulose; GE Healthcare Life Sciences) backed by a nylon membrane (Amersham Hybond-N+; GE Healthcare Life Sciences). The nitrocellulose membrane retains protein-bound RNA and the nylon membrane retains free RNA. The end point (30 min) was chosen after determining that incubations times of 20, 30, or 60 min gave indistinguishable results for all proteins assayed. After application of samples, the membranes were washed three times with 20-µl wash buffer containing 100 mM KCl, 5 mM MgCl2, and 20 mM Tris-HCl (pH 7.5), then dried and quantified using a PhosphorImager and the program ImageQuant (GE Healthcare Life Sciences).
Splicing time courses for the Nc mt LSU and Nc ND1m introns were done by pre-incubating 32P-labeled precursor intron RNA (50 or 200 nM; 0.13–0.4 Ci mmol−1) with protein (25 or 100 nM) in a 100-µl reaction containing 100 mM KCl, 5 mM MgCl2, 20 mM Tris-HCl (pH 7.5), 1 mM DTT, 0.1 mg/ml BSA, and 10% glycerol for 10 min on ice, followed by 5 min at reaction temperature. Reactions were initiated by adding 1 mM GTP-Mg2+. Portions (8 µl) were removed at different times, and the reaction terminated by adding 50 mM EDTA, followed by phenol-CIA extraction and mixing 10 µl of sample with 10 µl of 2× gel loading dye (95% formamide, 0.02% SDS, 0.02% bromophenol blue, 0.01% xylene cyanol, and 1 mM EDTA). End-point splicing assays were done similarly in reaction medium containing 25 or 100 mM KCl for 60 min. Splicing assays comparing wild-type CYT-18 and chimera CYT-18/Sc mtTyrRS proteins were also done with higher concentrations of 32P-labeled precursor RNA (200 nM, 0.13–0.4 Ci mmol−1) and protein (100 nM or 500 nM dimer) and with 200 nM unlabeled precursor, 500 nM protein, and 500 nM [α-32P]GTP (3,000 Ci mmol−1; PerkinElmer). In all cases, samples were analyzed by electrophoresis in a denaturing 4% polyacrylamide gel, which was dried and quantified with a PhosphorImager, and data were analyzed by using ImageQuant TL.
|
10.1371/journal.ppat.1007763 | On the evolutionary ecology of multidrug resistance in bacteria | Resistance against different antibiotics appears on the same bacterial strains more often than expected by chance, leading to high frequencies of multidrug resistance. There are multiple explanations for this observation, but these tend to be specific to subsets of antibiotics and/or bacterial species, whereas the trend is pervasive. Here, we consider the question in terms of strain ecology: explaining why resistance to different antibiotics is often seen on the same strain requires an understanding of the competition between strains with different resistance profiles. This work builds on models originally proposed to explain another aspect of strain competition: the stable coexistence of antibiotic sensitivity and resistance observed in a number of bacterial species. We first identify a partial structural similarity in these models: either strain or host population structure stratifies the pathogen population into evolutionarily independent sub-populations and introduces variation in the fitness effect of resistance between these sub-populations, thus creating niches for sensitivity and resistance. We then generalise this unified underlying model to multidrug resistance and show that models with this structure predict high levels of association between resistance to different drugs and high multidrug resistance frequencies. We test predictions from this model in six bacterial datasets and find them to be qualitatively consistent with observed trends. The higher than expected frequencies of multidrug resistance are often interpreted as evidence that these strains are out-competing strains with lower resistance multiplicity. Our work provides an alternative explanation that is compatible with long-term stability in resistance frequencies.
| Antibiotic resistance is a serious public health concern, yet the ecology and evolution of drug resistance are not fully understood. This impacts our ability to design effective interventions to combat resistance. From a public health point of view, multidrug resistance is particularly problematic because resistance to different antibiotics is often seen on the same bacterial strains, which leads to high frequencies of multidrug resistance and limits treatment options. This work seeks to explain this trend in terms of strain ecology and the competition between strains with different resistance profiles. Building on recent work exploring why resistant bacteria are not out-competing sensitive bacteria, we show that models originally proposed to explain this observation also predict high multidrug resistance frequencies. These models are therefore a unifying explanation for two pervasive trends in resistance dynamics. In terms of public health, the implication of our results is that new resistances are likeliest to be found on already multidrug resistant strains and that changing patterns of prescription may not be enough to combat multidrug resistance.
| Antibiotic resistance and, in particular, multidrug resistance (MDR) are public health threats. Multidrug resistant infections are associated with poorer clinical outcomes and higher cost of treatment than other infections [1, 2] and there is concern that the emergence of pan-resistant strains (pathogens resistant to all available antibiotics) will render some infections untreatable [3].
From the point of view of finding effective treatment options, multidrug resistance is particularly problematic because resistance to different antibiotics tends to be concentrated on the same strains: positive correlations between resistance to different drugs have been found in multiple species (including Streptococcus pneumoniae, Neisseria gonorrhoeae, Staphylococcus aureus, Escherichia coli, Klebsiella pneumoniae, Pseudomonas aeruginosa and Mycobacterium tuberculosis) [2]. In other words, the frequency of MDR strains is higher than we would expect from the frequencies of individual resistance determinants if these were distributed randomly in the population (‘MDR over-representation’).
Understanding the causes of this MDR over-representation is important for limiting the impact of resistance. A number of possible explanations have been suggested (Table 1) [2], but the extent to which these processes contribute to the trend remains uncertain. Many of the proposed mechanisms are specific to subsets of antibiotics and/or species. The pattern of MDR over-representation, on the other hand, is pervasive: correlations have been observed between resistance to antibiotics acting through different mechanisms, and between chromosomal and mobile genetic element (MGE) associated resistance determinants [2]. Explanations for MDR over-representation must therefore be either sufficiently general or sufficiently diverse to account for this pervasiveness.
In this paper, we approach the problem of explaining MDR over-representation in terms of strain ecology: explaining why resistance to different antibiotics is often seen on the same strain requires an understanding of the competition between strains with different resistance profiles. For models of such competition to be credible, they must capture observed trends in resistance dynamics whilst being ecologically plausible. Developing models that fulfil these criteria has not been trivial: sensitive and resistant strains compete for the same hosts and simple models of competition therefore predict that the fitter strain will out-compete the other (‘competitive exclusion’) [19]. However, this is rarely observed: resistance frequencies have remained intermediate over long time periods in a number of species. For example, sustained intermediate resistance frequencies are observed in Europe for various antibiotics and numerous species, including E. coli, S. aureus and S. pneumoniae (European Centre for Disease Prevention and Control Surveillance Atlas, available at https://atlas.ecdc.europa.eu). Stable coexistence is also observed in surveillance data from multiple other locations (Centre for Disease Dynamics, Economics and Policy, available at https://resistancemap.cddep.org/AntibioticResistance.php). For further review of evidence for stable coexistence, see references [19, 20].
Recent work has explored the role of i) host population structure [21–23], ii) pathogen strain structure [20, 21] and iii) within-host dynamics [24] in maintaining the coexistence of antibiotic sensitivity and resistance. In this paper, we identify a structural similarity in the first two categories of model. In these models, coexistence arises through a combination of two factors. First, the presence of groups within the host or pathogen population in which the evolutionary dynamics of resistance are approximately independent from the other groups. Second, the presence of variation in the benefit gained from resistance between these groups, so that antibiotic resistance is selected for in some groups while sensitivity is selected for in others. We show that if this variation is correlated for different antibiotics, models with this structure also predict high levels of association between resistance to different antibiotics: all resistance determinants will tend to be found where the fitness benefit gained from resistance is the greatest. The observed high frequency of multi-drug resistance is therefore in line with ecologically plausible models of coexistence, making these models a parsimonious explanation for both trends.
In this section, we discuss competitive exclusion and previously proposed coexistence mechanisms in the context of multidrug resistance. We identify a structural similarity in plausible models of coexistence [20–22] and show that, in a multidrug context, models with this structure predict MDR over-representation. The model we present captures the dynamics of a bacterial species which is mostly carried asymptomatically (e.g. E. coli, S. aureus or S. pneumoniae), so the probability of a host being exposed to antibiotics does not depend on whether the host is infected with the pathogen [13]. Key results, however, are also applicable when this is not the case (see Discussion).
In this section, we explore how introducing additional complexity to the simplified model affects our predictions about association between resistance determinants and nestedness.
In this paper, we approach the question of explaining observed patterns of association between resistance to different antibiotics (‘MDR over-representation’) in terms of understanding the competition between strains with different resistance profiles. We consider recent models of the coexistence of antibiotic sensitive and antibiotic resistant strains [20–23] in which coexistence is maintained by heterogeneity in the fitness effect of resistance, arising either from heterogeneity in the rate of antibiotic consumption and/or difference in duration of carriage. We present a generalised version of these types of models, in which competition between antibiotic sensitivity and resistance is simplified to a series of independent sub-models (strata). We show that this model structure also gives rise to MDR over-representation because resistance to all antibiotics will be selected for in the strata where the fitness benefit of resistance (‘resistance proneness’) is the highest. Therefore, our results suggest that two pervasive trends in resistance dynamics, the robust coexistence of antibiotic sensitive and resistant strains and the over-representation of multidrug resistance, can both be explained by heterogeneity in the fitness effect of resistance within the host or pathogen population.
We first present a simplified model for conceptual insights and then explore how additional complexity affects predicted trends. Under the strong assumption of identical antibiotic prescription patterns in all strata and no recombination, this model predicts complete linkage disequilibrium (D′ = 1) between resistance to all antibiotics. Relaxing these assumption decreases the magnitude of linkage disequilibrium, giving rise to values of D′ similar to those observed in multiple bacterial datasets. High D′ is maintained even at unrealistically high recombination rates. A lower correlation in antibiotic consumption profiles between strata leads to lower values of D′. However, the effect is gradual and the magnitude of the decrease depends on whether the strata also differ in clearance rate. Thus, even in context where patterns of prescription differ considerably between host groups, we would still expect a degree of association between resistance determinants when variation in duration of carriage contributes to variation in the fitness effect of resistance.
Although the model builds on work exploring the stable coexistence of antibiotic sensitivity and resistance and coexistence is robustly observed in multiple datasets, the prediction that variation in the fitness effect of resistance leads to MDR over-representation does not require coexistence to be stable. We would expect MDR over-representation in the presence of fitness variation, even when this variation is not enough to maintain stable coexistence: for all antibiotics, the increase of resistance frequencies towards fixation would occur most rapidly in the populations with the greatest selection pressure for resistance. Under these circumstances, fitness variation would give rise to transient MDR over-representation.
Our results show that when variation in the fitness effect of resistance is present and when this variation is at least partially correlated for different antibiotics, it will give rise to MDR over-representation. The extent to which this mechanism accounts for observed patterns of MDR over-representation therefore depends on the extent to which this type of fitness variation is present in pathogen populations.
It is not entirely straightforward to evaluate how common variation in the fitness effect of resistance is. Wide-spread coexistence of sensitivity and resistance is not direct evidence for the pervasiveness of fitness variation because coexistence may not always arise through this mechanism. Although the majority of mechanisms proposed to date [20–23] work through fitness variation, other mechanisms are also possible [19]. In particular, recent modelling suggests that co-infection with sensitive and resistant strains gives rise to frequency-dependent selection for resistance and thus promotes coexistence [24]. However, the magnitude of this effect depends on the nature of within-host competition [24], for which there is limited data. Thus while theoretically plausible, the extent to which this mechanism contributes in practice is still unclear. It is worth noting that different coexistence mechanisms are not mutually exclusive. If coexistence arises through a combination of fitness variation and other mechanisms, we would a priori still expect the fitness variation to give rise to MDR over-representation.
In the work presented here, we consider fitness variation arising from heterogeneity in antibiotic consumption between host groups (hospitals vs communities, geographic regions, age classes) and from heterogeneity in duration of carriage between host groups (age classes) and between strains (pneumococcal serotypes). This is not an exhaustive list of possible sources of heterogeneity. For example, serotype does not fully account for heritable variation in pneumococcal duration of carriage [28], suggesting other genetic traits also play a role in determining carriage duration. In light of recent results suggesting wide-spread negative frequency-dependent selection in bacterial genomes [30, 31], it is not implausible to suggest these duration of carriage loci may also be under frequency-dependent selection. If so, diversity at these loci would create another source of variation in the fitness effect of resistance and hence promote coexistence and MDR over-representation. More broadly, variation in the fitness effect of resistance may arise through different mechanisms for pathogens with a different ecology than modelled in this work. For example, we have modelled a pathogen that is mostly carried asymptomatically and therefore exposed primarily to antibiotics prescribed against other infections. For pathogens where antibiotics prescribed due to infection with the pathogen itself contribute to a significant proportion of antibiotic exposure, the presence of strains differing in invasiveness would give rise to between-strain variation in antibiotic exposure and heterogeneity in the fitness effect of resistance. For bacterial species able to multiply both in hosts and in the environment, the sort of structure and heterogeneity considered in this work may also arise from differences between environmental niches.
This study does not fully address the role antibiotic prescription patterns in MDR over-representation: we highlight two important remaining questions. Firstly, in the modelling framework used in this study, the distribution of drug consumption within a stratum (i.e. a well-mixed population) does not have an impact on MDR over-representation (S1 Text Section 5). In other words, the presence of host groups consuming antibiotics at different rates only promotes MDR over-representation if there is very little transmission between these host groups: individual-level correlation in antibiotic exposure is not predicted to promote multi-drug resistance. We have not explored this result in detail—it may arise because the model predicts competitive exclusion within a stratum. Secondly, in contrast to the distribution of antibiotic consumption within a stratum, our results suggest that the distribution of antibiotic consumption between strata does matter: the prediction of MDR over-representation is sensitive to how correlated prescription profiles are and the extent of this sensitivity depends on whether variation in duration of carriage is also present. Relating these theoretical results to observed correlations in the antibiotic consumption between different host groups and to the extent of assortative mixing between these groups will provide additional insights into observed patterns of MDR (e.g. why the association between some drugs is higher than others).
The fitness variation model playing a role in MDR over-representation does not preclude a potential role for other mechanisms in contributing to the trend (Table 1). This study does not address the relative extent to which the different possible mechanisms contribute to MDR over-representation. This is for two reasons. Firstly, it is unclear what the patterns of MDR predicted by alternative mechanisms of MDR over-representation are. Secondly, we do not have a full understanding of which host and pathogen characteristics are relevant in defining the strata so it is difficult to directly address whether these traits are predictors of MDR. One alternative strategy for establishing the extent to which the fitness variation model contributes to MDR over-representation would be to assess patterns of association between resistance determinants in a single strain circulating in a well-mixed host population (i.e. a single stratum). The fitness variation model predicts no MDR over-representation (as defined by D′ > 0) under these circumstances. Therefore, if linkage disequilibrium is observed under these conditions, this would indicate that fitness variation is not the only mechanism of MDR over-representation. Furthermore, the magnitude of linkage disequilibrium could inform the relative contribution of the fitness variation mechanism: observing similar levels of linkage disequilibrium within strata and within the whole population would suggest the fitness variation is not a necessary mechanism for generating MDR over-representation.
From a public health perspective, the fitness variation model makes two concerning predictions. Firstly, we predict frequencies of pan-resistance will be high: in a perfectly nested set of resistance profiles, the frequency of pan-resistance is equal to the frequency of the rarest resistance. As a consequence, we would expect resistance arising in response to adoption of new antibiotics or increased usage of existing antibiotics to appear on already multidrug resistant lineages—an observation which has been made for the emergence of ciprofloxacin resistance in N. gonorrhoeae in the United States [32].
Secondly, our analysis has implications for the effectiveness of potential interventions against MDR. The variation in the fitness effect of resistance to different antibiotics need not be perfectly correlated for it to promote MDR over-representation. If the variation in fitness effect is maintained by multiple factors (e.g. differential antibiotic consumption between populations and variation in clearance rates), removing one of these factors (e.g. changing patterns of prescription so that consumption of different antibiotics is no longer correlated between host groups) may have limited impact on MDR over-representation.
The fitness variation model provides an explanation for MDR over-representation that is consistent with long term stability in resistance frequencies. This is relevant when considering temporal trends in resistance frequencies and predicting the future burden of resistance: other explanations for MDR over-representation (e.g. cost epistasis, correlated antibiotic exposure at the individual level—see Table 1) often require MDR strains to have an overall fitness advantage over strains with lower resistance multiplicity. This would imply that the higher than expected frequency of MDR is evidence for MDR strains out-competing other strains and thus suggest that MDR strains will eventually take over. Conversely, in the model we present, MDR strains are not out-competing other strains: all resistance frequencies are at equilibrium and MDR over-representation arises from the distribution of resistance determinants. It is worth noting, however, that even in the context of the fitness variation model, on a very long time-scale, we might expect the frequency of resistance to rise if bacteria are able to evolve resistance mechanisms that carry a lower fitness cost.
We show that previously proposed models in which coexistence of antibiotic sensitivity and resistance is maintained by heterogeneity in the fitness effect of resistance also predict high frequencies of multidrug resistance. The pervasive trends of coexistence and MDR over-representation can therefore be considered, at least partially, facets of the same phenomenon. We do not propose that the model we present fully explains observed patterns of association between resistance determinants. However, this effect should be considered when evaluating the role of antibiotic-specific MDR promoting mechanisms. From a public health point of view, the model we present is concerning because it predicts high frequencies of pan-resistance. On the other hand, heterogeneity in the fitness effect of resistance as an explanation for MDR over-representation allows reconciling this trend with long term stability in resistance frequencies.
The Maela pneumococcal dataset [33], collected from a refugee camp on the border of Thailand and Myanmar from 2007 to 2010, consisted of 2244 episodes of carriage, with associated antibiograms and carriage durations. Data were obtained from, and durations of carriage calculated by, Lees et al. [28] (S1 File). Data on antibiotic sensitivity was provided for ceftriaxone, chloramphenicol clindamycin, erythromycin, penicillin, co-trimoxazole (trimethoprim/sulfamethoxazole) and tetracycline. Ceftriaxone was excluded from the analysis because data was missing for a large proportion of isolates (44%). The Massachusetts pneumococcal dataset, collected as part of the SPARC (Streptococcus pneumoniae Antimicrobial Resistance in Children) project [34], was obtained from Croucher et al. (2013) [35] (data available from Croucher et al [35]). Croucher et al. reported minimum inhibitory concentrations (MICs) for penicillin, ceftriaxone, trimethprim, erithromycin, tetracycline and chloramphenicol. Tetracycline and chloramphenicol were excluded from the analysis because data was missing for a large proportion of isolates (47% and 67% respectively). Non-sensitivity was defined in accordance to pre-2008 Clinical and Laboratory Standards Institute breakpoints [36]. For both datasets, ‘resistance’ as used throughout the paper refers to non-sensitivity. The four hospital datasets were obtained from Chang et al. [2] (S2 File). All data were analysed anonymously.
If the frequency of resistance to antibiotic a is pa and the frequency of resistance to antibiotic b is pb, the coefficient of linkage disequilibrium between resistance to antibiotics a and b is Dab = pab − papb, where pab is the frequency of resistance to both a and b. The normalised coefficient D a b ′ is given by: D a b ′ = D a b m i n ( p a p b , ( 1 - p a ) ( 1 - p b ) ) if Dab < 0 and D a b ′ = D a b m i n ( p a ( 1 - p b ) , ( 1 - p a ) p b ) if Dab > 0.
In general the sign of D′ is arbitrary because it depends on which alleles are chosen for the calculation. We consistently calculate D′ using the frequency of resistance: positive D′ therefore means resistance to one antibiotic is associated with resistance to the other, while negative D′ means association between sensitivity and resistance.
All described models were implemented in Wolfram Mathematica (version 11.2.0.0). Modelling results are numerical solutions at t = 100000 (equilibrium is reached considerably earlier, see Fig H in S1 Text). For computing D′, numerical results for strain frequencies have been rounded to the nearest 10−10 to ensure strain frequencies for absent strains are zero (as opposed to zero within numerical error). The code is provided as a supporting file.
To test the effect of relaxing the assumption that the pathogen dynamics can be divided into non-interacting sub-models, we include three additional models.
First, we model the dynamics of resistance to three antibiotics (i.e. eight possible resistance profiles) spreading in a host population consisting of five host groups. The antibiotics make up different proportions of total antibiotic consumption (20, 35 and 45% of total antibiotic consumption rate τ). The pathogen experiences a different clearance rate within each host class p (μp). In addition, sub-strain with resistance profile g experiences clearance from antibiotic exposure at rate τg which depends on its resistance status: τg = τ(ia0.20 + ib0.35 + ic0.45), where ia = 1 if g is sensitive to antibiotic a and 0 otherwise. Resistance to each antibiotic decreases transmission rate by a factor of c. Uninfected hosts of class p (Up) are therefore infected at rate c n g β [ ( 1 - m ) I g , p + m 4 ∑ x ∈ P ′ I g , x ], where ng is the number of antibiotics strain g is resistant to, m is a parameter that sets the extent of mixing between the classes and P′ is the set of population classes excluding p. The dynamics of strain g within population p are thus described by:
d I g , p d t = c n g β [ ( 1 - m ) I g , p + m 4 ∑ x ∈ P ′ I g , x ] U p - ( τ g + μ p ) I g , p (10)
Second, we model the dynamics of resistance to three antibiotics in a single host population in pathogen with five strains differing in clearance rate (i.e. eight resistance profiles and five strains, giving a total of 40 possible sub-strains) with recombination at the duration of carriage locus. Strain i is cleared at rate μi and, as above, sub-strains with resistance profile g experience clearance from antibiotic exposure at rate τg which depends on its resistance status: τg = τ(ia0.20 + ib0.35 + ic0.45). Resistance to each antibiotic decreases transmission rate by a factor of c. Balancing selection is modelled similarly to Lehtinen et al. [20], by scaling transmission rate of strain i by a factor ψi which depends on the strain’s prevalence: ψ i = ( 1 - [ ∑ x I x , i 1 - U - 1 5 ] ) k, where k is a parameter setting the strength of balancing selection and U is the uninfected host class. Recombination at the duration of carriage locus is modelled by allowing hosts infected with strain i with resistance profile g to transmit strain j with resistance profile g at a rate r∑x Ix,j. Recombination therefore decreases the transmission of strain i with resistance profile g by ρg,i = rIg,i ∑x ∑y Ix,y and increases it by κg,i = r∑y ∑x Ig,y Ix,i. Note that the recombination rate parameter r captures the probability of co-infection, the probability of recombination occurring and the probability of transmitting the recombinant sub-strain. The dynamics of strain i with resistance profile g are described by:
d I g , i d t = c n g ψ i β [ I g , i - ρ g , i + κ g , i ] U - ( τ g + μ i ) I g , i (11)
The third model is the same as the one above, with the exception that recombination occurs at the resistance loci instead of the duration of carriage locus. It is therefore described by Eq (11), but the expressions for ρ and κ are different. We define resistance profile g a ′ as a resistance profile otherwise identical to g, but with the other allele at locus a (i.e. if g is sensitive to antibiotic a, g a ′ is resistant), Ng,a as the set of resistance profiles with the same allele at locus a as profile g and N g , a ′ as the set of resistance profiles with the different allele at locus a than profile g. Hosts infected with strain i with resistance profile g transmit a strain i with a resistance profile g a ′ at rate r ∑ j ∑ x ∈ N g , a ′ I x , j. Recombination can occur at any of the three resistance loci (we assume recombination rates are low enough to ignore the possibility of recombination occurring at multiple loci at the same time). Recombination therefore decreases the transmission of strain i with resistance profile g by ρg,i = 3rIg,i ∑x ∑y Ix,y and increases it by κ g , i = r ( I g a ′ , i ∑ y ∑ x ∈ N g , a I x , y + I g b ′ , i ∑ y ∑ x ∈ N g , b I x , y + I g c ′ , i ∑ y ∑ x ∈ N g , c I x , y ).
The parameter values for the results presented in Fig 3 are: c = 0.95, β = 2, {μ1, ‥, μ5} = {1.2, 1., 0.8, 0.6, 0.4}, τ = 0.12 and k = 5.
To test the effect of relaxing the assumption that all host groups consume different types of antibiotics in identical proportions, we model the dynamics of resistance to two antibiotics in a population consisting of ten host groups. The dynamics within each host group are represented by Eq 3, with parameter values cβ = 0.95 and cμ = 1 for both antibiotics, β = 2, and, unless otherwise stated μ = 1. There is no transmission between these host groups. For both drugs, five of these host groups consume the antibiotic at a rate which selects for resistance when μ = 1 (τhigh = 0.075), and five at a rate which selects for sensitivity when μ = 1 (τlow = 0.025). There are therefore six different ways in which the consumption rates of the two antibiotics can be combined: all populations consuming the first drug at a high rate also consume the second drug at high rate (Spearman’s rho: 1); four out of the five populations consuming the first drug at a high rate consume the second drug at a high rate (Spearman’s rho: 0.6); etc. We run a simulation for each of these six possible configurations. To test the effect of additional variation in the resistance proneness of strata, we introduce variation in the clearance rate of these populations: the five host groups consuming the first drug at the high rate now have different clearance rates (evenly spaced between a maximum and minimum clearance rate), and similarly for the five host groups consuming the first drug at the low rate. We run a simulation for each of these possible ways the consumption of the second drug can be distributed among these host groups.
|
10.1371/journal.pgen.1000119 | The Recombinational Anatomy of a Mouse Chromosome | Among mammals, genetic recombination occurs at highly delimited sites known as recombination hotspots. They are typically 1–2 kb long and vary as much as a 1,000-fold or more in recombination activity. Although much is known about the molecular details of the recombination process itself, the factors determining the location and relative activity of hotspots are poorly understood. To further our understanding, we have collected and mapped the locations of 5,472 crossover events along mouse Chromosome 1 arising in 6,028 meioses of male and female reciprocal F1 hybrids of C57BL/6J and CAST/EiJ mice. Crossovers were mapped to a minimum resolution of 225 kb, and those in the telomere-proximal 24.7 Mb were further mapped to resolve individual hotspots. Recombination rates were evolutionarily conserved on a regional scale, but not at the local level. There was a clear negative-exponential relationship between the relative activity and abundance of hotspot activity classes, such that a small number of the most active hotspots account for the majority of recombination. Females had 1.2× higher overall recombination than males did, although the sex ratio showed considerable regional variation. Locally, entirely sex-specific hotspots were rare. The initiation of recombination at the most active hotspot was regulated independently on the two parental chromatids, and analysis of reciprocal crosses indicated that parental imprinting has subtle effects on recombination rates. It appears that the regulation of mammalian recombination is a complex, dynamic process involving multiple factors reflecting species, sex, individual variation within species, and the properties of individual hotspots.
| In most eukaryotic organisms, recombination—the exchange of genetic information between homologous chromosomes—ensures the proper recognition and segregation of chromosomes during meiosis. Recombination events in mammals are not randomly positioned along the chromosomes but occur in preferential 1–2-kilobase sequences termed hotspots. Different species such as humans and mice do not share hotspots, although the same principles almost certainly regulate their placement in the genome. Hotspot positions and activities depend on genetic background and show sex-specific differences. In this study, we present a detailed analysis of recombination activity along the largest mouse chromosome, finding that recombination is regulated on multiple levels, including regional positioning relative to the chromosomal ends, local gene content, sex-specific mechanisms of hotspot recognition, and parental origin. Our results will contribute to further understanding of one of the most fundamental biological processes and are likely to cast light on several aspects of population genetics and evolutionary biology, as well as enhance our practical ability to define the genetic components of human disease.
| Genetic recombination is a fundamental process common to all eukaryotic organisms, which ensures proper chromosomal segregation of homologous chromosomes in meiosis and increases genetic diversity by creating new combinations of parental alleles at each generation. The process begins in the leptotene stage of meiosis I with the creation of double strand breaks on one chromatid by the topoisomerase-like protein Spo11, and is followed by resection of 5′-ends to leave 3′-overhangs which then displace existing strands on a non-sister chromatid. The resected regions are eventually repaired using the non-sister chromatid as a template, producing two types of recombination products: crossovers, and gene conversions without exchange of flanking markers (non-crossovers). According to the most widely accepted model of double-strand break processing [1], crossovers are predominantly produced by double-strand break repair (DSBR), and non-crossovers are predominantly produced by synthesis-dependent strand annealing (SDSA) [2].
In mammals, higher plants and yeast, recombination initiates prior to synapsis, and is required for successful chromosome pairing in meiosis I. The majority of recombination is localized to very limited intervals along the genome, termed hotspots, which in mammals are typically only 1-2 kilobase pairs (kb) long [3] and are surrounded by much longer regions (tens of kilobases or more) lacking recombination. When crossover rates are measured at individual hotspots on sperm samples [4], their activities vary over several orders of magnitude, from as high as 1–2 centimorgans (cM) [5] to below 0.001 cM. In contrast, hotspots are not believed to be present in organisms such as Drosophila and C. elegans where synapsis precedes recombination [6],[7], although local variation of recombination rates across large genomic regions exists in these organisms [8],[9].
Despite their apparent abundance, less than two dozen recombination hotspots have been experimentally analyzed [10]–[13] in humans and mice. The most intensely mapped mammalian regions are the H2 region of mouse Chromosome 17 [5],[14], the human HLA region of Chromosome 6 [3], and the Ath1 region of mouse Chr 1 [11]. The evidence emerging from these studies suggests that mammalian hotspots are not uniformly or even randomly located along chromosomes. They can occur in “torrid zones” of very high recombination, with clusters of hotspots within 100 kb [11], leaving long stretches of DNA (as much as a megabase or more) devoid of recombination.
Recombination positioning and activity differ significantly between the sexes, and their recombination maps can have different lengths in many species. The female map is about 1.7 times longer than the male map in humans [15],[16] and about 1.3 times in mice [17], and high-resolution sex-specific linkage maps in humans [18] and mice [17] show dramatic variation between male and female recombination rates along the chromosomes. Several explanations have been proposed for these sex differences, including haploid selection[19], different epistatic interactions for genes expressed paternally or maternally[20], and regional differences in the chromatin structure of male and female gametocytes [21]. Our own work has shown that a difference in crossover interference distances in Mb, related to the physical length of synaptonemal complexes at the pachytene stage of meiosis I [22], is a major factor underlying broad-scale sex differences in recombination rates. Sex specificity has also been detected at the level of individual hotspots [23], resulting from participation of both cis- and trans-acting factors [5].
In several species, including maize [24], humans [25]–[27] and mice [5],[23], genetic background can dramatically influence the placement and activity of hotspots. Humans and chimpanzees do not share hotspots, although their sequences are 98.6% identical [28],[29]. Differences in recombination activity between individual human males were detected even when the hotspot and its surrounding sequences were identical [27]. And, in perhaps the most extensive such study, the activity of the Psmb9 hotspot in mice is dependent on flanking sequences, even though the hotspot sequence itself is identical in both active and inactive haplotypes [5],[23].
Collectively, these findings emphasize the utility of defining the recombination landscape resulting from hotspots acting in a genetically defined background, a task that is impossible in humans but entirely feasible in experimental animals. Creating such high-resolution genetic maps is important for both theoretical and practical reasons. Studying one-generation recombination in a genetically defined system will provide an entrée to understanding how the recombination process is regulated, the mechanisms underlying sex specificity, and the role of hotspots in evolutionary processes. Better fine-scale genetic maps will also help optimize strategies for mapping and identifying genes underlying disease that rely on genome-wide association studies in humans and the analysis of quantitative trait loci in laboratory animals.
Several genome-wide mapping efforts in mice [30]–[32] have achieved near centimorgan resolution, the latest and most comprehensive one reaching an average resolution of 0.37 cM or 550 kb [17]. The goal of this study is to present the first detailed analysis of recombination on an entire chromosome of an experimental animal under genetically defined circumstances at a resolution power reaching <5 kb that enables detection of individual hotspots.
We studied sex-specific recombination rates along the entirety of mouse Chr 1 as they occurred in the meioses of C57BL/6J (B6) and CAST/EiJ (CAST) F1 hybrids of both sexes at an average resolution of 225 kb, and further refined the extended subtelomeric region of 24.7 Mb. To test for potential effects parental imprinting might have on recombination, the F1 animals were produced by reciprocal crosses, and then backcrossed to C57BL/6J. Mapping the location of crossovers in these backcross progeny provided information on the recombination events arising in the F1 hybrids. A total of 6028 progeny were genotyped, of which 1465 were offspring of female B6xCAST, 1537 of female CASTxB6, 1479 of male B6xCAST, and 1547 of male CASTxB6. In all, we detected and localized 5472 crossover events on Chr 1, reaching a genetic resolution of 0.017 cM in the combined offspring. The frequency with which chromosomes with different numbers of crossovers were observed is summarized in Table 1. We found significantly more multiple crossovers in female compared to male meiosis (p<10−13 by χ2 test) as described before [22].
Backcross offspring were genotyped in two consecutive rounds with single nucleotide polymorphism (SNP) assays developed using the Amplifluor system (see Materials and Methods). In the first round, all progeny DNAs were mapped over the entire chromosome at 10-Mb resolution. This was sufficient to detect virtually all crossovers, given the strong interference in mouse meiosis [33]. In the second round, the crossovers occurring in each interval were mapped using additional SNP markers to an average physical resolution of 225 Kb. To provide a sample of even more detailed information, recombinants in the subtelomeric 24.7 Mb were subjected to additional rounds of testing using a combination of SNP and simple sequence length polymorphism (SSLP) markers. Among the crossovers occurring in this region, 81.4% were mapped to under 100 kb resolution: 8.2% at 50–100 kb resolution, 33.5% at 20–50 kb resolution, 8.6% to a nearly hotspot resolution of 5–20 kb and 31.1% were mapped to <5 kb, ensuring hotspot level resolution. All markers used in this study, their positions according to NCBI Build 36, physical resolution and the number of crossovers in each interval are included in Table S1. Individual crossovers in five of the newly identified hotspots (shown in Table S1) were sequenced to determine exact locations of the chromatid exchange points within the limits of resolution provided by the locations of internal SNPs.
In total, the sex-averaged genetic map length of Chr 1 in the B6xCAST cross was 90.9 cM, which represents an average rate of 0.469 cM/Mb across 193.8 Mb, excluding the centromere adjacent 3 Mb for which no sequence information is available according to NCBI sequence build 36.
At 225 kb resolution, recombination activity was distributed very unevenly along the chromosome, forming alternating domains of higher and lower activity (Figure 1A). Recombination activity was found in only 64% of all intervals along the chromosome, the remaining 36% being completely devoid of recombination. In several places along the chromosome, recombination activity tended to be clustered in runs of consecutive intervals all of which were active, forming “torrid zones”. The most concentrated of them were 1.4–6.1 Mb long and were located at 37–41 Mb, 51–52.4 Mb, 72–74.8 Mb, 81.6–83 Mb, 131.4–132.8 Mb, and 189.5–195.6 Mb (red boxes in Figure 1A).
Correspondingly, intervals devoid of recombination activity tended to cluster in “cold zones”, the largest of which was over 6 Mb long. These were most prominent around 44.6–46.8 Mb, 48.6–51 Mb, 84.8–88.0 Mb, 96–97.8 Mb, 102.6–105.6 Mb, 110–116 Mb, 119–121.6 Mb, 149.2–151.4 Mb, 158.6–160.2 Mb (blue boxes in Figure 1A).
We did not detect any significant correlation along the chromosome between the locations of torrid and cold zones and traditional cytological banding patterns (Figure 1B).
To test the extent to which the recombination properties of a chromosome are evolutionarily conserved, we compared our results, obtained in a cross of only two strains, with the recombination map of Shifman et. al. [17]. The Shifman map was prepared at an average 550 kb resolution using the progeny of heterogeneous stock (HS) mice which merge the genetic backgrounds of eight mouse strains, including C57BL/6J but not CAST/EiJ. The two crosses have similar regional distribution of recombination along the chromosome, but do not share a substantial fraction of hotspots, if any.
Regional conservation between the two crosses was indicated by the significant correlation of recombination rates along the chromosome when tested at long intervals (r = 0.87 at 8.75 Mb resolution, Pearson correlation). However, this correlation decreased markedly when smaller intervals (4.4 Mb, 2.2 Mb, 1.1 Mb and at the maximum resolution of 0.55 Mb) were compared (Figure 1C). At the half-megabase scale, we found only a weak regional correlation (r = 0.38).
These estimated correlations are somewhat attenuated by the sampling variation in the estimates of recombination rates, and this attenuation increases at higher resolution, since the sampling variation is greater at higher resolution (due to smaller numbers of observed recombination events in smaller intervals). But for the sample sizes in these studies, the attenuation in the estimated correlations is negligible (on the order of 1/1000), and so cannot account for the large observed decrease in correlation from the 8.75 Mb scale to the 0.55 Mb scale.
Long regions of very low or no recombination were evident in both crosses and provided the strongest parallels between the crosses. These regions include those around 43–50 Mb, 96–106 Mb, 111–116 Mb and several smaller regions between 141–152 Mb. The lack of recombination in these regions cannot be attributed to inversions, which would prevent the survival of recombinants. Two main reasons speak against this possibility. First, some parents in the mixed genetic background will inevitably have the same orientation of the region in question if it were inverted in some of the eight strains, and therefore recombination would be detected in their progeny. Second, some intervals in these regions are not totally devoid of recombination in both crosses but have very low rates.
In addition to local variation in recombination rates, genetic background also plays a role in determining overall recombination rates. The genetic map length of Chr 1 was ∼31% higher in HS mice than in our two-strain cross. The reasons for this significant difference are uncertain. The lack of local correlation indicates that this difference is not simply due to an increased use of the same hotspots in HS mice. The present genetic data [22] agree with counts of the average number of chiasmata per meiosis during spermatogenesis among inbred strains [34] and counts of MLH1 foci marking sites of crossing over on Chr 1 [35]. It might be possible that recombination in a very heterogeneous genetic background is quite different from that seen in crosses of inbred strains. The importance of genetic background in recombination is also suggested by substantial differences between the crosses' recombination rates at specific intervals. For example, in the 24.7 Mb region that was mapped at considerably greater resolution (see below), recombinational activity was often present in one mouse cross (B6xCAST or HS) but not the other.
We found an overall positive correlation between gene density and recombination along the entire chromosome over megabase distances (r = 0.557 at 10 Mb). However, this effect diminished over shorter distances (r = 0.164 at 500 kb) (Table 2). At 200 kb, the correlation was low (r = 0.079) but statistically significant. Moreover, this positive correlation was not uniform along the chromosome but was restricted to only some regions, and statistically significant only for the region between 100–150 Mb (maximum correlation r = 0.877 at 5 Mb for the sex-average data). In this region, the positive correlation was still detected, and statistically significant, at 200 kb (r = 0.278). For the first and second 50-Mb segment (3–50 and 50–100 Mb), the correlation was positive but not statistically significant, whereas the correlation for the last region (150–194 Mb) was slightly negative up to 2Mb but not statistically significant. The 24.7-Mb part of the last segment was mapped to higher resolution (see below) and showed slightly negative correlation between gene density and recombination at 200 kb which disappeared at 50 kb.
Recombination tended to avoid gene deserts larger than 1.5 Mb but showed a tendency of clustering at their borders. The average rate in large gene deserts totaling 59.77 Mb (shown in Figure S1) was 0.26 cM/Mb compared to 0.55 cM/Mb in the remaining 134.02 Mb of non-deserts (p<10−99 by χ2 test) and 0.467 cM/Mb over the entire chromosome. The average rate was 0.80 cM/Mb in the 0.5–0.7 Mb border regions surrounding large gene deserts (p<10−51) and rapidly decreased beyond that to become statistically indistinguishable from the average chromosome rate (p = 0.596).
Similar correlation was found over the entire chromosome between exon density and recombination (r = 0.566 at 10 Mb and r = 0.126 at 500 kb, Table S2) and transcription start sites and recombination (r = 0.585 at 10 Mb and r = 0.121 at 500 kb, Table S3). However, the correlation was not statistically significant at 200 kb (r = 0.043, p = 0.101 for exons and r = 0.026, p = 0.204 for transcription start sites). In these two comparisons, most of the positive correlation was statistically significant for the region between 100–150 Mb but not for the rest of the chromosome. In the 24.7-Mb region mapped to higher resolution, both exon density and transcription start sites were slightly negatively correlated with recombination down to 50 kb (r = −0.045 and r = −0.071, respectively) and this effect was statistically significant for transcription start sites (p = 0.021).
Two striking examples of torrid zones that occur in large introns provide evidence that recombination is not restricted to intergenic regions. The first one consists of at least six hotspots in the 218-kb long second intron of Pbx1 (pre B-cell leukemia transcription factor 1, located at 169.995–170.268 Mb, NCBI Build 36), which is also a hotspot for translocations associated with acute lymphoblastic leukemia in humans [36],[37]. The second torrid zone includes at least three hotspots in the 80-kb long third intron of Esrrg (Estrogen receptor-like receptor gamma, located at 189.309–189.915 Mb).
We observed a simple, negative exponential relationship between the crossover rate among intervals and the likelihood of seeing hotspots of that activity. Among intervals averaging 225 Kb in length, recombination rates (expressed as cM/Mb to correct for variations in interval length) varied continuously over almost three orders of magnitude, from 0.017 cM/Mb (the lower limit of detection in this cross) up to 10 cM/Mb. Intervals with differing recombination rate were not equally likely; instead, when they were placed in rank order of recombination activity, the rates were distributed in a simple exponential manner where Rn, the recombination rate in the nth ranked interval was equal to kecn, where k and c are constants (Figure 2A). Figure 2B, which is also an exponential function, describes the cumulative recombination rate among rank-ordered intervals. A similar exponential relationship for the cumulative recombination rate was reported by McVean et al [38] for the human genome.
These exponential relationships indicate that nearly 50% of all recombination activity occurred in only 7.6% of the intervals while 22.2% of the intervals accounted for 80% of all recombination activity. Similar findings that a high percentage of all recombination is concentrated in a small fraction of chromosome intervals have recently been reported for the human genome [39]. The interval fractions become even smaller with decrease in interval size (see below). This result, which suggests that the majority of all recombination events occur in a relatively small fraction of the chromosome, has important practical implications for genetic mapping strategies. The conclusion that follows is that a moderate size cross should be optimal for mapping genes and QTLs because adding more offspring will not substantially increase the resolution power. The result provides an experimental ground to something that mouse geneticists have known intuitively for some time-if a gene cannot be mapped with the first few hundred offspring, the best strategy is to move to another cross if that is at all possible.
High-resolution mapping further emphasizes the uneven distribution of recombination activities among intervals (Figure 3A and Figure S2).
The 24.7-Mb telomere-proximal segment between 168.8–193.5 Mb had a genetic length of 22.7 cM. This accounts for a relative recombination rate of 0.92 cM/Mb, which is about twice the average rate of the entire chromosome. When it was mapped further to an average resolution of 75 kb, the distribution of recombination activities among intervals remained continuously variable as in the 225 kb intervals. However, as expected from the punctate location of hotspots, a smaller fraction of the genome-52% compared to 64% at 225 kb resolution–contained all recombination. Indeed, 50 percent of all recombination occurred in 16 intervals spanning only 1.8% of the segment length, with each of these intervals having an activity of 0.34 cM or more.
Recombinations in eight of these sixteen most active intervals were mapped down to 20–45 kb resolution while those in the remaining eight intervals marked with red circles on Figure 3A were mapped down to ∼3 kb resolution. All but one of the eight intervals contained a single hotspot, which was separated from the closest adjacent hotspot by at least 30 kb of sequence. The notable exception was the presence of two hotspots only 5 kb apart in the third intron of the Esrrg gene (Figure 3B).
Distances between adjacent intervals with recombination rates of 0.34 cM or more varied over three orders of magnitude in genomic terms, ranging from 5 kb to 5 Mb (1.52 Mb on average). The variation was much smaller in genetic terms, from 0.37 to 2.44 cM, or an average of 1.26 cM.
As interval sizes become smaller, it becomes increasingly likely that an interval contains only one hotspot. This provides a means of estimating the total number of hotspots in this 24.7-Mb segment, and by extension the total number in the genome. For this, the number of intervals showing any recombination activity was plotted as a function of interval size and the resulting trend lines extrapolated to a 5kb interval size, the minimal distance we found between adjacent individual hotspots (Figure 3C, results summarized in Table S4). This yielded an estimate of one hotspot per 108 kb on average, or about 228 hotspots accounting for all recombination in this segment among 6028 meioses. As expected from the exponential relationship described above, more active hotspots occur less frequently. On average, those with rates higher than 0.1 cM are likely to occur once per 425 kb, and those with rates higher than 0.2 cM, about once per megabase. These results are obviously tempered by the fact that they were obtained for one genetic combination in a region of the genome whose recombination rate is higher than the genome wide average.
To the extent this region is representative of the rest of the genome, its hotspot density provides an estimate of the total number of hotspots in the entire mouse genome that are active in this B6xCAST cross. We have made this estimation by relating the genetic length of the 24.7-Mb region to the total genetic length of the mouse genome. We assume that genetic lengths (measured in cM) will be more relevant than physical lengths (measured in Mb) because of the uneven distribution of recombination along the chromosome and the existence of long regions devoid of recombination. This calculation, using the Dietrich et al [30] sex-average map length of 1361 cM for the same C57BL/6JxCAST/EiJ cross, results in an estimate of about 13,670 hotspots (228/22.7×1361) across the mouse genome.
A recent study [40] typing 8.23 million SNP markers detected about 40,000 haplotype blocks in 12 classical inbred mouse strains based on ancestry inferred from representative strains of the four main mouse subspecies. Although the haplotype block boundaries were not always well defined, to the extent that they represent bona fide historical sites of recombination, the scales of these two estimates are not far apart. Our study should be considered a minimum estimate as it measured recombination from contemporary hotspots in one generation of a cross involving only two inbred strains, and was limited by the sensitivity of detection of 6028 meioses. The estimate of Frazer et al [40] suggested a higher number of hotspots in the genome of classical mouse inbred strains because it is not limited to contemporary hotspots and reflects the behavior of historical hotspots generating recombination over many generations in a variety of genetic backgrounds.
The most recent estimate [41] using more than 3.1 million SNPs has identified 32,996 hotspots in the human population, which is in the range of these estimates for the mouse genome.
The two sexes differed at all levels of organization of recombination. Overall recombination rates were higher in females than males; recombination was distributed differently along the chromosome in males and females, and there were also sex-specific hotspots.
The female recombination map of Chr 1 was 99.5 cM, or 1.21 times longer than the male map which was 82.3 cM, with average recombination rates over the entire chromosome of 0.51 and 0.42 cM/Mb, respectively. These differences were statistically significant (p<10−6 by Fisher's exact test). Among 225 Kb intervals, there was an overall positive correlation between female and male rates (r = 0.64) along the chromosome. This correlation did not change significantly at larger interval sizes up to 8 Mb. The underlying reason why the correlation did not increase with interval size was the substantial variation in distribution of recombination along the chromosome (Figure 4A), which included differences in both the number and relative recombination activity of intervals.
Recombination activity was spread over a larger fraction of the chromosome in females than in males. In females, 57.1% of intervals were recombinationally active compared to only 42.2% in males (a ratio of 1.35). This differential was apparent at all activity levels; 80% of all activity occurred in 23.2% of female versus 13.6% of male intervals, and 50% occurred in 8.23% of female versus 4.65% of male intervals.
These sex differences in the relative rates of recombination were regionally controlled (Figure 4B). Female recombination rates were higher in the centromere-proximal 27 Mb and in the region between 79–178 Mb, whereas male recombination rates were higher in the telomere-proximal 178–197 Mb region and generally, but not in the entirety, of the region between 27–79 Mb.
To study regional effects in more detail, we examined the switch between higher female and higher male recombination found in the fine-mapped 24.7 Mb sub-telomeric region. Female recombination rates were generally higher than those in males in the region between 169–178 Mb, with an abrupt transition to the opposite case in the adjacent region between 178–194 Mb where males had higher recombination (Figure 5 and Figure S3). Interestingly, the switch occurs in a region of very low recombination in both sexes. Overall, the difference between the two sexes was highly significant over the entire region (p<10−4).
Although the sexes share a substantial fraction of hotspots, there are many considerable differences in activity. Commonality of hotspot usage was indicated by the observation that comparisons at multiple interval sizes did not change the correlation between the two sexes (r = 0.62). However, there were also specific sex differences in hotspot activity that were independent of regional control. Among the 28 intervals with sufficiently high recombination (>0.2cM) to provide sufficient numbers of crossovers for statistically significant analysis, 18 showed sex-specific differences after adjustment for multiple testing (Table 3). Among these 18, eleven showed at least some activity in both sexes, seven being markedly more active in females and four in males (p<0.01, q<0.1). Seven of the 18 were detected in only one sex, four in females and three in males. The latter group indicates that some hotspots may be truly sex specific, or at least that the differences in their activity are so great (>10 times) that recombination was not detected in the low-activity sex even in several thousand meioses.
Importantly, this sex specificity of individual hotspots is not constrained by regional controls. For example, the hotspot at 173.967 Mb is more active in males despite lying in the midst of a female predominant region, and the hotspot at 190.204 Mb, which is considerably more active in females, nevertheless lies in a male predominant region.
To address the broader question of how the total numbers and relative activity of hotspots differ between male and female meioses, we compared the two sexes across the female and male predominant segments of the subtelomeric 24.7 Mb region by extrapolating the resolution dependent trend lines for activity down to 5 Kb. Interestingly, the two regions gave distinct answers; greater female recombination in the proximal segment largely resulted from an increased number of hotspots, whereas in the distal segment, greater male recombination was primarily the result of increased recombination in a comparable number of hotspots (Table 4). In the proximal 9.8 Mb, where females had twice the recombination rate of males (9.0 cM vs. 4.2 cM), they had twice as many hotspots as well (72 vs. 34) that were somewhat more active, while in the distal 16 Mb where females have a significantly lower recombination rate than males (12.4 cM vs 19.8 cM), there were similar numbers of inferred hotspots (91 vs. 88) in the two sexes, but males had higher average recombination rates per hotspot.
These sex differences largely apply to lower activity hotspots, those less than 0.2 cM. The inferred numbers of hotspots with rates of up to 0.2 cM were significantly higher in females than in males over the entire 24.7 Mb (Table 4). However, this inequality did not hold for higher activity hotspots; both sexes had the same number of hotspots more active than 0.2 cM.
Fine mapping of crossover exchange points within hotspots made it possible to identify the parental chromosome initiating recombination and thereby show that the two parental chromatids are under independent recombinational control.
The locations of all 457 crossover events in five of the nine hotspots mapped to <3 kb resolution (marked with full red circles on Figure 3A) were further mapped using all available SNPs. In each case, the sites of crossing over were distributed over distances ranging from 500 to 2000 bp, which is a typical size for a hotspot [3] (Dataset S1). In some cases, recombination activities were distributed along the entirety of the hotspots regions following a single normal distribution, but in others they appeared to be the sum of two overlapping bimodal distributions. Distinguishing between the two distributions depended on the availability of SNPs for precisely mapping recombination events near the hotspot center. When such conveniently positioned SNPs were available, we observed that crossover events were predominantly located at the two sides of the hotspot, with very little or no recombination at the center (Figure 6B). According to the currently valid models of recombination, bimodal distribution will be observed when double strand breaks initiate in very narrow regions, and the crossover exchange points which are located at the sites of resolution of the Holliday junctions migrate sufficiently away from the initial sites of double strand breaks. Our finding that a bimodal distribution was observed when the necessary SNPs were available for detection suggests that this is likely to be the case for most hotspots.
For the hotspot at 186.3 Mb, the availability of particularly suitable SNPs (Figure 6A) allowed us to deduce that for this hotspot the B6 and CAST chromatids are under independent, sex-specific recombinational control. The sites of crossing over within the hotspot were quite different when the crossover products were B proximal-C distal v. C proximal-B distal. This was true for F1 animals derived from both reciprocal crosses, i.e. there were no imprinting effects. Among the 16 crossovers arising in female meioses, all B-C exchange points were positioned centromere-proximal to the center of the hotspot, whereas all C-B recombinants crossed over in the centromere-distal part. Thus, the center of the hotspot was of CAST origin in all crossovers (Figure 6C), indicating that, in this cross, recombination events in females only initiated on the B6 chromosome [5]. In males, which have 5.6 times higher recombination at this hotspot, there was also a strong bias towards initiation on the B6 chromosome, although the effect was not absolute. Crossover events of both types were distributed on both sides of the central region, indicating that recombination could initiate on either parental chromatid (Figure 6D). However, initiation on the B6 chromatid was 2.5 times more frequent than on the CAST chromatid.
Our results for the 186.3 hotspot clearly show that the overall control of recombination at a hotspot is the sum of distinct controls for each chromatid, and that this distinction applies to issues of both sex specificity and absolute recombination rates.
Examining 225 Kb intervals over the entire chromosome to compare F1 hybrids derived from the reciprocal crosses of B6xCAST and CASTxB6 provided statistically significant evidence for parent-of-origin effects on recombination activities in both sexes (p = 0.013 for reciprocal males and p = 0.009 for reciprocal females). The direction of imprinting was not uniform, and imprinting was only detected by finding a statistically significant excess of hotspots showing a preference for recombination in one direction of the cross or the other. In no case did we find absolute imprinting, where recombinants were significantly absent from one direction of the cross. A statistically significant difference was also detected in the fine mapped 24.7-Mb region of the chromosome in males (p = 0.001), but the difference was only marginally significant in females (p = 0.07). None of the higher activity hotspots in this region showed significant parent-of-origin effects after correction for multiple testing; rather, imprinting effects were restricted to medium- and low-activity hotspots. (See Tables S5 and S6).
However, although we detected slight but significant cumulative differences between reciprocal crosses in 225 Kb intervals in both female and male meiosis, and in male meiosis in the telomere-proximal 24.7 Mb, no one interval gave significant evidence for a difference in recombination rate between the reciprocal crosses. It is likely that the effects may be subtle and only recognizable statistically when data is accumulated across large chromosomal regions. Individual intervals, when considered on their own, showed recombination rate differences between the reciprocal crosses that could reasonably be explained by chance variation, but overall there were many more intervals with suggestions of recombination rate differences than could reasonably be explained by chance variation.
Additional data obtained from the backcross animals provided the first genetic evidence in mammals that genetic interference, which regulates the spacing of crossovers, does not affect the relative locations, one to the other, of the two distinct outcomes of the recombination process, crossing over and gene conversions not associated with crossing over.
Gene conversions arising in male meioses were detected in three of the fine-mapped hotspots by genotyping every SNP across each hotspot among 1365 male backcross progeny (Table 5). Only eleven conversions were found, six conversions not associated with crossovers (noncrossovers) and five conversions associated with simultaneous crossovers at the same hotspot. In the best mapped hotspot at 186.3 Mb, all five events we detected were positioned in the central part of the hotspot. The three noncrossovers were located between positions 1135–1311 bp on Figure 6B, and the two conversions associated with crossovers spanned between positions 877–1311 bp. For all three hotspots, the apparent frequencies of non-crossover conversions were lower (5–11 times) than crossover frequencies at the same hotspots, however these ratios must be interpreted with caution as while we were able to detect all crossovers, we were only able to detect the sample of conversions occurring at sites of available SNPs. The relative ratios of crossovers to noncrossover conversions in several human and mouse hotspots have shown considerable variation, from more than 12∶1 to 1∶4 [2],[5],[25],[42]. Given the positions of the available markers, the actual conversion frequencies could be much higher than detected. From SNP locations we could deduce that the minimum-maximum length for noncrossover conversion tracts was 9–279 bp. In contrast, conversion tracts associated with crossing over at the same hotspots had a minimum-maximum span of 199–1196 bp. Both estimates are of similar scale to those reported at the human DNA3 hotspot, 55–290 bp for conversion tracts not associated with crossovers and ∼460 bp for conversion tracts associated with crossing over [25].
The six progeny chromosomes carrying noncrossover conversions contained seven crossovers located elsewhere along the chromosomes. In four cases the distances between crossovers and conversions were significantly longer, 95–120 Mb, than the minimal male interference distance of 57 Mb between two crossovers observed in the 3026 male meioses used in this study [22]. However, in three cases the crossovers and conversions were only a few megabases apart, the closest distance being 1.12 Mb. We conclude that the process of genetic interference limiting the proximity of crossovers, one to another, does not limit the proximity of crossovers and non-crossover conversions. Our finding is in agreement with the lack of interference between crossovers and non-crossover conversions originally found in yeast [43].
This study presents the first high-resolution, comprehensive investigation of recombination as it occurs over an entire mammalian chromosome in a defined genetic background. As such, it provides material for further research, and as one might hope, generates as many questions as it provides insights.
The distribution of recombination along chromosome 1 provides genetic evidence that at least two levels of control regulate positioning of crossover events in mice; one is at a regional scale and another at the level of hotspot activity. This result is most apparent when comparing the genetic map created in the cross between B6 and CAST with the map reported for HS mice [17]; the two crosses share regional patterns of recombination but few if any hotspots. McVean et al [38], using linkage disequilibrium data, previously came to the same conclusion regarding human recombination.
In the case of mice, this substantial regional variation in the distribution of recombinational activities allowed us to examine the sex specificity of this phenomenon. Male recombination is concentrated at the telomere-proximal region, whereas female recombination is more evenly distributed along the chromosome. Importantly, however, the two sexes appear to share similar pattern of megabase-scale regions containing or lacking recombination as well as substantial portion of their hotspots within this regional variation, although at different activity levels.
The question then arises as to what the source of this regional variation might be as it is only to some extent related to exon density and not related to the other obvious biological feature of chromosomes-cytological banding patterns. The existence of alternating regions of high and low recombination suggests that regional recombinational activity might be an intrinsic property of genomic content. However, the general observation of high male recombination in subtelomeric regions suggests that positional effects, i.e. regional location relative to centromere and/or telomere may also play a critical role. Deciding between these possibilities may require comparisons of recombination patterns among chromosomes and between organisms carrying substantial chromosomal rearrangements.
Our data clearly show a multi-layered control of sex differences in recombination. First, averaging across the entire genome, females have an overall higher recombination rate than males. We have shown in another study [22] that the underlying reason for this is the crossover interference distance, which is shorter in females than in males when measured in megabases, allowing female chromosomes to accommodate more multiple crossovers. This difference in interference distances corresponds to differences in the length of the synaptonemal complex at pachynema [35],[44] and the synaptonemal complex length covaries with crossover/chiasma numbers [45]. Interference distances are the same in the two sexes when measured in microns of synaptonemal complex length, but the lesser compaction of female chromosomes results in fewer Mb of DNA per micron of length and hence greater opportunities for multiple crossing over.
The sexes also differ in the regional control of crossing over and the positioning of crossovers along the chromosome. Female recombination is distributed more evenly along the chromosome with alternating regional domains of higher and lower activity from centromere to telomere. In contrast, male recombination is more strongly localized, with two prominent peaks–one at the telomere-proximal region between 178–197 Mb and another at 27–79 Mb. It should be noted that the distance between the centers of the two male peaks equals the average intercrossover distance in male meiosis [22].
The sexes also differ at the local level in the usage of hotspots. Increased male recombination is associated with increased hotspots activity rather than an increase in the number of hotspots, whereas increased female recombination is associated with an increase in the number of hotspots of medium and low activity. These differences in hotspot usage are then reflected in the fact that the fraction of the chromosome (i.e. the number of 225 Kb intervals) exhibiting recombination is appreciably greater in females.
Finally, beyond these broad scale and regional effects there are truly sex-specific hotspots that may be found anywhere, including male specific hotspots in regions of predominantly female recombination and vice versa.
Our results examining mouse recombination show striking similarity to the features of sex specificity of recombination described in a human population-dramatic megabase-scale sex differences, similar overall use of hotspots by the two sexes, and examples of hotspots used mainly by one or the other sex [39].
The molecular origins of the sex effects must be complex, at the least involving differences in the nature of chromatin compaction during meiosis, the regional organization of chromatin, and sex-specific factors influencing the choice and activity of hotspots during meiosis. Given that the same chromosomal DNA sequences are the substrates for recombination in male and female meioses, these differences must reflect the existence of differentially transcribed, trans-acting factors controlling various aspects of recombination, but their identity is entirely unknown. Equally enigmatic are the biological functions and/or evolutionary selective pressures that underlie these differences. Do they have a primary function, or are they secondary consequences of other, underlying aspects of meiosis?
The exponential relationship between the frequency and activity of hotspots of different activity classes implies a probabilistic component to the determination of hotspot activity. This could result from a simple mechanism involving the accumulation of “units” that each contribute to the free energy requirement of hotspot activation. In the hope of promoting further discussion of what this “unit” might be, we here propose one possible formulation of the problem which suggests that an exponential function will be observed if two conditions prevail. The first condition requires that the relative activity of hotspots depends on the number of “units” they acquire. In this case, the probability of acquiring u units will be Pu = (P1)u, where each unit has a nearly equal but independent probability of being acquired. The second condition would require that each unit contributes a nearly equal increment of free energy, so that the free energy available to initiate recombination, ΔG, is proportional to u. Then, given the familiar relationship ΔG = −RT lnk, k (which we interpret as proportional to the forward rate constant of the initiating step) becomes proportional to eΔG = ecu. This formulation has the utility of focusing attention on the challenge of identifying the physical nature of a “unit”, which in principle could represent anything from formation of a single hydrogen bond to the assembly and/or disassembly of nucleosomes.
We found fairly strong evidence that parent of origin effects, i.e. imprinting, influence hotspot behavior. However, this is not expressed in a simple on-off manner as it is in many cases of imprinting control of gene expression where one parental allele is virtually silenced relative to the other. The failure to detect any overall preference for one parental direction vs. the other (B6xCAST vs. CASTxB6) likely reflects variation among hotspots as to which parental chromatid initiates recombination more frequently and hence which parental direction is favored. The imprinting effect on recombination was only apparent as a tendency when combining data from across the chromosome and could not be detected at statistically significant levels at any single hotspot, even when taking the issue of chromatid specificity into account. In females, the hotspot at 186.3, which only activated on the B6 chromatid, failed to show any recombination bias between reciprocally generated F1 animals. Our finding is somewhat surprising because it has been well established that methylation imprints at maternally or paternally expressed genes are erased during primordial germ cell development [46],[47] and reestablished during gametogenesis. The possible role of imprinting in recombination has been discussed previously [48],[49]. Despite a lack of prior evidence that it does occur, these authors argued that imprinting should play a role in recombination as it is the only process in ontogenesis that requires recognition and contact between homologous chromosomes. Additionally, the possibility holds attraction as a means of enabling the distinction between sister and non-sister chromatids, an essential feature of meiosis.
Finally, we are left with one of the ultimate questions in recombination biology; what makes a hotspot a hotspot? Several aspects of this question have been elucidated in yeast [50] where three classes of hotspots can be distinguished. Unfortunately, although the identification of a series of new hotspots does provide new experimental material, we are still far from adequately answering this most critical question in mammals, which has already been addressed extensively with limited success by others [26],[51],[52]. The most definitive progress has been made in identifying nucleotide motifs that could explain a fraction of recombination activity based on LD data [52] and recently confirmed by crossover mapping [39].
Previously, elucidation of the possibilities for cis and trans regulation of recombination activity in mammals [3],[5],[23] has relied on qualitative data. In humans, much higher recombination in females than in males has been reported for the TAP2 hotspot [53]. The most detailed investigation of cis and trans control of hotspot activity has involved the mouse Psmb9 hotspot. Shiroishi et al [23] established that this hotspot is active only in female meiosis and only when in the context of a particular surrounding chromosomal segment. When the centromere-proximal part of the active segment was replaced, hotspot activity was lost. In males, replacing the centromere-distal segment resulted in additional hotspot activation. Baudat and de Massy[5] have extended this analysis to present evidence that trans as well as cis acting factors regulate Psmb9 activity. The one refinement we can offer is the realization that, as exemplified by the hotspot at 186.3 Mb, the control of crossing over is chromatid specific. The control of a “hotspot” is, in effect, the sum of controls of the individual chromatids present at meiosis. Exploring this question in detail requires the ability to distinguish, quantitatively, the activity of each separate chromatid.
In conclusion, our data present a picture of recombination patterns along a chromosome that are controlled by a dynamic, complex regulatory system, with multiple levels of regulation depending on species identity, genetic variation, sex-specific mechanisms of recognition, and usage of specific hotspots. Only a fraction of all potentially available sites are used in a given F1 hybrid between two inbred strains, presumably as a function of the combined genetic contributions of both parents.
Improving our understanding of the structures and mechanisms bringing about these multiple layers of regulation for one of the most fundamental of biological processes is likely to cast light on several aspects of population genetics and evolutionary biology, as well as enhance our practical ability to define the genetic components of human disease.
C57BL/6J and CAST/EiJ were obtained from The Jackson Laboratory, Bar Harbor, USA. F1 hybrids were produced by reciprocal crosses in which either strain was the female or male parent. These hybrids were then backcrossed to C57BL/6J and recombination was detected in their progeny. All parents and F1 hybrids were genotyped for three markers on each chromosome to ensure strain identity using DNA isolated from tail tips.
To prepare DNA for genotyping, mouse spleens were digested in 900 µl buffer containing 50 mM KCl, 10 mM Tris-HCl, pH 8.3, 2.5 mM MgCl2, 0.1 mg/ml gelatin, 0.45% v/v Nonidet P40, 0.45% v/v Tween 20, and 60 µg/ml proteinase K overnight with occasional shaking. After digestion, the pH of the samples was adjusted by adding 100 µl of 100 mM Tris-HCl, pH 8.0. These digests were stored at −80°C. Samples were diluted 20x in 10 mM Tris-HCl, pH 8.0 for genotyping. All progeny were genotyped at 10 Mb resolution using previously described assays [54] for single nucleotide polymorphisms (SNPs) based on Amplifluor technology [55]. Individuals with a gap of >20 cM or >35 Mb between typed markers were omitted from subsequent analyses. Recombination was detected as a transition from homozygous to heterozygous genotype or vice versa. New Amplifluor assays were developed for the subsequent rounds of genotyping using the publicly available SNP database of the Mouse Phenome Project (http://phenome.jax.org/pub-cgi/phenome/mpdcgi?rtn=snps/door). In the second round, all recombinants detected were mapped at 200 kb resolution. In the subsequent rounds, recombinants were mapped to increased resolution until reaching the maximum hotspot resolution. In each round, the flanking markers from the previous round were retyped to confirm the validity of the recombinants. All detected conversions were confirmed by sequencing. This approach ensured extremely low error rate. A list of all markers used in this study is available as part of the Online Supporting Material (Table S1). The positions of all markers are in accordance with NCBI Build 36.
All the analyses were performed using R (http://www.r-project.org/) on the untransformed data (i.e. numbers of crossovers per interval). To compare recombination rates between groups (between the sexes, or between the two reciprocal crosses within one sex), first we tested whether there exists any difference in any interval across all intervals of the entire chromosome. An omnibus likelihood ratio test was used to compare the probability of the data if the recombination rate is allowed to be different across groups within each interval, versus the probability when the rates are forced to be the same across the groups for all intervals. A significant difference between the two groups indicates a difference in the recombination rate for at least one interval. The distribution and significance of the test statistics were determined via permutation method (>10,000 permutations). Then we tested the differences within individual intervals to see where the signal, if any, was coming from. Both likelihood ratio tests and Fisher exact tests were implemented and they produced similar p-values. These p-values were then transformed into q-values based on Storey and Tibshirani [56]. A q-value cutoff of 0.1 (equivalent to a false discovery rate (FDR) of 10%) was used to determine significant intervals.
The exon and transcript data was downloaded from the UCSC MySQL server (http://genome.ucsc.edu/FAQ/FAQdownloads#download29) using data from NCBI Build 36 of the mouse genome. The density is the fraction of the genome within transcribed sequences or exon coding regions, respectively, calculated in 50 Kbp blocks. Transcription start site density represented the number of 5′-gene ends per 50 kb. For the exon and transcript coverage, overlapping was treated as a continuous exon or transcript. Transcriptional starts only considered unique start sites; i.e., if two or more transcripts had a common start site, the site was only counted once. Correlation was calculated using the Pearson's product-moment correlation between the normalized recombination rate (cM/Mb) and the genomic feature (i.e., gene density, exon density, transcription start sites). The significance of the correlation was determined by 1000 bootstrap iterations, counting the number of correlations with an absolute value greater than the absolute value of the original correlation. Repetition of the bootstrap analysis found the results to be robust and no significant improvement was observed when using more than 1000 iterations. |
10.1371/journal.pgen.1007797 | Genome-wide identification of RETINOBLASTOMA RELATED 1 binding sites in Arabidopsis reveals novel DNA damage regulators | Retinoblastoma (pRb) is a multifunctional regulator, which was likely present in the last common ancestor of all eukaryotes. The Arabidopsis pRb homolog RETINOBLASTOMA RELATED 1 (RBR1), similar to its animal counterparts, controls not only cell proliferation but is also implicated in developmental decisions, stress responses and maintenance of genome integrity. Although most functions of pRb-type proteins involve chromatin association, a genome-wide understanding of RBR1 binding sites in Arabidopsis is still missing. Here, we present a plant chromatin immunoprecipitation protocol optimized for genome-wide studies of indirectly DNA-bound proteins like RBR1. Our analysis revealed binding of Arabidopsis RBR1 to approximately 1000 genes and roughly 500 transposable elements, preferentially MITES. The RBR1-decorated genes broadly overlap with previously identified targets of two major transcription factors controlling the cell cycle, i.e. E2F and MYB3R3 and represent a robust inventory of RBR1-targets in dividing cells. Consistently, enriched motifs in the RBR1-marked domains include sequences related to the E2F consensus site and the MSA-core element bound by MYB3R transcription factors. Following up a key role of RBR1 in DNA damage response, we performed a meta-analysis combining the information about the RBR1-binding sites with genome-wide expression studies under DNA stress. As a result, we present the identification and mutant characterization of three novel genes required for growth upon genotoxic stress.
| The Retinoblastoma (pRb) tumor suppressor is a master regulator of the cell cycle and its inactivation is associated with many types of cancer. Since pRb’s first description as a transcriptional repressor of genes important for cell cycle progression, many more functions have been elucidated, e.g. in developmental decisions and genome integrity. Homologs of human pRb have been identified in most eukaryotes, including plants, indicating an ancient evolutionary origin of pRb-type proteins. We describe here the first genome-wide DNA-binding study for a plant pRb protein, i.e. RBR1, the only pRb homolog in Arabidopsis thaliana. We see prominent binding of RBR1 to the 5’ region of genes involved in cell cycle regulation, chromatin organization and DNA repair. Moreover, we also reveal extensive binding of RBR1 to specific classes of DNA transposons. Since RBR1 is involved in a plethora of processes, our dataset provides a valuable resource for researches from different fields. As an example, we used our dataset to successfully identify new genes necessary for growth upon DNA damage exerted by drugs such as cisplatin or the environmentally prevalent metal aluminum.
| The first molecular function assigned to the human tumor suppressor Retinoblastoma (pRb) was that of a transcriptional repressor controlling entry into S-phase. It was shown that pRb binds and inhibits the function of E2F-DP transcription factors and that phosphorylation of pRb by CDK-cyclin complexes disrupts this interaction, releasing E2F-controlled genes from repression [1].
Since then, a wealth of additional functions in cell proliferation, differentiation, environmental response and genome stability have been discovered for the family of pRb related proteins in various organisms [2–5]. To date, more than 200 interactors of human pRb are listed in the BioGRID database [6], reflecting the multi-functionality of this molecular hub. Although there is evidence for a role outside the nucleus, e.g. in the cytoplasm to regulate nuclear import of viral proteins [7] and at mitochondria where pRb seems involved in the control of apoptosis [8], most functions of pRb-type proteins are chromatin associated [5].
The pRb–E2F module originated before the divergence of the plant and animal lineages, likely in the last common ancestor of all eukaryotes [9]. While there exists only one pRb homolog in C. elegans, i.e. lin-35, there are two homologs, Rbf1 and Rbf2, in Drosophila and in humans, pRb is member of a small gene family comprising pRb (p105), p107 and p130 [10]. In the model plant Arabidopsis thaliana, there is only one homolog, termed RBR1 (RETINOBLASTOMA RELATED 1) and its loss is female gametophytic lethal [11].
In their function as transcriptional regulators of the cell cycle, pRb-type proteins not only control G1-S transition in proliferating cells but also operate at other phases, i.e. as repressors of G2 and M phase genes in response to DNA damage [12] as well as in G0 to control quiescence as part of the DREAM complex.
The human DREAM complex consists of DP, pRb-like (p130 or p107), E2F and the Multivulval class B (MuvB) core-complex, comprising five additional proteins, LIN9, LIN37, LIN52, LIN54, and RBAP48. When cells exit G0 and reenter the cell cycle, the DREAM-complex is disassembled upon phosphorylation of p130/p107 and the MuvB-core-complex sequentially associates with other transcription factors, like B-Myb and Fox-M1 to activate different sets of genes important for subsequent phases of the cell cycle [13,14]. In Drosophila, a homologous complex, termed dREAM (Drosophila Rbf, E2F2 and Myb-interacting proteins), has been characterized but, in contrast to mammals, the fly pRb-type proteins and the only Myb transcription factor have been shown to concomitantly bind the MuvB-core complex [15,16].
Recently, different versions of an Arabidopsis DREAM-like complex were described by Kobayashi and co-workers and have been implicated in the control of mitotic genes [17]. Arabidopsis contains five three-repeat MYBs (MYB3R), that function as transcriptional activators and/or repressors of mitosis [17–19]. Since members of both classes of MYB proteins were found to interact with DREAM components, these data suggest the existence of activating and repressive DREAM-complexes depending on which MYB3R transcription factor is present [4].
In addition to its role in controlling progression through the cell cycle, RBR1, like its animal counterparts [20], is involved in developmental decisions, e.g. by interacting with non-E2F transcriptions factors such as SCARECROW (SCR) to control asymmetric cell divisions in the root meristem [21,22] and FAMA during stomata development [22,23]. Furthermore, RBR1 has been attributed a role in promoting the meiotic fate of the megaspore mother cell by repressing the expression of the stem cell factor WUSCHEL [24].
Like human pRb, which interacts with different chromatin modifiers, such as histone deacetylases and methyltransferases [25], the SWI-SNF complex [26,27], and the Polycomb repressive complexes (PRC) [28,29], Arabidopsis RBR1 has also been linked to chromatin modification [30]. For example, RBR1 has been shown to complex with the PRC2 subunits FIE [31] as well as MSI1 [32], and a RBR1-PRC2 cooperation has been proposed to allow the switch from late embryogenesis to autotrophic seedling development [33].
Another facet of Arabidopsis RBR1 is its role in DNA damage response (DDR) [34,35], which involves two different modes of action. On the one hand RBR1 binds to and represses DDR genes and by that likely links their regulation to activation of E2F and entry into the cell cycle [34,35]. On the other hand, RBR1 partially co-localizes with γH2AX, a marker for double strand breaks [34,35], and is required for the recruitment of the DNA repair protein RAD51 to DNA lesions [34]. In addition to its role at damage-induced double strand brakes (di-DSBs), RBR1 has also been shown to localize to SPO11-dependent foci in early meiotic prophase, most likely reflecting processed sites of programmed double strand breaks (p-DSBs), i.e. sites of cross-over formation [36].
Thus, like its animal homologs, Arabidopsis RBR1 is a multi-functional protein. However, despite several studies on specific aspects, a comprehensive view of RBR1 action has been precluded by the lack of genome-wide DNA binding data. Using an optimized ChIP protocol on Arabidopsis cell culture material, we now generated the first comprehensive collection of direct RBR1 targets. These data allowed us to obtain insights into the properties of RBR1 binding in mitotically active cells revealing the localization of RBR1 to genes and transposable elements (TEs). Furthermore, we then used this dataset to identify three new genes to be required for growth under genotoxic stress providing functional evidence for the power of our genome-wide study.
One likely reason for the lack of genome-wide RBR1-ChIP data is the indirect binding of RBR1 to DNA, i.e. via E2F and other transcription factors/chromatin modifiers. To close this gap, we introduced several changes to our standard plant ChIP protocol [37]. Starting out from liquid nitrogen ground tissue, we performed a double fixation step using Di(N-succinimidyl) glutarate (DSG) followed by formaldehyde to fix protein-protein- as well as protein-DNA interactions, a strategy successfully applied to improve ChIP results in human cell lines [38]. While the cross-linker Ethylene glycol bis(succinimidyl succinate) (EGS), has been used in ChIP experiments of the transcription factor CRY2 in Arabidopsis [39], we found DSG, a different length cross-linker, to work best for RBR1 among several long-range cross-linkers tested. Another important change to the standard protocol is the use of a douncer, which greatly enhances the release of nuclei prior to chromatin purification (S1A Fig, Material and Methods).
The protocol was applied to two replicates of an exponentially growing Arabidopsis cell culture (MM2d cells, S1B Fig) [40] using a RBR1-specific antibody [41]. ChIP signal to noise ratio and local resolution was high as verified by testing known RBR1 targets as well as negative controls by qPCR (S1C Fig). When analyzed on a whole genome level by pyro-sequencing, both ChIP replicates showed highly significant overlap (Fig 1, S2 Fig).
RBR1-bound domains were not only found to be associated with genes but also with transposable elements (TEs, Fig 1A). While in both replicates RBR1-binding is mainly found in the 5’ region of genes, the marking of TEs does not follow an evident pattern (Figs 1A, 1C and 1D, S2 Fig). A meta-gene profile analysis revealed that the peak of RBR1-binding mostly lies within 200 bp upstream of the transcriptional start of genes, but is also often found within the 5’ end of the transcribed region (Fig 1C). For further characterization of the RBR1-targeted loci, we used the overlap of both replicates (Fig 1B, S1 Table), i.e. a total of 937 genes and 475 transposons representing a robust core set of RBR1-bound elements. Since in other organisms pRb-type proteins have been shown to be associated with origins of replication (ORI) [42,43], we related our data to an ORI set of Arabidopsis [44]. While there is a small, but still significant overlap between gene-associated ORIs and our RBR1-gene set (P(X> = 66) = 3.24E-05), this is not the case for TEs (P(X> = 1) = 0.452, S3 Fig).
When we addressed the binding of RBR1 to TEs in detail, a clear overrepresentation of DNA-transposons was found while retrotransposons were nearly absent (Fig 2A). At the level of individual TE families, we found a more than 10-fold enrichment of Simplehat1, Simplehat2, Simpleguy1, Arnoldy1 and Arnoldy2 in the RBR1 dataset when compared to the whole genome frequency (Table 1). These families are all Miniature Inverted–Repeat TEs (MITEs), i.e. very short elements of the non-autonomous type, which do not carry any ORF. Unless inserted into the transcribed region of another gene, we assumed these elements to be transcriptionally silent. Indeed, none of the MITEs of our list of RBR1-bound TEs showed clear transcriptional activity in the wildtype or was significantly upregulated in ddm1 or met1 mutants in a recent whole genome transcriptional study of TEs [45] (S2 Table).
We next asked if RBR1-marked MITEs, when inserted inside or close to a gene, would necessarily influence that gene’s expression. However, neither AT1G65985, a gene that carries a Simpleguy1 transposon (AT1TE80690) in the second intron, nor AT1G60020, where a Simplehat2 transposon (AT1TE72980) is inserted into the upstream-region, showed upregulation in hypomorphic mutants of RBR1 (rbr1-2, Fig 2B–2D). In line with this, we do not see enrichment of RBR1-bound TEs in proximity of genes compared to the whole genome TE dataset (S3 Table).
Among the significantly overrepresented RBR1-bound TE-families, there is only one autonomous DNA-transposon family, VANDAL21. Its RBR1 decorated members, including the well-characterized TE Hiun [46], show transcriptional activity which is significantly upregulated in ddm1 and met1 mutants [45] (S2 Table). Hiun has been show to carry 3 ORFs, VANA, VANB and VANC [46] (Fig 2E). VANA encodes a protein with high sequence similarities to MURA-type transposases and VANC functions as a DNA demethylation factor implicated in the escape of Hiun from epigenetic silencing. Interestingly, the RBR1-peak marks the upstream region of VANB (Fig 2E) and an analysis of VANB-expression showed a significant upregulation in rbr1-2 mutants in comparison to the wildtype (Fig 2B). So far, a biological function has not yet been assigned to VANB. However, our observation that this ORF is under RBR1 control suggests a possible connection to the cell cycle and/or cell differentiation.
A motif analysis of the RBR1 bound domains associated with TEs using the MEME-Chip software [47] detected three highly overrepresented motifs (Fig 2F, see S1 Appendix for corresponding probability matrices). The most enriched motif (TE-motif 1) fits the consensus sequence described for the Arabidopsis E2F-transcription factor family, i.e. WTTSSCSS, where W stands for A or T and S stands for C or G [48] (S4A Fig). Matches to this more degenerate consensus motif were subsequently found in 77% of all RBR1-targeted TE associated domains.
In contrast to genes, the transposon-associated RBR1-preaks are often broad, sometimes overlapping the entire TE (S2B Fig). This finding is in accordance with the observation that transposon-associated E2F-sites are frequently organized in a microsatellite structure [49]. Consistently, when we quantify the number of WTTSSCSS motifs per RBR1-bound TE associated domain, we see an average of 8.6 and a median of 6 motifs per WTTSSCSS bearing domain, confirming a repetitive organization (Table 2). For comparison, the average of WTTSSCSS occurrences is 1.9 with a median of 1 in gene-associated domains.
To follow up the question if clustering is specific for TE-motif 1, we performed a MCAST-analysis (Motif Cluster Alignment and Search Tool-analysis) including all three overrepresented motifs. A total of 218 clusters were identified in the RBR1-marked TE associated domains 22 of which contained only TE-motif 1, while pure TE-motif 2 or TE-motif 3 clusters occurred merely once. In most clusters however all 3 motifs were present in a variety of different layouts (S2 Appendix).
The S-Phase genes PCNA1, ORC3, ORC1A, MCM5 and MCM2, the DNA repair genes RAD51 and BRCA1 as well as the cyclin-dependent kinases CDKB1;1 and CDKB1;2 have been shown to be RBR1 targets by gene specific RBR1-ChIP-qPCR experiments [23,24,33–35,50,51] and all of them are present in our core dataset of 937 genes (Fig 1B, S1 Table) highlighting the quality of the RBR1-ChIP result.
To get a whole genome view on RBR1-controlled processes in proliferating cells, we performed GO-term enrichment studies of the RBR1-bound genes. In agreement with an evolutionary conserved role of pRb-type proteins [5,52], several analyses using different algorithms showed highly significant enrichment of GO terms like cell cycle, DNA repair, DNA replication and chromatin organization (e.g. Table 3).
We also compared our RBR1-ChIP data with published gene lists covering different areas of interest (S5 Fig, S4 Table). These comparisons revealed that more than two-thirds of the Arabidopsis core replication machinery show RBR1-binding in our assay as well as more than one third of the main cell cycle genes, including RBR1 itself (S5A Fig and S5B Fig). A little less pronounced but still highly significant, is the overlap with genes involved in DNA repair and chromatin organization (S5C Fig and S5D Fig).
When we compared RBR1-bound loci with previously published RBR1 RNAi transcriptomes, there was considerable overlap with genes upregulated upon RBR1 depletion (Fig 3A). Almost half of the genes upregulated in roots of a RBR1-RNAi line in which an antisense RNA is specifically expressed in the root meristem [35], are bound by RBR1 in cell culture. Also the overlap with transcriptome data from young leaves using an inducible RNAi construct against RBR1 [53] is highly significant (roots, P(X> = 38) = 1.126e-33; young leaves, P(X> = 122) = 5.516e-39). For the latter dataset, representing a time-course after RNAi induction, the overlap is mainly seen with genes upregulated at 12 and 24 hours after RNAi induction (hai) but only marginally with the 3 and 6 hai-datasets (Fig 3B). Thus, it apparently takes more than 6 hours till RBR1 silencing and subsequent upregulation of direct RBR1 targets becomes evident.
Taken together, these analyses indicate that our data are a reliable whole genome representation of RBR1-controlled genes in mitotically active cells, covering the area of cell cycle, especially replication, DNA repair and chromatin organization and provide a valuable resource for further studies.
In cell cycle regulation, RBR1, like pRb-type proteins in other organisms, has been shown to complex with E2F and MYB transcription factors [17,51,54]. Hence, we next compared the targets of these transcriptional regulators with those of RBR1. Since RBR1 is best known as a repressor of E2F-controlled genes, we first related our dataset to several published E2Fa target datasets (see below) and genes having the E2F-binding consensus WTTSSCSS within 400 bp upstream of their translational start site (Fig 3C and S6 Fig). We chose 400 bp upstream of the start codon as a reasonable distance since up to that limit the WTTSSCSS motif was shown to be overrepresented in a group of E2Fa-DPa upregulated genes [48]. For non-protein coding genes 400 bp upstream of the beginning of the gene were used. The E2Fa datasets in our comparisons included two transcriptional datasets from seedlings containing genes with increased transcription upon E2Fa-DPa over-expression [48,55], data from technically diverse E2Fa chromatin purification experiments using cell culture (ChIP, Chromatin Affinity Purification (ChAP) and Tandem Chromatin Affinity Purification (TChAP)) [56] as well as results of a DNA affinity purification sequencing approach (DAP) from young leaves [57].
Several conclusions can be drawn from this comparative analysis: First, the overlap of RBR1-targets with E2Fa-DP responsive genes (S6B Fig and S6C Fig) as well as with genes showing direct E2Fa association (ChIP, ChAP, TChAP and DAP, Fig 3C, S6A Fig and S6D Fig) is very large, in accordance with an important role of E2F transcription factors in RBR1 targeting. Second, the WTTSSCSS site on its own, even if positioned in the 5’ region of a gene is not sufficient for RBR1 binding as only a fraction of all genes having a WTTSSCSS motif within 400 bp up-stream of the start codon were associated with RBR1 in our ChIP experiment. Yet, this is similar for E2Fa target genes (Fig 3C and S6 Fig) as noticed before [48,55–57]. The third general observation is, that RBR1 as well as E2Fa also bind to genes that do not contain a consensus WTTSSCSS site in their 5’ region, indicating that binding can occur to a more degenerate or even completely different motif (see below). Finally, there are WTTSSCSS containing, RBR1-bound genes which are apparently not regulated by E2Fa, possibly reflecting a dependency on the developmental context and/or control by other members of the E2F family.
In a second set of comparative analyses, we related our RBR1-ChIP with whole genome ChIP data for MYB3R3, a repressive MYB-transcription factor that has recently been shown to be part of a plant DREAM-like complex (Fig 4A) [17]. Notably, more than half of the MYB3R3 gene targets also exhibit RBR1 binding and almost a quarter bind E2Fa in addition to RBR1 based on an E2Fa dataset from Verkest et al. [56]. A meta-analysis locating the position of RBR1 binding with respect to the center of E2Fa and MYB3R3 bound sites, revealed peaks centered around the same position (S7 Fig), which is in accordance with all three proteins being part of a DREAM complex regulating the same genes. Although there is a weak preference of MYB3R3 for M-Phase associated genes (MYB3R3: P(X> = 66) = 3.78E-52; RBR1: P(X> = 48) = 1.38E-15) and RBR1 binding is slightly more pronounced for S-Phase associated genes (MYB3R3: P(X> = 30) = 1.03E-14; RBR1: P(X> = 58) = 7.88E-24), the general picture is that both MYB3R3 and RBR1 bind to and potentially control S- as well as M-Phase genes (Fig 4B and 4C).
Next, we performed a GO-term enrichment analysis of the genes bound by RBR1 and MYB3R3 (RBR1/MYB3R3-overlap), by RBR1-only and by MYB3R3-only (S8 Fig). This revealed an enrichment of the GO-terms DNA replication, chromatin organization, chromosome segregation, DNA recombination, DNA repair and mitosis in the RBR1-only and in the RBR1/MYB3R3-overlap groups while of these GO-terms only mitosis is enriched in the MYB3R3-only class (S8B Fig). Furthermore, when we compared these three gene sets with genes upregulated in a myb3r1 myb3r3 myb3r5 triple mutant [17], we saw preferential overlap with the MYB3R3-only group (S8A Fig). Thus, there seems to be a sub-class of mitotic genes under MYB3R control that are not co-repressed by RBR1, an example being the cytokinesis specific syntaxin KNOLLE [58] (S2C Fig).
Since not all RBR1-bound genes appeared to be targets of E2F and/or MYB3R3 (see above), we performed a search for overrepresented motifs in our RBR1-ChIP data. A MEME-ChIP analysis with standard settings identified six motif-clusters. The most significantly enriched motif of each cluster is shown in Fig 5 (gene-motif 1–6, see S1 Appendix for corresponding probability matrices).
To estimate the genome-wide frequency of these motifs as well as the distribution among the MYB3R3- and/or RBR1-marked protein coding genes, we performed a FIMO-analysis (FIMO, Find Individual Motif Occurrences) [59] using 400bp upstream of the start codon as target sequences. Since FIMO counts all individual motif occurrences, we also clustered overlapping sites and summed up the clusters reducing the number of counts especially for repetitive motifs like gene-motif 1 and gene-motif 3 (Table 4).
This analysis showed that sequences matching the repetitive motifs 1 and 3 occur at high frequency within 400 bp upstream of the translational start of genes on a genome-wide level. Gene-motif 1 strongly resembles the so-called GAGA-motif (see S4 Fig for motif alignments), which has been described as an element of PREs (polycomb responsive elements) in animals [61,62] and plants [63,64] while gene-motif 3, also named translocon1-motif (TL1, GAAGAAGAA), has been shown to be bound by TBF1, a heat-shock factor-like protein associated with the expression of defense response genes [65]. Whereas RBR1 function has not yet been documented for drought stress or ABA-signaling, processes associated with the less frequent gene-motif 4 (ACGTGKC) [66,67], a very similar motif has been found enriched in plant PREs as well [64] and related to a motif called G-box (CACGTG). A third match to motifs described as relevant for PREs is gene-motif 5, which resembles the so-called telobox (AAACCCTAA) [64]. In addition, teloboxes, which consist of 1.3 units of the Arabidopsis telomere repeat, are enriched in promoters of components of the translational machinery and can be bound by AtPurα [68]. Noteworthy, complex formation of AtPurα and E2Fa has been documented in Arabidopsis [69] and the mammalian homolog, Purα, has been shown to directly interact with pRb [70] suppressing the transcriptional activity of E2F-1 [71].
As expected from our comparative studies with E2Fa, gene-motif 2, which matches the E2F consensus, is the strongest enriched motif in the complete RBR1 dataset as well as in the RBR1/MYB3R3 and the RBR1-only fraction from the RBR1/MYB3R3 comparison (Table 4). Furthermore, gene-motif 6, which shows very significant enrichment in the RBR1/MYB3R3 and the MYB3R3-only set, contains the core of the MSA-element (AACGG) found in the promoter of mitotic genes and known to be bound by MYB3R transcription factors [72].
Since DNA repair is among the highly enriched GO-terms in our RBR1 core dataset and since it was shown that at least a few DDR genes are under direct RBR1 transcriptional control [34,35], we decided to zoom into this role of RBR1 as a functional test case of our work.
Exposure to stresses, such as DNA damage, usually causes a cascade of transcriptional responses making it difficult to separate primary from subsequent and/or indirect effects. We hypothesized that combining the criteria “transcriptional upregulation upon DNA damage” and “gene bound by RBR1” might be a valid approach to identify so far uncharacterized, yet important DDR genes. We also postulate that “gene bound by RBR1” might be better suited than the criterion “upregulated upon loss/reduction of RBR1 activity” since the reduction of RBR1 by RNAi or the use of the hypomorphic rbr1-2 mutant only resulted in a rather weak upregulation of DDR genes [34,35].
It was previously proposed that the regulation of DDR genes by RBR1 could represent a priming mechanism, i.e. the coupling of DDR gene expression to the cell cycle might open the chromatin of these genes in dividing cells, in which DNA damage is especially critical. This opening of the chromatin would then make them easily and fast accessible for other, DDR specific transcriptional regulators, such as SOG1 [34].
For our analysis, we made use of publicly available transcriptional profiles of various Arabidopsis tissues after treatment with DNA damaging agents [73–82, GEO series GSE5620 and GSE5625]. We extracted the transcriptionally upregulated genes from 32 experiments (S5 Table) and calculated the overlap with our RBR1-ChIP dataset as well as a reference list of genes involved in DNA repair (S4 Table). In total, 8907 genes were found to be upregulated in DNA stress experiments, 307 of which are RBR1 targets according to our analysis. As shown in Fig 6A there is an about tenfold enrichment of genes involved DNA damage repair in the RBR1 bound subset (11.1%) compared to the non-RBR bound group (0.9%) of transcriptionally upregulated genes providing proof of concept for the validity of our approach.
Fig 6B displays all genes that are upregulated under DNA stress in more than three experiments and that are also bound by RBR1. In search of genes with a not yet described role in DDR, we selected four candidates based on the availability of homozygous insertion lines, i.e. AT1G04650, AT2G45460, AT3G20490 and AT5G46740 for further analysis (Fig 6B, black label; Material and Methods). To verify if RBR1 binding to these genes indeed reflects transcriptional inhibition, we monitored their expression in the wildtype and rbr1-2 mutants using qRT-PCR. As for known DNA-damage regulators like BRCA1 and RAD51, we see a slight, yet significant upregulation of all four candidate genes in rbr1-2 mutants (S9 Fig).
After confirming the absence of full-length transcripts in mutant lines of the candidate genes (S10 Fig), we analyzed them in root growth assays on different DNA damaging agents. In a first set of experiments, we used bleomycin and cisplatin since we have previously shown that rbr1-2 mutants are sensitive both toxins [34] (Fig 7). Bleomycin induces double strand breaks, which can be repaired by non-homologous end joining (NHEJ) and homologous recombination (HR). Cisplatin also causes DNA breaks and in addition, DNA cross-links, which require homology-dependent DNA repair. In addition we tested for root-growth on hydroxy urea (HU) containing media to complement our set of DNA damaging drugs with an agent causing replication stress, which eventually leads to double strand breaks in S-Phase that can be repaired by HR as well (S11 Fig).
Whereas plants mutant for AT2G45460 grew like the wildtype on bleomycin, cisplatin as well as HU containing media, the loss of any of the other genes resulted in different patterns of hypersensitivity to these three DNA damaging drugs (Fig 7, S11 Fig). While this work was in progress, the closest rice homolog of AT1G04650 was shown to participate in meiotic recombination and designated MEICA1 (Meiotic Chromosome Association1) [83]. More recently, AT1G04650 itself was found to be an interactor of the anti-crossover factor FIDGETIN-LIKE-1 (FIGL1) in Arabidopsis and therefore named FLIP (FIDGETIN-LIKE-1 INTERACTING PROTEIN) [84]. While FLIP’s crossover limiting role in meiotic recombination has been clearly demonstrated, a likely analogous function in DNA damage repair has only been speculated on. AT5G46740 will be referred to as UBP21 (Ubiquitin-specific protease 21), according to the nomenclature introduced by Yan et al. [85] and AT3G20490 will be called KNOTEN1 (KNO1, German for “to knot, to tie together”) since the mutant shows an accumulation of DNA lesions upon genotoxic stress (see below).
Mutants in KNO1 showed a strong growth inhibition on cisplatin, were only mildly affected by HU and displayed no significant growth reduction on bleomycin at the concentrations used in our assay (Fig 7, S11 Fig). For the flip lines, we observed a clear mutant phenotype on cisplatin as well as bleomycin although the growth inhibition on cisplatin was less pronounced than for kno1. Growth of flip mutants on HU was mildly, yet significantly reduced, similar to that of kno1 plants. Mutants for UBP21 were not affected by HU, but showed a mild growth inhibition on cisplatin and bleomycin containing plates, the latter being slightly more effective.
We further tested all hypersensitive lines for recovery growth after treatment with 1.5 mM aluminum (S12 Fig). Bioavailable aluminum (e.g. as Al3+) is a toxin plants are frequently exposed to on acidic soils [86] and previous work has indicated that it also induces DNA breaks [87]. Significant reduction in recovery growth was seen for the kno1 mutants from day three after treatment onwards. Also the flip mutant lines showed a clear trend towards growth reduction on aluminum. However, the result was statistically significant only for line flip-3 after 4 days. In contrast, the ubp21 mutants did not show any obvious reduction in recovery growth after aluminum treatment at the conditions tested.
To analyze if the observed hypersensitivity of kno1, flip and ubp21 plants to genotoxic agents is indeed due to enhanced DNA damage, we monitored γH2AX foci as a marker for double strand breaks after short term incubation in media with and without cisplatin or bleomycin [88]. While wildtype plants only showed few γH2AX-foci upon DNA stress under these conditions, we observed an enhanced accumulation of foci in the mutant lines (Fig 8). Whereas the damage for kno1-1 and flip-2 on cisplatin was slightly more severe than the damage on bleomycin at the conditions tested, the opposite was true for ubp21-1, which showed a stronger accumulation of γH2AX-foci on bleomycin than cisplatin in agreement with its slightly higher sensitivity towards bleomycin in the root growth assay.
Next, we asked if the genes identified might be involved in signaling of DNA damage. We therefore used qRT-PCR to check the respective mutant lines for expression of the SOG1 targets CYCB1;1 and RAD51, known to be transcriptionally upregulated upon DNA lesions (Fig 9A and 9B). While expression of CYCB1;1 in upb21-1 is at wildtype level in stressed and non-stressed plants, it’s upregulation upon cisplatin treatment is significantly less pronounced in the kno1-1 line and significantly more upregulated in flip-2 mutants when compared to the wildtype (Fig 9A). A similar trend is seen for RAD51 (Fig 9B). This result indicates that kno1-1 plants have problems in transmitting a DNA damage-induced signal, which normally leads to RAD51/CYCB1;1 upregulation, while the finding for flip-2 is in accordance with an impaired repair process, where the plant shows a compensatory response by transcriptional upregulation of repair pathway components.
Finally, we analysed RAD51 localization as a marker for the assembly of the HR repair machinery upon treatment with cisplatin or bleomycin (Fig 9C). In upb21-1 and flip-2 mutants RAD51 localized in a wildtype-like pattern, indicating that RAD51-mediated homology search still takes place in these mutants, whereas no clear RAD51-foci were seen in kno1-1 mutants. The reason for this could be the reduced RAD51 transcription. However, since RAD51 upregulation is only reduced but not abolished in the kno1-1 line, the lack of RAD51 foci might also indicate an additional function of KNO1 in the proper recruitment of the repair machinery to lesion sites.
Taken together, using “RBR1 binding” and “transcriptional upregulation upon DNA stress” as combined criteria is an efficient approach to identify new genes involved in different aspects of the DNA damage response.
Here, we present the first genome-wide RBR1-ChIP dataset for plants. Using proliferating cells of Arabidopsis, we identified a core set of 937 genes and 475 TEs marked by RBR1. The high reliability of our dataset is indicated on the one hand by the GO-term enrichment results, which are in accordance with Rbf1/Rbf2-ChIP results from flies [89–91] and ChIP results for human pRb-type proteins [92–94] and on the other hand by a strong overlap with gene sets regulated by proteins known to form a complex with RBR1, like E2Fa-DP and MYB3R3. Our data reveal a preferential RBR1-binding to the 5’ end of genes, as expected for a transcriptional regulator and a strong enrichment of the E2F consensus sequence WTTSSCSS in the RBR1-bound domains.
However, it needs to be tested if the analysis of different tissues/cell types will complete the list of RBR1-targets, especially since previously identified RBR1-targets involved in developmental processes were not detected by our approach [21,22,50]. We assume that this discrepancy is not due to technical constraints of the established ChIP protocol, which resulted in very reproducible and strong signal enrichments, but rather indicates that RBR1 binding to, and repression of developmental targets is temporally and/or spatially restricted.
For Arabidopsis, there are several indications of interplay between RBR1 and PRC2, a chromatin associated complex implicated in the stable repression of genes turned off during developmental progression [95,96]. On the one hand, the PRC2 components MSI1 and FIE have been shown to interact with RBR1 [31,32]. However, MSI1 likely also acts independently of PRC2 since it is an essential part of the CAF-1 complex [97] and by homology to the mammalian RBBP4 it is a putative component of a plant DREAM complex [17]. In support of this notion, transcriptional control of MET1 in the endosperm depends on MSI1-RBR1 but is independent of PRC2 [32]. On the other hand, three of the DNA sequences enriched in our RBR1-ChIP, i.e. the very frequent and repetitive gene-motif 1 (GAGA) as well as gene-motif 4 (ACGTGKC) and the telobox-like gene-motif 5, resemble DNA motifs that were recently shown to contribute to PREs in plants [64]. Nevertheless, when we compared the RBR1 targets with gene sets marked by either FIE or by H3K27me3, the chromatin mark reflecting PRC2 action, we did not find significant overlap on a genome wide level. On the contrary, relating lists of RBR1 bound and H3K27me3 decorated genes, we see significantly less overlap than expected by chance indicating a mutually exclusive pattern [64,98,99] (S13 Fig). Thus, either the overrepresentation of similar motifs in RBR1 and PRC2 bound domains is due to some higher order similarity between both gene sets or concomitant/interdependent gene repression by both regulators takes place only transiently and therefore is not seen by analysis of data from different tissues.
Our ChIP data indicates that RBR1 is recruited to E2F-sites that have been picked up and amplified by TEs in Arabidopsis. It has been reported that spreading of TEs with E2F-binding sites in microsatellite structure also occurred in other Brassicaceae species [49]. Interestingly, it is not always the same TE family showing this sequence motif expansion, although there is a clear bias for DNA-transposons, more specifically MITES. This suggests that the local retention of E2F/RBR1 is beneficial regarding TE amplification, as the E2F sequence motif occurs in different TE families even in closely related species and therefore must have accumulated after their evolutionary separation (Henaff 2014). Since MITES do not carry any ORF, this advantage is likely unrelated to RBR1’s role in transcriptional control. Because replication timing correlates with chromatin accessibility [100], one possibility is, that phosphorylation of RBR1 at G1/S and thus dissociation from E2F might lead to a decompaction of chromatin structure at the start of S-phase allowing for early replication of the affected loci and thus, giving a higher chance for multiplication by transposition from a newly replicated chromatid to a yet unreplicated site. Alternatively, local RBR1 accumulation might be beneficial for DNA-repair upon transposition of MITEs. In this respect, it is noteworthy not only that the mobilization of the DNA transposon Sleeping Beauty in human cell lines depends on Xrcc3/Rad51C, a complex that functions during homologous repair (HR), and on Ku70/Ku80, a key player in non-homologous end-joining (NHEJ), but also that Sleeping Beauty transposase directly interacts with the Ku70/Ku80 hetero-dimer [101]. Conversely, the involvement of pRb in canonical NHEJ and HR has been described, and it has been shown that pRb interacts with the ku70/ku80 hetero-dimer as well [102,103]. Also for Arabidopsis, we and others have seen that RBR1 is involved in DNA repair at the site of the lesion [34,35], partially co-localizing with RAD51, a major player in HR [34]. It is known that TEs use and modify the cellular machinery of the host at several levels to promote their own survival [104,105]. Thus, RBR1 might provide a link to the recombination/repair machinery required for stable MITE transposition in the host genome.
Additionally, in case of the non-MITE transposon Hiun, which belongs to the VANDAL21 family, we see a different example of interplay between the host machinery and the transposon, since one of the Hiun-localized ORFs is transcriptionally controlled by RBR1 and therefore potentially activated during G1/S phase, the moment when DNA transposons excise and mobilize [106].
To make further use of the information derived from our genome-wide study, we combined the core-set of RBR1-bound genes with transcriptional data from DNA stress experiments. This led to the identification of three genes with so far unknown function in protection against DNA damage. At the beginning of this study, only two of the four genes analyzed had a functional annotation based on homology, i.e. AT5G46740 as ubiquitin-specific protease (UBP21) and AT2G45460 as SMAD/FHA domain-containing protein, while KNO1 (AT3G20490) and FLIP (AT1G04650) were described as genes of unknown function. With the new set of mass annotation provided by Araport11 [107], KNO1 became annotated as putative Rho GTPase-activating protein and FLIP as Holliday junction resolvase, but both annotations are still lacking experimental support in Arabidopsis. Holliday junction resolvases function in meiotic as well as somatic HR in different organisms [108] and FLIP has recently been shown to act as a suppressor of meiotic crossovers in complex with FIGL1 [84]. A role in damage induced HR has not been shown so far, yet is very likely in light of our findings.
Our results show that KNO1 is needed after DNA damage to efficiently up-regulate and probably also localize components of the HR repair machinery like RAD51. In humans genotoxin-induced DNA damage stimulates nuclear Rac1, a Rho GTPase required for the activation of stress kinases [109]. However, plants do not possess Rac1 orthologs, but a plant specific family of Rho-type GTPases (Rop) instead [110], which to our knowledge has not yet been linked to DDR. Notably, KNO1 as well as UBP21 are among 146 recently identified direct targets of the major DNA damage related transcription factor SOG1 [111], adding further support for their role in DDR. In addition, ubiquitin-specific peptidase 21 (USP21) from human, which like UBP21 of Arabidopsis is a ubiquitin carboxyl-terminal hydrolase, has been shown to de-ubiquitinate and stabilize BRCA2 to promote efficient RAD51 loading at DNA double-strand breaks [112] and it is tempting to speculate if UBP21 has a similar role. However, although we cannot exclude subtle quantitative effects, we still see RAD51 loading in ubp21 mutants. Further studies are needed to unravel the exact molecular function of Arabidopsis KNO1, FLIP and UBP21 in somatic cells to fully understand their here discovered requirement under DNA damaging conditions.
Here we have presented the first genome wide RBR1-binding study in plants. We show, that RBR1 associates with TEs, especially MITEs, and with genes highly enriched for GO-terms like cell cycle, replication, chromatin and DNA repair in actively dividing cells. However, previously described developmental RBR1 targets remain unmarked. To investigate RBR1's proposed role as a potential integrator of cell cycle regulation with developmental processes, genome-wide RBR1 distribution at specific developmental time points and in defined cell types will be beneficial and can be achieved by applying our optimized ChIP protocol in combination with FACS or INTACT methods [113,114].
Further, our results demonstrate a vast commonality of genes bound by E2Fa, MYB3R3 and RBR1. In this respect, it will be valuable to gain and integrate information on genome wide binding of the different MYB3R and E2F transcription factors as well as RBR1 in distinct cell types as well as upon short and long term DNA damage. Recently, an involvement in DDR has been shown for repressive MYB3R proteins, but so far only the impact on G2/M genes has been analyzed in detail [115]. A comprehensive, context-dependent analysis will reveal if specific compositions of the DREAM complex govern the expression of the same genes in different cellular contexts or if different DREAM complexes have distinctive targets.
Finally, the here presented set of RBR1-controlled genes is a valuable resource that can be exploited to identify new genes involved e.g. in cell cycle control, chromatin remodeling and DNA repair as exemplified by our successful approach to reveal new DNA damage regulators.
Triplicates of the Arabidopsis MM2d cell culture, ecotype Landsberg erecta [40], were collected 3 days after sub-culture and frozen at -80°C. The material was homogenized by thorough grinding using mortar and pistil with permanent addition of liquid nitrogen. 300mg of the powder was dissolved in 10ml fixation buffer (10mM Hepes pH = 7.6, 0.5M Sucrose, 5mM KCl, 5mM MgCl2, 5mM EDTA, 14mM β-Mercapto-ethanol, 2.5mM DSG [Sigma, Di(N-succinimidyl) glutarate], protease-inhibitor [Roche, cOmplete tablets], 0.6% Triton X-100) and incubated for 1 hr at room temperature (RT) with gentle agitation (turning wheel). 300μl formaldehyde (Sigma, 37% FAA solution) was added to reach a final concentration of 1% FAA and incubated for exactly 5 min with gentle agitation. To stop the crosslinking reaction, 1ml of 2.5M Glycine was added to the solution and mixed immediately. The cross-linked material was transferred on ice and nuclear isolation was performed using a dounce tissue grinder set (Sigma, 100ml D0189). The lysate was filtered through a miracloth mesh and centrifuged in a swinging rotor using 50ml Falcon tubes at 3000g for 10 min at 4°C. The pellet was dissolved in 300μl nuclear isolation buffer (10mM Hepes pH = 7.6, 0.5M Sucrose, 5mM KCl, 5mM MgCl2, 5mM EDTA, 14mM β-Mercapto-ethanol, protease-inhibitor [Roche, cOmplete tablets]) by gentle shaking and overlayed onto a 600μl 15% Percoll solution (HEPES pH8.0 10mM, 15% (v/v) Percoll (pH8-9), 0.5M sucrose, 5mM MgCl2, 5mM KCl, 5mM EDTA) in a 1.5ml tube. After centrifugation at 3000g for 5 min at 4°C all supernatant was removed and the pellet was dissolved in 500μl nuclear lysis buffer (50mM Tris-HCl, pH7.5, 0.1% SDS, 10mM EDTA) without generating bubbles and vortexed thoroughly for 1 min. Sonication was carried out using a cooled Diagenode Bioruptor with a 45 sec ON– 45 sec OFF cycle for 2x15 min (water was changed between the cycles). Sonication efficiency was tested by de-crosslinking 20μl of the sonicated solution overnight and migrating on a 1.5% agarose gel. Sonication should lead to fragmentation of all gDNA to a fragment size of 200–500nt length (in case of remaining high size gDNA the nuclei solution needs to be further sonicated and tested for proper fragmentation). 100μl of the fragmented chromatin was then incubated in 1ml ChIP dilution buffer (15mM Tris-HCl, pH7.5, 150mM NaCl, 1% Triton-X-100, 1mM EDTA) over night at 4°C using a turning wheel with magnetic beads (Merck, Magna ChIP Protein A+G Magnetic Beads), pre-incubated for 1 hr in 500ul ChIP dilution buffer at 4°C with 1μg of affinity-purified anti-RBR1 [41] and anti-E2Fa antibody [116], respectively. Beads without antibody were used as negative control. The next day the beads were washed once by resuspending/pipetting first and subsequently for 15 min on a turning wheel with 500μl of the following washing buffers: (1) at 4°C: ChIP-dilution buffer (see above), (2) at 4°C: Low Salt buffer (20 mM Tris-Cl pH 8.0, 0.1% SDS, 1% Trition X 100, 2 mM EDTA, 150 mM NaCl), (3) at 4°C: LiCl buffer (20 mM Tris, pH 8.0, 0.25 M/0.5M LiCl, 1% NP40/Ipepal, 1% deoxycholate, 1 mM EDTA), (4) at room temperature: TE. The immuno-complex was eluted from the beads using an incubation with twice 250μl elution buffer (prepare fresh: 0.1M NaHCO3, 1% SDS) for 10 min at 65°C with gentle agitation. To de-crosslink the DNA, 20μl of 5M NaCl was added and left at 65°C over night, together with the input material (1% of the quantity used in the IP, volume was adjusted to 500μl using elution buffer). The next day, Proteinase K was added (20μl Tris, pH 6.5, 10μl 0.5M EDTA, 20μg Proteinase K) and incubated at 45°C for 2 hrs. DNA was isolated using Phenol-Chloroform purification and precipitation by Na-acetate/EtOH. The pellet was re-suspended in 20–50μl TE from which 1μl was used for a single qPCR reaction.
ChIP-seq analyses were performed with two biological replicates. For each replicate, 1 ng of immunoprecipitated (IP) and genomic (INPUT) DNA were used to prepare libraries with the MicroPlex Library Preparation kit (Diagenode). Quality of libraries was validated using 2100 Bioanalyzer (Agilent). Multiplexed libraries were sequenced using a HiSeq 2000 system (Illumina) with single-end 50-bp reads. Following a FASTQC (version 0.11.5) quality control, reads were mapped onto the TAIR10 Arabidopsis thaliana genome assembly using Bowtie (version 0.3 [117]) run in the sensitive mode, allowing one mismatch and randomly choosing one map position in case of multiple matching. MACS (version1.4.2 [118]) was used for peak detection including INPUT DNA as control and using the following parameters: Effective genome size = 120 Mbp, tag size = 50 bp, bandwidth = 150 bp, P value cutoff for peak detection = 1e-05, MFOLD range = 10,30. The average peak width is 930 bp (replicate 1) / 670 bp (replicate 2) and the median is at 600 bp (replicate 1) / 450 (replicate 2). Peaks were assigned to a gene or TE using an iterative procedure: (1) peak overlaps with gene or TE by at least 150 bp, (2) peak overlaps with gene or TE by at least 50 bp, (3) peak overlaps with 150 bp up-stream or downstream sequence of a gene/TE.
The MetaGene Profile was generated using the tool makeMetaGeneProfile.pl of the Homer Software [119]. Venn diagrams were generated using the VENN diagram generator designed by Tim Hulsen at http://www.biovenn.nl [120]. The test for statistical significance of the overlap between two groups of genes was calculated using the phyper function in R [121]. Data sources for comparative analyses are given in the text or the respective figure or table legends. GO-term enrichment analysis was done with PANTHER Version 13.0 [122], Fisher’s exact test was selected as test type and the Bonferroni correction has been applied to the P values. We used MEME SUITE [123] for Motif-analysis, i.e. the tools MEME-ChIP [47], including MEME [124] and DREME [125] for motif discovery, as well as AME [60] for calculation of motif enrichment, FIMO [59] to count individual motif occurrences and MCAST [126] to perform a motif cluster analysis. The Integrative Genomics Viewer (IGV) [127] was used to display signal distribution over representative genes and TEs.
Plants were germinated and grown on vertical plates containing half Murashige and Skoog (1/2 MS) medium under long day light conditions (16h) at 22°C for 6 days. Chemicals used in this study are bleomycin (bleocin, Duchefa), cisplatin (Nacalai Tesque) and hydroxyurea (Sigma-Aldrich). Seedlings were transferred to medium with or without 0.3 μg/ml bleomycin, 20 μM cisplatin, 2 mM hydroxyurea or 2.5 mM hydroxyurea and grown for 6 days further. The position of the primary root tip was marked daily for each plant. After 6 days, plates were photographed and root length was measured using ImageJ software. Data are presented as mean ± SD (n > 30). Significant differences from wildtype were determined by Student’s t-test: *, p < 0.05.
For the aluminum recovery growth assay, plants were germinated and grown on vertical plates containing ½ MS medium for 6 days. Seedling were then transferred to 1.5 mM Al-containing hydroponics (water solution (pH 4.2) consisting of 1 mM KNO3, 0.2 mM KH2PO4, 2 mM MgSO4, 0.25 mM (NH4) 2SO4, 1 mM Ca(NO3)2, 1 mM CaSO4, 1 mM K2SO4, 1 μM MnSO4, 5 μM H3BO3, 0.05 μM CuSO4, 0.2 μM ZnSO4, 0.02 μM NaMoO4, 0.1 μM CaCl2, 0.001 μM CoCl2) prepared as previous described [128,129] and treated for 12 hrs. Treated seedlings were planted on vertical ½ MS medium and allowed to grow for 5 days. The position of the primary root tip was marked daily for each plant. After 5 days, plates were photographed and root length was measured using ImageJ software. Data are presented as mean ± SD (n > 30). Significant differences from wildtype were determined by Student’s t-test: *, p < 0.05.
10-day-old seedlings were transferred to ½ MS liquid medium containing 3μg/ml bleomycin or 50μM cisplatin. Incubation time was 3 hrs. Root tip spreads and immunostaining was subsequently performed as described earlier in Friesner et al [130]. γH2AX immunostaining was conducted with a rabbit anti-γH2AX antibody (1:600), provided by Dr. Charles White, and a goat Alexa Fluor488 anti-rabbit antibody (Life Technologies, Carlsbad, CA, USA) was used as secondary antibody in a 1:300 dilution. For the observation of RAD51, we used a rat anti‐RAD51 antibody, provided by Dr. Peter Schlögelhofer, in a 1:500 dilution together with a Cy3 anti‐rat antibody (Thermo Fisher Scientific; Cat.# A‐10522) at 1:300. Imaging was done with a Leica TCS SP8 inverted confocal microscope at 40X magnification. The excitation light for the fluorophores was emitted by a diode 405 nm laser, an argon laser at 488 nm and a DPSS laser (561 nm).
RNA was extracted from 10-day-old Arabidopsis seedlings or inflorescence material using the RNeasy Plant Mini Kit from Qiagene according to the instructions of the manufacturer. cDNA synthesis was performed using a Transcriptor First-Strand cDNA Synthesis kit for RT-PCR according to the manufacturer’s instructions (Roche). The cDNA produced was used in semi-quantitative PCR experiments to test for presence of mRNA. Quantitative PCR was performed with a Roche LightCycler 480 SYBR Green I Master with 0.5 μM specific primers and 0.1 μg of first-strand cDNAs. PCR reactions were conducted with the LightCycler 480 Real-Time PCR System (Roche) under the following conditions: 95°C for 5 min; 45 cycles of 95°C for 10 sec, 60°C for 10 sec and 72°C for 15 sec. Cq calling was done using the second derivative maximum method. Target-specific efficiencies were calculated as the mean of all reaction-specific efficiencies for a given target. Reaction-specific efficiencies were deduced using LinRegPCR 2015.2 [131,132]. Data were quality-controlled, normalized against at least three reference genes, and statistically evaluated using qbasePLUS 3.0 [133]. Primers used for genotyping, semi-quantitative RT-PCR and qRT-PCR are listed in S6 Table.
Sequence data for genes characterized in this article can be found in the EMBL/GenBank data libraries under accession numbers AT3G12280 (RBR1), AT3G20490 (KNO1), AT1G04650 (FLIP), AT5G46740 (UBP21) and AT2G45460.
The RBR1-ChIP-seq data generated in this publication have been deposited in NCBI's Gene Expression Omnibus [77] and are accessible through GEO Series accession number GSE108741.
The mutant lines used in this study were provided by the Nottingham Arabidopsis Stock Centre (NASC [134]) and the Versailles Arabidopsis Stock Center (http://publiclines.versailles.inra.fr). They are part of the SALK line collection [135] or the FLAG line collection (http://publiclines.versailles.inra.fr), respectively.
AT3G20490 (SALK_023330C is kno1-1, SALK_023527 is kno1-2)
AT1G04650 (SALK_037387C is flip-2, SALK_119229C is flip-3)–note, that we renamed our mutant lines to be congruent with the numbering used by Fernandes et al. [84].
AT5G46740 (SALK_205928C is ubp21-1, SALK_201584C is ubp21-2)
AT2G45460 (SALK_142111C is at2g45460-1, FLAG_519A08/EHGTV204T3 is at2g45460-2).
|
10.1371/journal.pgen.1002991 | Muscle-Specific Splicing Factors ASD-2 and SUP-12 Cooperatively Switch Alternative Pre-mRNA Processing Patterns of the ADF/Cofilin Gene in Caenorhabditis elegans | Pre–mRNAs are often processed in complex patterns in tissue-specific manners to produce a variety of protein isoforms from single genes. However, mechanisms orchestrating the processing of the entire transcript are not well understood. Muscle-specific alternative pre–mRNA processing of the unc-60 gene in Caenorhabditis elegans, encoding two tissue-specific isoforms of ADF/cofilin with distinct biochemical properties in regulating actin organization, provides an excellent in vivo model of complex and tissue-specific pre–mRNA processing; it consists of a single first exon and two separate series of downstream exons. Here we visualize the complex muscle-specific processing pattern of the unc-60 pre–mRNA with asymmetric fluorescence reporter minigenes. By disrupting juxtaposed CUAAC repeats and UGUGUG stretch in intron 1A, we demonstrate that these elements are required for retaining intron 1A, as well as for switching the processing patterns of the entire pre–mRNA from non-muscle-type to muscle-type. Mutations in genes encoding muscle-specific RNA–binding proteins ASD-2 and SUP-12 turned the colour of the unc-60 reporter worms. ASD-2 and SUP-12 proteins specifically and cooperatively bind to CUAAC repeats and UGUGUG stretch in intron 1A, respectively, to form a ternary complex in vitro. Immunohistochemical staining and RT–PCR analyses demonstrate that ASD-2 and SUP-12 are also required for switching the processing patterns of the endogenous unc-60 pre-mRNA from UNC-60A to UNC-60B in muscles. Furthermore, systematic analyses of partially spliced RNAs reveal the actual orders of intron removal for distinct mRNA isoforms. Taken together, our results demonstrate that muscle-specific splicing factors ASD-2 and SUP-12 cooperatively promote muscle-specific processing of the unc-60 gene, and provide insight into the mechanisms of complex pre-mRNA processing; combinatorial regulation of a single splice site by two tissue-specific splicing regulators determines the binary fate of the entire transcript.
| Muscle is a specialized organ with specialized contractile apparatuses. A number of genes encoding contractile apparatus-related proteins undergo muscle-specific pre–mRNA processing. However, the molecular mechanisms and consequences of muscle-specific alternative pre–mRNA processing remain largely unknown. In this study, we reveal regulation mechanisms of pre–mRNA processing of the unc-60 gene locus, encoding two tissue-specific isoforms of ADF/cofilin in C. elegans. The unc-60A and unc-60B genes share only the first exon, and UNC-60B protein is specifically expressed in muscle. We visualize the tissue-specific processing patterns of the unc-60 pre–mRNA with green and red fluorescent proteins in living worms. We provide genetic, biochemical, and immunohistochemical evidence that muscle-specific RNA–binding proteins ASD-2 and SUP-12 cooperatively bind to specific motifs in intron 1A to retain intron 1A, which leads to skipping of exon 2A through 5A and splicing between exon 1 and 2B. Consistently, disruption of the splicing factors leads to expression of UNC-60A in muscle and suppresses paralysis of an unc-60B-specific mutant. Our study raises a model of step-by-step execution of complex co-transcriptional pre–mRNA processing and provides insight into the fate decision of the entire transcript.
| Alternative pre-mRNA processing is a major way to produce a number of different mRNAs and proteins from one gene [1], [2]. Recent transcriptome analyses by deep sequencing estimated that more than 90% of human multi-exon genes undergo alternative processing and most alternative processing events are regulated in tissue-specific manners [3], [4]. These alternative pre-mRNA processing events are classified into seven elementary events: cassette exons, mutually exclusive exons, alternative 5′ splice sites, alternative 3′ splice sites, intron retention, alternative first exons and alternative polyadenylation sites [5], [6]. A variety of tissue-specific splicing factors and RNA secondary structures have been shown to regulate these elementary events in the minigene context or by knockdown and/or knockout experiments [7], [8], [9]. However, pre-mRNA processing in multicellular organisms is often complex due to various combinations of the elementary events and the molecular mechanisms by which tissue-specific factors regulate such complex alternative processing of the entire gene in vivo remain to be elucidated.
Muscle is one of tissues in which many genes undergo tissue-specific pre-mRNA processing [3], [4]. A number of muscle-specific protein isoforms are expressed by alternative pre-mRNA splicing and play adapted roles depending on the specific properties of muscle fiber types [10], [11], [12]. For instance, tissue-specific splicing generates functionally distinct isoforms of tropomyosin [13] and troponin T [14]. Global analyses of splicing patterns during development of heart and skeletal muscle revealed that splicing transitions of these genes occur at specific times [15], [16]. Bioinformatics analyses have revealed candidate cis-elements regulating muscle-specific splicing patterns [16], [17], [18]. In addition, several trans-acting splicing factors are known to regulate muscle-specific alternative splicing. These include muscleblind-like (MBNL) [19], RBFOX family [20], CUGBP and ETR-3 like factor (CELF) family [21], polypyrimidine tract binding protein (PTB) [22] and hnRNP H [23]. However, how multiple splicing factors coordinate regulation of specific splicing events is poorly understood.
Alternative processing of the uncoordinated (unc)-60 gene in Caenorhabditis elegans provides an excellent model of muscle-specific and complex pre-mRNA processing of genes related to contractile apparatuses. The unc-60 gene encodes two homologous proteins, UNC-60A and UNC-60B [24], which are members of the actin depolymerising factor (ADF)/cofilin family of actin-binding proteins that promote rapid turnover of the actin cytoskeleton [25]. The unc-60 gene consists of a common first exon and two separate series of downstream exons, exons 2A through 5A for UNC-60A and exons 2B through 5B for UNC-60B (Figure 1A). Alternative choices of exons 2A–5A or exons 2B–5B result in tissue-specific expression patterns of the two ADF/cofilin isoforms: UNC-60A protein is expressed in most embryonic cells throughout embryogenesis and predominantly expressed in non-muscle tissues, while UNC-60B protein is mainly detected in body wall muscles [26]. Our biochemical and genetic studies demonstrated that the UNC-60 isoforms have distinct biochemical properties in the regulation of actin dynamics [27], [28] and different in vivo functions during development and in muscle organization [26], [29].
The structure of the unc-60 gene and its expression patterns raise a question as to how the first exon and the two series of downstream exons are properly spliced in a tissue-specific manner. We previously reported genetic evidence that an RNA-binding protein SUP-12, which has only one RNA-recognition motif (RRM), is required for generation of muscle-specific UNC-60B mRNA [30]. However, the molecular mechanism by which SUP-12 regulates the muscle-specific alternative processing of the unc-60 gene remains unclear. In this study, we applied a transgenic alternative splicing reporter system [31], [32], [33] to visualize muscle-specific alternative processing patterns of the unc-60 pre-mRNA. We demonstrate that repression of excision of the intron between exon 1 and exon 2A is the fate-determining event for the unc-60 transcript. We provide genetic and biochemical evidence that SUP-12 and another muscle-specific splicing regulator Alternative-Splicing-Defective-2 (ASD-2), a member of the signal transduction and activation of RNA (STAR) family of RNA-binding proteins [34], cooperatively repress excision of the first intron through specific binding to the intron. Our data provide in vivo evidence that combinatorial regulation of a single splice site by two tissue-specific splicing regulators determine the binary fate of the entire transcript that can potentially be processed into two alternative isoforms.
In order to visualize the binary processing patterns of the unc-60 transcript in vivo, we intended to construct a pair of fluorescence alternative processing reporter minigenes. If the intron between exon 1 and exon 2A (hereafter called intron 1A) is excised prior to selection of exon 2B, it would be impossible to produce UNC-60B mRNA. We therefore assumed that excision of intron 1A should be repressed until exon 2B is transcribed in tissues where UNC-60B is expressed. On the basis of the assumption, we constructed an asymmetric pair of reporter minigenes, unc-60E1-E2A-RFP and unc-60E1-E3B-GFP. The unc-60E1-E2A-RFP cassette, carrying unc-60 genomic DNA fragment from exon 1 through exon 2A (Figure 1B, top panel), was designed to monitor excision of intron 1A via expression of RFP-fusion protein (UNC-60A-RFP). If intron 1A is retained (UNC-60-I1A), RFP would not be expressed due to a premature termination codon in intron 1A (Figure 1B, top panel). On the other hand, the unc-60E1-E3B-GFP cassette, carrying unc-60 genomic DNA fragment from exon 1 through exon 3B (Figure 1B, bottom panel), was designed to monitor UNC-60B-type processing via expression of GFP-fusion protein (UNC-60B-GFP). An intact UNC-60A isoform (UNC-60A-full) would be expressed in tissues where UNC-60A is expressed (Figure 1B, bottom panel).
We successfully visualized the alternative expression of the UNC-60 isoforms with the unc-60 reporter cassettes under the control of the unc-51 promoter that directs expression in a broad variety of tissues [35], [36]. As expected, the expression patterns of UNC-60A-RFP and UNC-60B-GFP varied between muscle and non-muscle tissues (Figure 1C, 1D). Non-muscle tissues including the nervous system and intestine expressed UNC-60A-RFP (Figure 1C, 1D, left panels), and muscle tissues such as body wall muscles and pharyngeal muscles expressed UNC-60B-GFP (Figure 1C, 1D, right panels). This result is consistent with our previous immunohistochemical studies showing that UNC-60A and UNC-60B proteins were detected in non-muscle and muscle tissues, respectively [26], [37]. We checked splicing patterns of mRNAs derived from the unc-60 reporter cassettes by cloning and sequencing reverse transcription-polymerase chain reaction (RT-PCR) products, and confirmed that the four mRNA isoforms schematically shown in Figure 1B were actually generated in the transgenic worms (data not shown).
To focus on the muscle-specific control of the unc-60 processing, we utilized myo-3 promoter to drive expression of the unc-60 reporter specifically in body wall muscles. Transgenic worms with an integrated transgene allele ybIs1831 [myo-3::unc-60E1-E2A-RFP myo-3::unc-60E1-E3B-GFP] predominantly expressed UNC-60B-GFP in body wall muscles (Figure 1E), consistent with the unc-60 reporter expression in muscles (Figure 1C, 1D). We therefore used the myo-3 promoter for further analyses described below.
To test whether muscle-specific repression of UNC-60A-RFP and expression of UNC-60B-GFP from the unc-60 reporter are similarly regulated by a muscle-specific splicing regulator SUP-12 to the endogenous mRNAs for UNC-60A and UNC-60B isoforms [30], we crossed the reporter allele ybIs1831 with a presumptive null allele sup-12 (yb1253) [38]. As expected, the reporter worms clearly turned the colour from Green to Red in the sup-12 background (Figure 2A), confirming that SUP-12 is required for the muscle-specific expression profile of the unc-60 reporter.
In a previous study, we identified SUP-12 as a co-regulator of mutually exclusive exons of a fibroblast growth factor receptor gene egg-laying-defective (egl)-15 [38]. In the case of repression of egl-15 exon 5B, SUP-12 functions as a muscle-specific partner of the Fox-1 family proteins ASD-1 and FOX-1 [31], [38]. We therefore speculated that other regulator(s) may also be involved in the muscle-specific regulation of unc-60. As direct interaction between SUP-12 and ASD-1 in a yeast two-hybrid system had been reported in a worm interactome study [39], we screened for a putative co-regulator of the unc-60 reporter by knocking down genes encoding possible SUP-12-interactors ASD-1, ASD-2, ETR-1, MEC-8, R02F2.5 and W02A11.3, deposited in the database (http://interactome.dfci.harvard.edu/). We performed RNA interference (RNAi) by feeding the reporter worms with bacterial clones targeting the six genes, and found that knockdown of asd-2 led to expression of UNC-60A-RFP (Figure S1).
We previously identified ASD-2, an RNA-binding protein belonging to the STAR family, as a regulator of muscle-specific and developmentally regulated alternative splicing of a collagen gene let-2 [32], [33]. The asd-2 gene has alternative first exons and a non-lethal allele asd-2 (yb1540) has a nonsense mutation in the asd-2b-specific first exon (Figure 2B), which is used in body wall muscles and pharyngeal muscles [32]. The unc-60 reporter worms exhibited weak Red phenotype in the asd-2 (yb1540) background (Figure 2C) and body wall muscle-specific expression of ASD-2b cDNA rescued the colour phenotype (Figure 2D), confirming that asd-2b is involved in the muscle-specific regulation of the unc-60 reporter. To investigate subcellular localization of ASD-2, we raised polyclonal antibodies against recombinant full-length ASD-2b protein and stained wild-type and asd-2 (yb1540) worms with a purified immunoglobulin G (IgG) fraction (Figure 2E, 2F). Nuclei of body wall muscles, which are aligned along the dorsal and ventral periphery, are stained in the wild type (Figure 2E) and not in asd-2 mutant (Figure 2F). In Western blotting, the same antibody detected a major band with an apparent molecular weight of 56 kDa in wild-type and not in asd-2 (yb1540) lysate (Figure 2G). These results indicated that ASD-2b is the major isoform and is predominantly localized in the nuclei of body wall muscles. RNAi by micro-injecting double-stranded RNA (dsRNA), a more effective method than feeding dsRNA-expressing bacteria, led to a stronger Red phenotype (Figure 2C), suggesting trace remaining activity of ASD-2 in asd-2 (yb1540) mutant.
To confirm splicing patterns of mRNAs derived from the unc-60 reporter minigenes in body wall muscles, we performed RT-PCR analysis with minigene-specific primer sets (Figure 2H). In the wild-type background, UNC-60B-type mRNA, UNC-60B-GFP, was predominantly generated from unc-60E1-E3B-GFP (Figure 2H, middle panel, lane 1). A transcript derived from unc-60E1-E2A-RFP was almost undetectable (Figure 2H, top panel, lane 1), presumably due to rapid degradation of a non-productive mRNA isoform, UNC-60-I1A, by nonsense-mediated mRNA decay (NMD) [40]. On the other hand, the amount of UNC-60B-GFP was reduced and UNC-60A-type mRNAs, UNC60A-RFP and UNC-60A-full, were detected in asd-2 and sup-12 mutants (Figure 2H, lanes 2 and 3), consistent with their colour phenotypes shown in Figure 2C and 2A, respectively. These results confirmed that both SUP-12 and ASD-2 are responsible for switching the processing patterns of the unc-60 reporter from UNC-60A-type to UNC-60B-type in body wall muscles.
The experiments described above indicate that each of the unc-60 reporter minigenes, even the shorter one, carries sufficient regulatory elements for ASD-2 and SUP-12 to switch from non-muscle-type to muscle-type processing. As regulatory elements for alternative splicing are often evolutionarily conserved in introns among nematodes [31], [32], [38], [41], we searched for conserved stretches in unc-60 intron 1A in the Caenorhabditis genus. Alignment of nucleotide sequences available in WormBase (http://www.wormbase.org/) revealed that CTAAC repeats and TGTGTG stretch are highly conserved just upstream of the splice acceptor site (Figure 3A). To evaluate the roles of these elements in the muscle-specific processing of the unc-60 reporter, we constructed two pairs of modified unc-60 reporter minigenes M1 and M2. In the M1 pair, CTAAC repeats were mutagenized to CAAAC (Figure 3B). In the M2 pair, TGTGTG were mutagenized to TATATA (Figure 3B). Disruption of either of the two elements resulted in Red phenotype (Figure 3C), phenocopying sup-12 mutant (Figure 2A) and asd-2 (RNAi) worms (Figure 2C). RT-PCR analysis of mRNAs derived from the mutant reporters revealed that both M1 and M2 mutations increased production of UNC-60A-RFP (Figure 3D, top panel) and decreased expression of UNC-60B-GFP (Figure 3D, bottom panel), consistent with their colour phenotypes. These results confirmed that the colour phenotypes observed with the mutant reporters are due to altered patterns of pre-mRNA processing. We concluded that both CUAAC repeats and UGUGUG stretch are required for muscle-specific repression of intron 1A excision. Notably, expression of UNC-60A-full mRNA from the M1 and M2 mutants of unc-60E1-E3B-GFP minigene increased compared to the wild-type minigene (Figure 3D), indicating that the repression of intron 1A excision via CUAAC repeats and UGUGUG stretch is a crucial event to switch the processing patterns of the entire unc-60E1-E3B-GFP minigene from UNC-60A-type to UNC-60B-type.
To confirm direct and specific binding of ASD-2 and SUP-12 to the cis-elements in unc-60 intron 1A in vitro, we prepared radiolabelled RNA probes containing the intact sequence (WT) or those with mutations as in the mutant reporters (M1 and M2) (Figure 4A) and recombinant full-length ASD-2b and full-length SUP-12 proteins (Figure 4B) to perform electrophoretic mobility shift assays (EMSAs) (Figure 4C, 4D). Recombinant ASD-2b protein shifted the mobility of WT (Figure 4C, lanes 1–6) and M2 (Figure 4D, lanes 18–22) probes in a dose-dependent manner and not of M1 probe (Figure 4D, lanes 1–5), demonstrating direct and specific binding of ASD-2b to CUAAC repeats. On the other hand, recombinant SUP-12 protein shifted the mobility of WT (Figure 4C, lanes 13–18) and M1 (Figure 4D, lanes 6–9) probes to a similar extent in a dose-dependent manner and less efficiently of M2 probe (Figure 4D, lanes 23–26) to a less extent, demonstrating direct and specific binding of SUP-12 to UGUGUG stretch. The result also indicated that SUP-12 could bind to other site(s) in the probes with a lower affinity.
We next asked whether ASD-2b and SUP-12 cooperatively bind to unc-60 intron 1A RNA. We analyzed supershifts of the mobility of the unc-60 intron 1A probes by the combination of ASD-2b and SUP-12 in EMSAs (Figure 4C, 4D). ASD-2b efficiently supershifted the mobility of WT probe at lower concentrations in the presence of SUP-12 (Figure 4C, lanes 7–12) compared to ASD-2b alone (lanes 1–6). In the same way, SUP-12 supershifted the mobility of WT probe at lower concentrations in the presence of ASD-2b (lanes 19–24) compared to SUP-12 alone (lanes 13–18). These results indicated that ASD-2b and SUP-12 cooperatively form a stable ASD-2b/SUP-12/RNA ternary complex with unc-60 intron 1A RNA. ASD-2b failed to supershift the mobility of M1 probe (Figure 4D, lanes 10–17), indicating that CUAAC repeats are essential for the ternary complex formation. SUP-12 less efficiently supershifted the mobility of M2 probe (Figure 4D, lanes 31–34) compared to WT probe (Figure 4C, lanes 21–24) in the presence of ASD-2b, indicating that UGUGUG stretch is involved in the ternary complex formation.
We finally asked whether ASD-2b and SUP-12 can preform a complex in the absence of unc-60 intron 1A by pull-down experiments (Figure 4E). Glutathione-S-transferase (GST)-fused full-length ASD-2b protein pulled down a substantial amount of recombinant full-length SUP-12 protein in the absence of target RNAs (Figure 4E, lane 2) and wild-type (WT) unc-60 intron 1A (unc-60-I1A) RNA enhanced the pull-down efficiency in a dose-dependent manner (lanes 3, 4). On the other hand, GST-fused monomeric RFP (mRFP) protein failed to pull down SUP-12 protein even in the presence of unc-60-I1A RNA (lanes 10–13), demonstrating specific interaction between ASD-2b and SUP-12. M1 and M2 mutant unc-60-I1A RNAs less effectively enhanced the interaction between ASD-2b and SUP-12 (lanes 5–8), consistent with their weaker or no ability to form a ternary complex (Figure 4D). We therefore concluded that ASD-2b and SUP-12 can weakly interact with each other and that unc-60 intron 1A RNA promotes the formation of the stable ASD-2b/SUP-12/RNA ternary complex by providing juxtaposed CUAAC repeats and UGUGUG stretch that are specifically recognized by ASD-2b and SUP-12, respectively.
We examined whether ASD-2 regulates muscle-specific pre-mRNA processing of the endogenous unc-60 gene. We have demonstrated that ASD-2 and SUP-12 cooperatively switch alternative processing of the unc-60 reporter from UNC-60A-type to UNC-60B-type in body wall muscles. If this model can be applied to the endogenous unc-60 gene, worms depleted of asd-2 function should ectopically express UNC-60A in place of UNC-60B in body wall muscles. Indeed, RT-PCR analysis of the endogenous UNC-60 mRNAs revealed that relative amount of UNC-60B mRNA was decreased in asd-2 (yb1540); asd-2 (RNAi) worms (Figure S2). To further test the splicing change in body wall muscles, we investigated expression of UNC-60A protein by immunohistochemistry (Figure 5A, 5B). In wild-type worms, UNC-60A was undetectable in body wall muscles (Figure 5A, encircled) but was detected in other tissues (Figure 5A, left panel). Knockdown of the asd-2 gene resulted in ectopic expression of UNC-60A in body wall muscles (Figure 5B, encircled), confirming that ASD-2 determines the processing patterns of the endogenous unc-60 gene in body wall muscles.
Our previous work demonstrated that sup-12 mutation strongly suppressed structural defects of body wall muscles and paralysis of UNC-60B-specific mutant, unc-60B (su158) [30]. The deletion allele su158 lacks exons 3B and 4B (Figure 1A), and suppression of the phenotypes by sup-12 mutation was likely due to ectopic expression of UNC-60A [30]. We therefore investigated whether knockdown of the asd-2 gene also suppresses phenotypes of unc-60B (su158) mutant. Wild-type worms exhibited sinusoidal locomotion (Figure 5C, left panel), and actin filaments were organized in a striated pattern (Figure 5C, right panel). On the other hand, unc-60B (su158) worms were almost paralyzed (Figure 5D, left panel) with severe disorganization of actin filaments (Figure 5D, right panel). We found that asd-2 (yb1540); unc-60B (su158) double mutant slightly restored motility and actin filament organization (Figure 5E). Since asd-2(RNAi) worms showed a severer colour phenotype than asd-2(yb1540) allele (Figure 2C), we further knocked down remaining activity of asd-2 by RNAi. As expected, asd-2 (yb1540); unc-60B (su158); asd-2 (RNAi) worms restored sinusoidal locomotion (Figure 5F, left panel) and actin filament organization was greatly improved (Figure 5F, right panel). We confirmed by immunohistochemistry that asd-2 (yb1540) mutation and/or asd-2 (RNAi) resulted in ectopic expression of UNC-60A in body wall muscles in the unc-60B (su158) background (Figure S3). Transgenic expression of UNC-60A (Figure 5G) as well as UNC-60B (Figure 5H) in body wall muscles restored sinusoidal locomotion of unc-60B (su158) mutant, indicating that UNC-60A can exert, at least in part, functions of muscle-specific UNC-60B isoform and that possible splicing change in other genes are not required for the phenotype suppression. These observations demonstrate that ASD-2 is a bona fide regulator of the muscle-specific pre-mRNA processing of the endogenous unc-60 gene as well as SUP-12.
Finally, we analyzed splicing patterns of mature and partially spliced RNAs from the endogenous unc-60 gene (Figure 6). For this experiment, we used wild-type and sup-12 (yb1253) worms because asd-2 (yb1540) mutation exhibited weaker effect on the unc-60 reporter. In the wild type, mature UNC-60A and UNC-60B mRNAs were almost equally detected (Figure 6A, lane 3), while the latter was hardly detectable in sup-12 mutant (lane 4), consistent with the result with the reporter (Figure 2H) and our previous study [30]. To analyze processing patterns of UNC-60B RNAs in body wall muscles, we amplified partially spliced RNAs carrying intron 2B, 3B or 4B with a forward primer in exon 1 and intronic reverse primers (Figure 6B). Partially spliced RNAs committed to UNC-60B, in which exon 1 was connected to exon 2B, were detected in the wild type (all panels, lane 3, bands 2 and 3) but were undetected in sup-12 mutant (lane 4), consistent with the result shown in Figure 6A. These results indicated that SUP-12 is required for proper splicing between exon 1 and exon 2B in muscles. In sup-12 mutant, all the introns, including intron 1A, were excised in the only detected RNAs (Figure 6B, all panels, lane 4, band 1), while in the wild type, intron 1A is retained in the longest detected RNAs (all panels, lane 3, band 1), indicating that SUP-12 represses excision of intron 1A.
We next analyzed partially spliced RNAs from the UNC-60A region (Figure 6C, 6D). Although the detected RNAs derived from this region were mixture of those in muscles and in non-muscle tissues, we assumed that differences in their relative amounts could be attributed to functions of SUP-12 in muscles. With a forward primer in intron 1A and a reverse primer in exon 5A (Figure 6C), we detected eight RNA species in sup-12 mutant (lane 4, bands 1–6). These RNAs were all the theoretical intermediates in the UNC-60A processing. In the wild type (lane 3), two of the RNAs (bands 3 and 6) predominated, suggesting that SUP-12 represses their production. In these RNAs, intron 1A alone (band 6) or introns 1A and 2A were retained (band 3), supporting the idea that SUP-12 represses excision of intron 1A, and weakly of intron 2A, even after introns 3A and 4A are excised. We then analyzed the partially spliced RNAs with the forward primer in exon 1 and intronic reverse primers in introns 2A, 3A and 4A (Figure 6D). All the two (top panel, band 1–2), four (middle panel, bands 1–4) and eight (bottom panel, bands 1–7) theoretical intermediate RNA species were detected in sup-12 mutant (lane 4), and relative amounts of the partially spliced RNAs to the pre-mRNAs (band 1) in the wild type (lane 3) and sup-12 mutant (lane 4) were in good accordance with the idea that excision of introns 1A and 2A is facilitated in the absence of SUP-12. All these analyses of the partially spliced RNAs supported the model that SUP-12 represses excision of intron 1A to preserve exon 1 until exon 2B is transcribed in muscles.
In this study, we have provided genetic and biochemical analyses of the mechanisms for regulation of the muscle-specific alternative processing of the unc-60 pre-mRNA. Figure 7 illustrates models of the pre-mRNA processing deduced from this study. In non-muscle tissues (Figure 7A), intron 1A and the other introns are excised during or after transcription and UNC-60A mRNA is generated. The order of intron removal is not strictly regulated as suggested by the presence of all the theoretical partially spliced RNAs (Figure 6C, 6D). In muscles (Figure 7B), ASD-2b and SUP-12 cooperatively bind to CUAAC repeats and UGUGUG stretch, respectively, in intron 1A to repress excision of intron 1A and weakly of intron 2A during transcription of the UNC-60A region. When UNC-60B-specific region is being transcribed, exon 1 is readily spliced to exon 2B, and introns 3B and 4B are also readily removed in the order of transcription (Figure 6B). Introns 3A and 4A are properly and rapidly excised during the UNC-60B processing (Figure 6C) likely due to their small sizes (53 nt and 60 nt, respectively). This may explain why exon 1 is not aberrantly spliced to exons 3A or 4A but is exclusively spliced to exon 2B to form UNC-60B mRNA. Regulation of tissue-specific alternative polyadenylation may also be involved in the fate-decision of the unc-60 transcript, although the results demonstrated above did not provide conclusive evidence that ASD-2 and/or SUP-12 regulate muscle-specific repression of the polyadenylation site for UNC-60A mRNA.
We have demonstrated that ASD-2 and SUP-12 cooperatively represses the 3′-splice site and not the 5′-splice site of intron 1A. Although C. elegans does not have a recognizable branch point consensus or a polypyrimidine tract [42], a putative branch site for intron 1A is the A at position -19, between CUAAC repeats and UGUGUG stretch (Figure 3A). This A is the first A upstream from the 3′ splice site and is close to the positions where the putative branch site A is frequently found [43]. It is therefore reasonable to suggest that formation of ASD-2b/SUP-12/RNA ternary complex sterically hinders U2 snRNP auxiliary factor (U2AF) bound to the 3′-splice site from recruiting U2 snRNP to the branch site. The situation is quite similar to muscle-specific repression of egl-15 exon 5B, where the Fox-1 family proteins and SUP-12 cooperatively bind to juxtaposed cis-elements overlapping a putative branch site [20], [38]. Recent microarray analyses of alternatively spliced exons in splicing factor mutants including sup-12 identified many other splicing events affected by multiple splicing factors [44]. Combinatorial regulation by multiple splicing factors may be the common feature in tissue-specific alternative pre-mRNA processing in C. elegans.
ASD-2 ortholog in Drosophila, Held out wings (How) [45], [46], [47], and that in zebrafish, Quaking A (QkA) [48], are known to be required for muscle development or activity by mutant analyses. Vertebrate orthologs of SUP-12, known as SEB-4 or RBM24, are also expressed in muscle tissues and have recently been shown to be involved in myogenic differentiation by knockdown experiments [49], [50], [51], [52], [53]. However, the target events that these orthologs regulate in muscles remain almost unclear. Considering the highly conserved amino acid sequences and their expression patterns, it is likely that the orthologs of ASD-2 and SUP-12 regulate alternative pre-mRNA processing to produce muscle-specific protein isoforms in higher organisms.
In this study, we have presented a model of complex alternative pre-mRNA processing of a gene generating two almost distinct mRNAs. An important aspect of this study is the successful application of a dichromatic fluorescence reporter system to analyze the complex alternative pre-mRNA processing. The asymmetric pair of fluorescence reporter minigenes utilized in this study offers an alternative option for visualizing complex processing patterns besides symmetric pairs of minigenes applied to mutually exclusive exons and cassette exons [32], [33]. Another example of evolutionarily conserved genes with a structure similar to the unc-60 gene is the cholinergic gene locus; genes encoding choline acetyltransferase (ChAT) and vesicular acetylcholine transporter (VAChT) share the common first exon, and the other exon(s) for VAChT reside in the first intron of the ChAT gene in mammals [54], Drosophila [55] and C. elegans [56]. The regulation mechanisms presented here would provide insight into the regulation of this kind of genes.
We demonstrated that ectopically expressed UNC-60A can compensate for the function of UNC-60B in sarcomeric actin organization in body wall muscles of unc-60B mutant. However, both UNC-60A and UNC-60B have characteristic actin-regulatory activities of ADF/cofilin in vitro with some quantitative differences [27], [28], [29]; UNC-60A has strong actin-monomer sequestering and only weak actin-filament severing activities, while UNC-60B has no actin-monomer sequestering and strong actin-filament severing activities. Although UNC-60A can compensate for the function of UNC-60B in body wall muscles, sarcomeric actin filaments in UNC-60A-complemented unc-60B mutant muscles still exhibit minor disorganization (unpublished data), suggesting that UNC-60B is a more suitable isoform. On the other hand, UNC-60B cannot compensate for the function of UNC-60A in the gonadal myoepithelial sheath [29]. This work and our previous works demonstrated that UNC-60A and UNC-60B are specifically adapted for functions in non-muscle and muscle cells, respectively, emphasizing that precise expression of appropriate ADF/cofilin isoforms, unravelled in this study, is important for development of tissue-specific actin-cytoskeletal structures [26], [29].
To construct the unc-60E1-E2A-RFP and unc-60E1-E3B-GFP cassettes, unc-60 genomic fragments spanning from exon 1 through 2A and exon 1 through 3B, respectively, were amplified from N2 genomic DNA and cloned into Gateway Entry vectors (Invitrogen) carrying either mRFP1 [57] or EGFP (Clontech) cDNA by using In-Fusion system (BD Biosciences). M1 and M2 mutations were introduced by mutagenesis with Quickchange II (Stratagene). Expression vectors were constructed by homologous recombination between the Entry vectors and Destination vectors [31], [33] with LR Clonase II (Invitrogen). Sequences of the primers used in plasmid construction are available in Table S1.
Worms were cultured following standard methods. Transgenic lines were prepared essentially as described [33] using lin-15 (n765) as a host or pmyo-2-mRFP as a marker. Integrant lines were generated by ultraviolet light irradiation as described previously [33], [58]. Images of fluorescence reporter worms were captured using a fluorescence stereoscope (MZ16FA, Leica) with a dual and-pass filter GFP/DsRed equipped with a colour, cooled CCD camera (DP71, Olympus) or a confocal microscope (Fluoview FV500, Olympus) and processed with Metamorph (Molecular Devices) or Photoshop (Adobe).
RNAi experiments by feeding were performed essentially as described [59]. Briefly, L4 hermaphrodites were transferred to agar plates seeded with bacteria expressing dsRNAs of target genes and their progeny were scored for colour and behavioural phenotypes or used for staining. For RNAi experiment by micro-injection, sense and anti-sense asd-2 RNAs were prepared as described preciously [32] and were annealed at room temperature and 1–5 µg/µl dsRNA was injected into the gonad of young adult hermaphrodites. Injected worms were cultured at 20°C and the colour phenotype of their progeny was evaluated.
Total RNAs were extracted from worms by using RNeasy Mini kit (Qiagen) and DNase I (Promega). RNAs (300–500 ng) were reverse transcribed using random hexamers and Superscript II (Invitrogen) according to manufacturer's protocol. PCR was performed essentially as described previously [31], [33]. For amplification of partially spliced RNAs, total RNAs were reverse transcribed with PrimeScript II and random hexamers (Takara), and amplified with BIOTAQ (Bioline) and analyzed by using BioAnalyzer (Agilent). Sequences of the RT-PCR products were confirmed either by direct sequencing or by cloning and sequencing. Sequences of the primers used in the RT-PCR assays are available in Table S2.
Denatured His-tagged full-length ASD-2b for immunization was purified from denatured bacterial lysate by using Ni-NTA agarose (QIAGEN). Cold-shock inducible expression vectors for His-GST-fused full-length ASD-2b and mRFP1 and FLAG-tagged full-length SUP-12 were constructed by using Destination vectors pDEST-Cold-GST and pDEST-Cold-FLAG (H.K.), respectively. GST-ASD-2b and FLAG-SUP-12 were purified by using Glutathione Sepharose 4B (GE Healthcare) and Anti-FLAG M2 Magnetic Beads (Sigma), respectively, and dialyzed against RNA binding buffer (see below). Purified proteins were separated by standard SDS-PAGE and stained with SimplyBlue SafeStain (Invitrogen).
Rabbit polyclonal anti-ASD-2b antiserum was generated with denatured recombinant His-ASD-2b protein by Operon Biotechnologies (Tokyo, Japan). IgG fraction (TD0135-02) was prepared from the antiserum by Medical & Biological Laboratories (Nagoya, Japan). Worm lysates were extracted from synchronized L1 larvae, separated by neutral polyacrylamide gel electrophoresis (NuPAGE, Invitrogen) and transferred to nitrocellulose membrane (Protran BA85, Whatman). Western blotting was performed with 15 µg/ml anti-ASD-2b (TD0135-02) or 1∶40,000-diluted anti-actin monoclonal antibody (Ab-1, Calbiochem) and 1∶1,000-diluted HRP-conjugated anti-rabbit IgG antibody (Pierce) or 1∶10,000-diluted HRP-conjugated anti-mouse IgM antibody (Calbiochem). Chemiluminescence signals (West Dura, Thermo) were detected by using LAS4000 (GE Healthcare).
For staining with anti-ASD-2b, mixed stages of N2 and asd-2 (yb1540) worms were fixed with Bouin's fixative (15∶5∶1 mixture of saturated picric acid, formalin and acetic acid) supplemented with 25% methanol and 1.25% 2-mercaptoethanol for 60 min at room temperature, washed with phosphate-buffered saline (PBS) and permeabilized with 5% 2-mercaptoethanol and 1% Triton X-100 in PBS at 37°C for 30 hours. Fixed worms were treated with blocking buffer (0.5% skim milk and 0.5% bovine serum albumin (BSA) in PBS) for 2 hours at room temperature and stained with 6 µg/ml anti-ASD-2b (TD0135-02) as a primary antibody in blocking buffer for 24 hours at room temperature and then with 2 µg/ml Alexa488-conjugated goat anti-rabbit IgG (Invitrogen) as a secondary antibody together with 1 µg/ml Hoechst 33258 (Hoechst) in blocking buffer for 2 hours at room temperature. Fluorescence images were captured by using a compound microscope (DM6000B, Leica) equipped with a colour, cooled CCD camera (DFC310FX, Leica) or an inverted fluorescence microscope (Nikon TE2000) equipped with a monochrome CCD camera (SPOT RT, Diagnostic Instruments, Inc). Staining with anti-UNC-60A and anti-MyoA were performed as described previously [30]. Actin filaments were visualized by staining with tetramethylrhodamine-phalloidin as described previously [60].
32P-labelled RNA probes were generated by in vitro transcription with [α32P] UTP (Perkin Elmer) and T7 RNA polymerase (Takara). Sequences of template oligo DNAs are available in Table S3. Gel-purified RNA probes alone or with increasing amounts of recombinant protein(s) were incubated in 25 µl of RNA binding buffer (20 mM HEPES-KOH (pH7.9), 150 mM KCl, 5% glycerol, 1% Triton X-100, 1 mM DTT and 0.1 mM PMSF) supplemented with 100 ng/µl E. coli tRNA and 50 ng/µl BSA for 30 min at 20°C. Each sample was separated on a non-denaturing 4% polyacrylamide gel and analyzed with a fluoro-imaging analyzer (FLA-3000G, Fuji Film).
His-GST-fused recombinant full-length ASD-2b and mRFP1 proteins were immobilized on glutathione sepharose 4B beads (GE Healthcare) and incubated with His-SUP-12 in 100 µl of pull-down buffer (20 mM HEPES-KOH (pH7.9), 150 mM KCl, 1% Triton X-100, 1 mM DTT and 0.1 mM PMSF) supplemented with 100 ng/µl E. coli tRNA, 50 ng/µl BSA and 0, 30, 100, or 300 nM of unc-60-I1A RNAs (Operon Biotechnologies) for 30 min at 20°C. The sequences of the unc-60-I1A RNAs: unc-60-I1A-WT, 5′-UUUUUGCCUAACCUAACCUAACCUAUGUGUGCCUGUUUU-3′; unc-60-I1A-M1, 5′-UUUUUGCCAAACCAAACCAAACCUAUGUGUGCCUGUUUU-3′; unc-60-I1-M2, 5′-UUUUUGCCUAACCUAACCUAACCUAUAUAUACCUGUUUU-3′. Beads were washed four times with 1 ml pull-down buffer. Bound proteins were eluted with LDS sample buffer and separated by NuPAGE (Invitrogen). Gels were stained with SimplyBlue SafeStain (Invitrogen) and detected and analyzed by using LAS4000 (GE Healthcare).
|
10.1371/journal.pgen.1005952 | Evolution of Social Insect Polyphenism Facilitated by the Sex Differentiation Cascade | The major transition to eusociality required the evolution of a switch to canalize development into either a reproductive or a helper, the nature of which is currently unknown. Following predictions from the ‘theory of facilitated variation’, we identify sex differentiation pathways as promising candidates because of their pre-adaptation to regulating development of complex phenotypes. We show that conserved core genes, including the juvenile hormone-sensitive master sex differentiation gene doublesex (dsx) and a krüppel homolog 2 (kr-h2) with putative regulatory function, exhibit both sex and morph-specific expression across life stages in the ant Cardiocondyla obscurior. We hypothesize that genes in the sex differentiation cascade evolved perception of alternative input signals for caste differentiation (i.e. environmental or genetic cues), and that their inherent switch-like and epistatic behavior facilitated signal transfer to downstream targets, thus allowing them to control differential development into morphological castes.
| Division of labor into reproductive queens and helper workers in the societies of ants, bees and wasps is achieved by phenotypic plasticity, which allows individuals to embark on discrete developmental trajectories in response to variable signals. These signals can be genetic, epigenetic or environmental, thereby resembling the extreme variation in signals for sex determination across multicellular animals. We show that common developmental pathways downstream of these input signals, including the conserved sex differentiation gene doublesex, regulate sex and caste-specific phenotypic differentiation in the ant species Cardiocondyla obscurior. Many different mechanisms of gene regulation have been implicated in controlling caste-specific development in social insects but these all depend on a higher-level genetic switch. We propose that highly conserved hub genes such as dsx, which can translate variable input signals into large transcription differences using intermediate-level regulators, are tightly linked with the repeated evolutionary transition to eusociality and caste polyphenism.
| The mechanisms underlying the evolution of phenotypic novelty are hotly debated [1–4]. A fundamental question is how small genetic or epigenetic changes can produce a set of simultaneous, complementary phenotypic changes required to generate new adaptive trait combinations. A key prediction of the ‘theory of facilitated variation’ [5] is that regulation acts on evolutionarily conserved switch mechanisms, which then modulate expression of target loci controlling development. This process may facilitate large and complex evolutionary steps because it brings together new combinations of inputs (internal or external stimuli) and outputs (phenotypes) but does not rely on evolution of genes involved in the processes per se. Importantly, the reliance of this mode of evolution on conserved genetic and developmental processes increases the likelihood that the outputs will be functionally integrated and thus non-lethal, similar to the ‘two-legged goat effect’, a striking example of phenotypic accommodation in which developmental robustness allows the animal to ‘adapt’ to a previously unselected bipedal lifestyle [6].
Evolution by facilitated variation may be especially important to the origin of developmental polyphenisms in which organisms develop into two or more discrete forms, since polyphenisms typically result from plastic activity of regulatory genes. Additionally, it is likely that regulatory mechanisms controlling one set of polyphenism are pre-adapted to evolve control over newly evolving polyphenisms, for two reasons. Firstly, such mechanisms’ pre-existing sensitivities to variable cues make it more likely that they will evolve the ability to perceive alternative gradients of novel cues, relative to constitutively expressed genes. Secondly, their downstream target genes already show inter-individual variability in expression, and the organism will thus already have evolved alternative responses to this variability.
Gerhart and Kirschner [5] made predictions about the properties of the “core components” which they hypothesize to be the principal drivers of evolutionary novelty, namely that these components should display both robustness and adaptability, as well as exploratory behavior, state-dependent expression and regulatory compartmentation. The sex differentiation pathways exhibit all these properties, making them prime candidates for facilitating the evolution of new forms of polyphenism. Some components of the sex differentiation pathway (such as the doublesex-mab3 (DM) gene family; [7–9]) are evolutionarily ancient and conserved across diverse metazoa, and thus could potentially be involved in generating novel polyphenism in multiple distantly related taxa. In insects, the sequence of sex determination has been called hourglass-shaped [10], with highly variable input signals and downstream targets, but a small set of conserved core regulatory genes including transformer (tra) and doublesex (dsx). doublesex is alternatively spliced depending on the presence of an active TRA protein, and its sex-specific isoforms act as transcription factors causing sex-specific gene expression and development through their differential effects on multiple downstream targets [11,12].
The two social insect ‘castes’—queens and workers—differ radically from one another in their developmental environment (e.g. nutritional environment) resulting in differences in size, fecundity, behavior and physiology. Ultimately, the evolution of caste polyphenism thus required concerted evolution of environmental input signals and corresponding developmental responses [13]. Eusociality has evolved at least twice within the Hymenoptera [14], but we presently lack a well-evidenced theory of the genetic mechanisms that allowed caste-specific gene expression to originate. There is increasing evidence that the evolution of polyphenism in ants, bees and wasps was achieved primarily through evolution of regulatory genes, rather than gene content or composition [15–17], but the core components involved are largely unknown. Here, we propose that conserved parts of the sex differentiation cascade, including the transcription factor doublesex, evolved sensitivity to new environmental input signals (e.g. nutritional signals), thereby triggering caste-specific gene expression that sends larvae on divergent developmental trajectories. To test this hypothesis, we identified dsx and its female- and male-specific isoforms, and measured their expression across life stages in the four discrete morphs (queens, workers, winged males and wingless males) of the ant Cardiocondyla obscurior. We find that dsx sex-specific isoforms are expressed both sex-specifically and morph-specifically in larvae, pupae and adults. Moreover, ninety other conserved genes with sex-biased expression showed morph-specific expression patterns during larval development, suggesting that co-option of the genes regulating sex differentiation via sex-specific alternative splicing was involved in the origin of morphologically distinct castes.
Queens and workers produced from inter-population crosses were heterozygous for diagnostic microsatellite markers, whereas emerging winged and wingless males as well as one sex mosaic individual expressing both male and female characters exclusively carried the maternal alleles (S1 Table). Although single locus complementary sex determination is unlikely because the species regularly engages in inbreeding [18], C. obscurior appears to use standard haplodiploid sexual reproduction.
The C. obscurior genome [19] has four paralogs containing the DM domain of doublesex (dsx) (pfam00751; Cobs_01393, Cobs_07724, Cobs_09254 and Cobs_18158), representing the ancestral state in holometabolous insects [20]. Sex-specific splice forms are only known from one paralog per species (e.g., in Apis [21] and Nasonia [20]), and the function of the others is unclear. In C. obscurior, only Cobs_01393 was differentially expressed in male and female larval RNAseq data (S2 Table). Moreover, Cobs_01393 had the highest sequence homology to functional dsx in other insects (S1 Fig). Finally, we found that Cobs_01393 was located within ~79 kb of prospero; microsynteny of prospero and dsx is conserved across the Hymenoptera [20]. We thus conclude that Cobs_01393 is the functional paralog of dsx.
We identified the full-length sequence and sex-specific isoforms of the functional paralog of dsx using 3’ rapid amplification of cDNA ends (RACE) (S2 Fig). The first four exons are identical in both isoforms. The DM domain (pfam00751) is located in exon 2 and the dsx dimerization domain (pfam08828) in exon 4. The female-typical isoform dsxF contains one exon specific to dsxF, whereas the male-typical isoform dsxM excludes that exon but includes two others that are absent in dsxF (Fig 1A and S3 Table). This splicing pattern, with a shortened female transcript, has been inferred for the fire ant Solenopsis invicta [22], and matches dsx sex-specific isoforms in Drosophila melanogaster and Apis mellifera, but not Nasonia vitripennis [20]. While the sex-signaling function of dsx is conserved across highly divergent lineages, recent evidence shows that dsx sequence evolves rapidly [23–25], causing substantial inter-specific variation in dsx splicing patterns. A higher level of divergence in dsx compared to other DM domain-containing proteins in our phylogenetic analysis confirms this result (S1 Fig).
We designed primers that spanned the exon boundary of the DM domain-containing exon (to measure the overall expression of both isoforms), as well as primers specific to both isoforms for use in RT-qPCR. We found significantly higher expression of the DM domain in adult males (pooling winged males and wingless “ergatoid” males; WM and EM) compared to females (pooling queens and workers; WO and QU) (nEM = 8, nWM = 8, nWO = 10, nQU = 10; Welch two sample t-test: t25.7 = -8.7, p<0.001, Fig 1B). Expression of the DM-domain was similar in queens and workers (t-test with Benjamini-Hochberg (BH) correction: p = 0.672), but higher in winged males compared to wingless males (p = 0.009).
We then compared expression of dsxF and dsxM across all four morphs in pupae (nEM = 10, nWM = 10, nWO = 9, nQU = 10) and adults (nEM = 7, nWM = 7, nWO = 7, nQU = 7; Fig 1C and 1D). We found morph-specific signatures of expression in both life stages for dsxF (ANOVA: pupae: F(3,35) = 42.33, p<0.001; adults: F(3,24) = 3.75, p = 0.024) as well as for dsxM (Kruskal Wallis rank sum test with df = 3: pupae: X2 = 30.2, p<0.001; adults: X2 = 22.6, p<0.001). Worker pupae showed significantly higher dsxF expression than queen pupae (pairwise t-test with BH correction: p = 0.013) and worker pupae and adults showed significantly lower dsxM expression than queen pupae and adults, respectively (Wilcoxon Tests with BH correction: pupae: p = 0.012; adults: p = 0.0014). Neither dsxF nor dsxM expression differed significantly between the two male morphs (dsxF: pairwise t-test with BH correction: pupae: p = 0.480, adults: p = 0.277; dsxM: pairwise Wilcoxon tests with BH correction: pupae: p = 0.481, adults: p = 0.805). However, overall expression of both isoforms was higher in winged compared to wingless males (Fig 1B). Our finding that dsx is differentially expressed and alternatively spliced across morphs in pupae and adults suggests that dsx might play a role in controlling polyphenic development.
To confirm that expression of dsx isoforms corresponds with phenotypic tissue differentiation, we used qPCR to analyze dsxM and dsxF expression in male and female-typical tissues dissected from aberrant “sex mosaic” individuals that express both male and female characters. C. obscurior sex mosaics are typically laterally separated into female and male halves, indicating that intersexuality is caused by single, early developmental aberrations such as anomalous fertilization events, loss of sex locus expression or inheritance of maternal effects [26–28]. The expression of dsxF and dsxM was male-typical in male tissue and female-typical in female tissue for all individuals except one, which had similar levels of dsxM in both tissue types (S3 Fig). As in previous studies [29,30], we only observed individuals possessing queen and winged male traits, or worker and wingless male traits; other trait combinations were absent (S4 Table), implying that common mechanisms control morph differentiation in males and females.
We analyzed published RNAseq data [31] from individual early 3rd instar larvae (QU, EM, WM, WO; n = 7 each) on an exon-level with DEXSeq [32]. We found morph-biased expression in each of the seven dsx exons, and confirmed sex-specific expression of the DM domain, dsxF, and dsxM in the early 3rd larval stage (S4 Fig and S5 Table). Overall, dsx expression was higher in males than in females, and higher in wingless morphs compared to winged morphs (EM > WM, WO > QU).
We hypothesized that other genes with sex-specific alternative splicing have been similarly co-opted for morph differentiation. Using a conservative false discovery rate of 0.005, DEXSeq analysis identified 179 exons of 91 genes with sex-biased expression (S6 Table). Dsx exon 5 (= dsxF) is ranked 5th among the top 10 differentially expressed exons and exons 6 and 7 (= dsxM) are the two most significant differentially expressed exons across all samples. To test for co-option of this set of exons into morph differentiation, we performed a hierarchical clustering analysis based on log-transformed exon counts. Queens and workers, as well as winged and wingless males, were clearly separated by the set of sex-biased exons, with the exception of two male samples that clustered with the wrong male morph (bootstrap node support: QU/WO = 75, WM/EM = 68) (Figs 2, S5 and S6 for bootstrap support for all nodes). Because terminal switch points for morph differentiation in male and female larvae may differ [31], misclassification of two male samples (WM34 & EM29) in hierarchical clustering may reflect higher plasticity in males compared to females at this particular developmental stage. Accordingly, in C. obscurior 3rd instar larvae, more genes are differentially expressed between queens and workers than between winged and wingless males [31].
To identify the sex-biased exons that most strongly affect separation between sexes and morphs, we performed a principal component analysis (PCA) of the 179 normalized exon counts. PC 1 separated sexes (29.9% explained variation), PC 2 (15.3%) and PC 4 (6.8%) separated female and male morphs, respectively (Fig 3; linear discriminant analysis using Wilk’s test on PCs 1, 2 and 4; factor sex: F(1,28) = 95.81, p < 0.001; factor morph: F(3,28) = 27.70, p < 0.001), while PC 3 (7.7%) did not separate between sexes or morphs (linear discriminant analysis using Wilk’s test on PC 3; factor sex: F(1,28) = 0.06, p = 0.80; factor morph: F(3,28) = 1.81, p = 0.17). From the 179 exons, we identified those with the strongest influence on sex (PC 1), female morph (PC 2), and male morph (PC 4) by extracting the exon loadings that fell in either the 10% or 90% quantiles for each PC (S6 Table). Using these lists, we identified dsx (replicating the RT-qPCR results) and seven other genes that showed both sex-specific and morph-specific alternative splicing, of which kr-h2 has a putative transcription factor function (Table 1). All eight genes are conserved across the Insecta, and a Gene Ontology (GO) term enrichment analysis with topGO [33] suggests that they serve basic metabolic and other core functions (S7 Table).
Our study suggests provides evidence that the sex differentiation pathway has been co-opted to control morph-specific development, as we predicted from the theory of facilitated variation. The major candidate gene dsx was alternatively spliced in males and females, and differentially expressed between queens and workers and between winged and wingless males. We independently replicated these results using qRT-PCR and RNAseq data from different individuals and life stages. Strikingly, we found that exons showing sex-biased expression were also differentially expressed between morphs, suggesting that dsx and other sex-biased genes mediate polyphenism within each of the sexes. The RNAseq analysis conservatively identified eight genes that have sex-specific and morph-specific alternative splicing; all of these genes were evolutionarily conserved and had GO terms associated with basic cellular functions. While dsx encodes sex-specific transcription factors and co-ordinates expression of a large number of downstream genes [34], except for a putative role of kr-h2 (see below) the other genes exhibit no transcription factor function. We confirmed that the sex-specific isoforms of dsx correlated with tissue type by analyzing male and female-typical tissue dissected from aberrant sex mosaic individuals. Finally, we reaffirmed that sex mosaics are always either hybrids of a queen and a winged male, or a worker and a wingless male, implying common morph differentiation control mechanisms in both sexes, especially regarding winglessness.
Interestingly, dsx has been shown to be a central hub gene involved in generating evolutionary novelty and polyphenism in other taxa. In a butterfly, genetic variation in dsx is associated with a heritable female-limited wing color/shape polymorphism, suggesting that dsx has been co-opted to control a novel, female-limited trait as well as maintaining its function in sex differentiation [24]. In the genus Drosophila, new localizations of dsx are thought to have facilitated the evolution of a novel male-limited trait (the sex combs), highlighting how the preexisting sex determination system was co-opted to produce a new polyphenism [35]. In the dung beetle Onthophagus taurus, RNAi experiments suggested that variation in dsx splicing mediates the difference in the presence of horns between males and females, and also controls a nutritionally dependent, male-limited polyphenism between large-horned and small-horned males [36]. A subsequent study of another horned beetle showed that different dsx isoforms control the sensitivity of the mandibles to juvenile hormone (JH), such that male mandibles are stimulated to grow by JH while those of females are not [37]. Thus it appears that dsx first evolved to mediate male-limited expression of horns by elevating the sensitivity of male horn tissue to JH [37] and perhaps also the IGF signaling pathway [38], and was then secondarily co-opted to control a nutrition-sensitive, male-limited polyphenism. The beetle dsx data are thus highly congruent with the theory of facilitated variation: the male polyphenism evolved using pre-existing genetic switches and developmental mechanisms to link a novel combination of stimuli and outputs (here, larval nutrition and horn phenotype).
Pre- and posttranscriptional genetic tools are not yet well established in ants but there is circumstantial evidence for similar links between dsx and JH in C. obscurior. A previous experiment showed that JH is involved in the development of larvae of both sexes into winged morphs [39], and the present study found differences in dsx splicing and expression between winged and wingless morphs. Thus, we speculate that the isoforms of dsx may mediate the responsiveness of developing tissues to JH, as hypothesized for beetles [37]. Significant differences in feminizer expression between queens and workers in the stingless bee Melipona [40], an upstream signal of dsx in bees [41], likewise suggests co-option of sex differentiation genes into caste differentiation in bees. There are no homologues of csd and feminizer in ants, because csd evolved in the Apis lineage by duplication of feminizer [42]. In ants, the closest homologue to feminizer is transformer. In C. obscurior, we could not detect morph-specific expression of the two transformer paralogues (tra1: Cobs_03145 and tra2: Cobs_18309), although they were expressed in a sex-specific manner.
In addition to dsx, we found a second sex-biased transcript with putative regulatory function. This ortholog to kr-h2 was alternatively spliced in queen and worker larvae, rendering the Kruppel homolog family a promising candidate for modulating plastic responses to the environment. kr-h2 has structural similarity to the JH-inducible transcription factor kr-h1 [43], which is involved in the initiation of metamorphosis in other insects [44,45]. kr-h2-induced differences in developmental timing may explain why metamorphosis is delayed in queens compared to workers [39], and further points to a link between sex-specific transcription, function in transcriptional regulation, sensitivity to JH, and evolutionary co-option into within-sex polyphenism.
We believe that the hypothesis advanced here, i.e. co-option of sex differentiation pathways into social insect caste polyphenism, is complementary to a previous theory regarding the proximate mechanisms underlying the origin of eusociality, termed the reproductive groundplan hypothesis (RGPH). Based on the ovarian ground plan hypothesis [46], the RGPH posits that eusociality arose via changes in the regulation of pre-existing gene sets relating to reproductive physiology and behavior, for example when genes involved in nest provisioning and brood care began to be expressed in unmated, non-reproductive individuals [47]. Research on the RGPH has stressed the importance of genes with nutrition-sensitive expression in delimiting the queen and worker “genetic toolkits”, in light of evidence that caste fate is nutrition-sensitive [48], that diet preference, reproduction and behavior are pleiotropically linked [49], and that some nutrition-related genes such as IRS and TOR influence caste fate [50]. Juvenile hormone, which is involved in regulatory feedback loops with some nutrition-related gene networks, has also been linked to caste differences [48,50]—including in our study species C. obscurior [31,39]—as well as to within-caste polymorphisms (e.g.[51]). Our hypothesis and the RGPH both argue that regulatory evolution caused conserved genes to acquire caste-specific expression. Our hypothesis is distinct in that it explicitly proposes that this regulatory evolution takes place in sex differentiation genes, but leaves the targets of these genes unspecified. By contrast, the RPGH makes predictions about which gene networks produce caste-biased phenotypes (e.g. ovary development, [52]), but makes no prediction regarding the identity of the regulatory sequences controlling these networks. Thus, the hypotheses do not overlap, and both may be correct. Analyses of potential regulatory links between the pathways presented here and those implicated with the RGPH will reveal to what extent they are connected.
Co-option of conserved genes involved primarily with sex differentiation in novel contexts allows functionally integrated gene networks to produce discrete phenotypes. Together with the horned beetle data reviewed above, our study suggests that core components of the sex differentiation pathway such as dsx can produce evolutionary novelty by acting as a switch for nutrition and JH-sensitive growth and development. Although many mechanisms of gene regulation have been implicated in controlling caste-specific development in social insects (e.g. methylation [53], transcription factors [31], small RNAs acting post-transcription [17], RNA editing [54] or structural chromatin modification [55]), all of these depend on some higher-level genetic switch to trigger differential activity. We propose that highly conserved hub genes such as dsx, which can translate variable input signals into large transcription differences using intermediate-level regulators, were the most basic mechanism responsible for the repeated evolutionary transition to eusociality and caste polyphenism.
Crosses between five queens of a C. obscurior population from Japan (JP) and five wingless males of a C. obscurior population from Brazil (BR) were set up by placing sexual pupae together with some brood and ~20 workers in plaster-filled Petri dishes. Nests were checked twice a week, provided with water, honey and pieces of dead insects and kept at constant conditions (12h 28°C light, 12h 23°C dark). We sampled emerging F1 hybrid QU, WO, EM and WM pupae and extracted DNA from the 10 parental and 71 F1 individuals (23 EM, 3 WM, 22 QU, 22 WO, 1 GY = gynandromorph, for sample sizes per family see S1 Table). Each individual was analyzed at three variable microsatellite loci (Cobs_1.1, Cobs_8.3, Cobs_8.4; for primer sequences see S8 Table). PCRs were performed using the BIO-X-ACT Short Mix (Bioline) and microsatellite analyses were carried out on an ABI PRISM (Applied Biosystems).
To find the functional dsx ortholog of C. obscurior, we identified DM domain-containing proteins of Drosophila melanogaster, Nasonia vitripennis, Apis mellifera, Pogonomyrmex barbatus, Acromyrmex echinatior and C. obscurior by BLASTp and tBLASTn analyses (S9 Table) and aligned them with MUSCLE [56]. We extracted the DM domain region from the manually corrected alignment (S7 Fig) and built a phylogenetic tree in MEGA [57], applying a WAG+G+I phylogenetic model and bootstrap resampling with 1,000 replicates (S1 Fig).
We reanalyzed previously published RNAseq data of larvae [31]. After removing adapter sequences with cutadapt and performing quality filtration with Trimmomatic, the reads were mapped against the reference genome with tophat2 (v2.0.8) and bowtie2 (v2.1.0) in sensitive mode. We generated count tables with HTseq based on the Cobs1.4 official gene set and used DESeq2 [58] to assess sex-specific expression of the four dsx paralogs following size factor normalization.
We applied RACE (Rapid Amplification of cDNA Ends) for identification of dsx isoforms. Total RNA was extracted from three females (QU adult, QU pupa, WO pupa) and three wingless males (one pupa, two adults) using the peqGOLD MicroSpin Total RNA Kit (peqlab). Transcription to cDNA was performed with the AffinityScript Multiple Temperature cDNA Synthesis Kit (Agilent Technologies), using the 3’ RACE Adapter GCGAGCACAGAATTAATACGACTCACTATAGGTTTTTTTTTTTTVN. 3’ RACE was performed in a nested PCR using two gene-specific 3’ primers (dsx4_for4, Co_dsx_p3_for, for primer sequences see S8 Table) and the 5’ primer provided in the First Choice RLM-RACE Kit (Ambion). PCRs were performed using the BIO-X-ACT Short Mix (Bioline) with the following protocol: 94°C (3 min), followed by 35 cycles 94°C (30 sec), 60°C (30 sec), 72°C (2 min) and a final elongation of 72°C (7 min). The products were purified with the NucleoSpin Gel and PCR Clean-up (Macherey-Nagel) and Sanger sequenced at LGC Berlin.
Total RNA was extracted from adults (8 EM, 8 WM, 10 QU, 10 WO) using the RNeasy Plus Mini Kit (Qiagen) and transcribed to cDNA using the AffinityScript Multiple Temperature cDNA Synthesis Kit (Agilent Technologies). Expression of the DM domain was quantified by qPCR using the primer pair dsx4_for4/dsx4_rev1 and normalized with two housekeeping genes (RPS2_new, RPL32; see S8 Table for primer sequences).
We further used qPCR to measure isoform-specific dsx expression (dsxM and dsxF) in pupae and adults of all four morphs, and in tissue from four sex mosaic pupae. We dissected the head and thoraces of the sex mosaics (for morphological descriptions see S4 Table) laterally into male and female halves and stored male and female tissue parts separately in RNAlater-ICE (Ambion), resulting in one female and one male sample per individual. We extracted total RNA from 9–10 pupae and seven adults of each of the four morphs, and from the sex mosaic tissue using the peqGOLD MicroSpin Total RNA Kit (peqlab) including a DNA digestion step with the peqGOLD DNase I Digest Kit (peqlab). After cDNA synthesis with iScript cDNA Synthesis Kit (Bio-Rad) we quantified gene expression of dsxF and dsxM using isoform specific, intron-spanning primers (dsxF: 4for/F5rev, dsxM: 4for/M5rev; see Fig 1A for position of primers) and two housekeeping genes (RPS2_new, Y45F10D_JO1). All qPCR reactions were performed in triplicates (repeatability was uniformly high, so we took the mean of the three replicates prior to analysis). Data analysis was carried out according to [59], using the geometric mean of the two housekeeping genes for normalization.
We analyzed published RNAseq data [31] from 3rd star instar larval QU, WO, WM and EM (n = 7 each) and assessed differential exon-specific expression with DEXSeq [32]. Raw reads were trimmed and passed through quality filtration as described in [31] and mapped to the reference genome Cobs1.4 [19] using STAR [60]. We corrected the dsx and tra gene model using the RACE results for dsx, and split the tra gene model into two paralogs (tra1 and tra2), as observed in other ants [42,61,62]. For all other genes we used gene models of the Cobs1.4 official gene set. We followed the default workflow of DEXSeq and tested for differential exon usage between males and females based on a false discovery rate of 0.005. In the resulting 179 sex-specific exons, we tested for morph-specific exon profiles using hierarchical clustering (implemented by the R function hclust using the ward.D2 method [63]) of pairwise Manhattan distances between log-transformed normalized exon counts. We assessed the support for each node in the cluster analysis using bootstrap resampling with 10,000 replicates using the pvclust package in R 3.1.2.
We conducted a PCA with normalized exons counts. We visually identified principal components that best separated between sexes (PC 1), female morphs (PC 2) and male morphs (PC 4) and confirmed that these components suffice to separate among sexes and morphs with linear discriminant analysis and subsequent Wilk’s tests in R 3.1.2. Based on loadings of exons on each component, we identified exons that fell in either 10% or 90% quantiles (S6 Table) as those with the strongest influence on PC 1, PC 2 and PC 4. From this list, we extracted only those genes that contained multiple exons with strong influence on both sex (PC 1) and morph (PC 2 and/or PC 4). This yielded a list of eight candidate genes showing alternative splicing between sexes as well as morphs (Table 1).
|
10.1371/journal.pgen.1001055 | Multivesicular Body Formation Requires OSBP–Related Proteins and Cholesterol | In eukaryotes, different subcellular organelles have distinct cholesterol concentrations, which is thought to be critical for biological functions. Oxysterol-binding protein-related proteins (ORPs) have been assumed to mediate nonvesicular cholesterol trafficking in cells; however, their in vivo functions and therefore the biological significance of cholesterol in each organelle are not fully understood. Here, by generating deletion mutants of ORPs in Caenorhabditis elegans, we show that ORPs are required for the formation and function of multivesicular bodies (MVBs). In an RNAi enhancer screen using obr quadruple mutants (obr-1; -2; -3; -4), we found that MVB–related genes show strong genetic interactions with the obr genes. In obr quadruple mutants, late endosomes/lysosomes are enlarged and membrane protein degradation is retarded, although endocytosed soluble proteins are normally delivered to lysosomes and degraded. We also found that the cholesterol content of late endosomes/lysosomes is reduced in the mutants. In wild-type worms, cholesterol restriction induces the formation of enlarged late endosomes/lysosomes, as observed in obr quadruple mutants, and increases embryonic lethality upon knockdown of MVB–related genes. Finally, we show that knockdown of ORP1L, a mammalian ORP family member, induces the formation of enlarged MVBs in HeLa cells. Our in vivo findings suggest that the proper cholesterol level of late endosomes/lysosomes generated by ORPs is required for normal MVB formation and MVB–mediated membrane protein degradation.
| The multivesicular body (MVB) sorting pathway provides a mechanism for the lysosomal degradation of membrane proteins, such as growth factor receptors. The formation of MVBs is unique in that the curvature is directed toward the lumen of the compartment rather than the cytosol. During MVB formation, the curvature-inducing proteins, such as clathrins, could not be involved in the inward invagination of the endosomal membrane. Under these circumstances, lipids have been assumed to play a role in the membrane invagination step by creating local membrane environments; however, the lipids involved in this step have not been fully elucidated. Here we demonstrate that cholesterol, an essential membrane component in animals, is critical for MVB formation and function. We found that disruption of OSBP–related proteins (ORPs), which have been proposed to function in cellular cholesterol distribution and metabolism, reduces the cholesterol content in late endosomes/lysosomes, leading to impaired MVB function. MVB sorting pathway is known to be involved in many processes, including growth factor receptor down-regulation, exosome secretion, antigen presentation, the budding of enveloped viruses, and cytokinesis. Our findings provide a novel link between cholesterol and these biologically important functions.
| The multivesicular body (MVB) sorting pathway provides a mechanism for the lysosomal degradation of membrane proteins and has a role in many processes, including growth factor receptor down-regulation [1], antigen presentation [2], developmental signaling [3], [4], the budding of enveloped viruses [5], and cytokinesis [6], [7]. MVBs form when the limiting membrane of the late endosomes invaginates and buds into the lumen of the organelle, selecting a subset of the proteins from the limiting membrane in the process [8], [9]. The MVB sorting machinery is constituted by proteins that form the endosomal sorting complexes required for transport (ESCRT-I, -II, and -III) [10], [11]. These ESCRT complexes are recruited sequentially to endosomal membranes where they function in sorting cargo and generating characteristic intralumenal vesicles. MVBs then fuse with lysosomes, resulting in degradation of their cargo. In addition to the ESCRT proteins, lipid molecules have been assumed to be involved in MVB formation by creating local microdomains in the endosomal membrane that induce the inward membrane curvature. For example, lysobisphosphatidic acid (LBPA) and ceramide were shown to induce the formation of internal vesicles in liposomes [12], [13]. Furthermore, treatment with anti-LBPA antibodies disrupts normal MVB formation in mammalian cells, suggesting that LBPA has a role in driving lumenal-vesicle formation at the cellular level [14].
In eukaryotes, different organelles within a cell generally have distinct cholesterol concentrations. Such differences are thought to be necessary for various biological functions ranging from membrane trafficking to signal transduction [15]. Obtaining the normal subcellular cholesterol distribution is thought to require a variety of intracellular cholesterol movements through vesicular and nonvesicular mechanisms [16], [17]. Recently, oxysterol-binding protein (OSBP) and OSBP-related proteins (ORPs) have been shown to mediate a number of cellular processes including signal transduction, lipid metabolism, vesicular trafficking and nonvesicular sterol transport [18]–[20]. OSBP was first identified as a high-affinity cytosolic receptor for oxysterols, such as 25-hydroxycholesterol [21]. Subsequently, most eukaryotes have been shown to have proteins homologous to OSBP, including 12 ORP-homologs in humans (OSBP and ORP1 to ORP11), four in C. elegans (this study; OBR-1 to OBR-4), four in D. melanogaster, and seven in the budding yeast S. cerevisiae (Osh1p to Osh7p) [19], [22]. Most ORPs share two highly homologous structural features: a PH domain at the amino-terminus and a ∼400-amino acid sterol-binding domain at the carboxy-terminus (Figure S1) [19]. The mammalian ORP family can be subdivided into six subfamilies (I–VI) based on gene organization and amino acid homology. Yeast ORPs share comparatively low sequence homologies with mammalian ORP proteins and are not classified into the ORP subfamilies, whereas C. elegans and D. melanogaster ORPs clearly fall into subfamilies I, II, IV and V based on the homology of the sterol-binding domains (Figures S1, S3, S4, S5, S6).
Many lines of evidence suggest that ORPs have a role in sterol distribution among intracellular organelles. Raychaudhuri showed that yeast ORPs (Osh4p, Osh5p, and Osh3p) have a role in transporting sterol from the yeast plasma membrane to the esterification compartment, ER [18]. In addition, the cholesterol distribution in yeast ORPs mutants was abnormal. A crystal structure analysis indicated that Osh4p is able to accommodate a variety of sterols including cholesterol [23]. In in vitro analyses, Osh4p and mammalian ORPs transferred sterols from donor to acceptor liposomes [18], [24]. In mammalian cells, the transport of newly synthesized cholesterol from the ER to the cell surface is enhanced by expression of ORP2 [25]. Although increasing evidence supports the involvement of ORP proteins in subcellular cholesterol distribution, knockout studies of ORPs in animals have not been reported, and consequently, the biological significance of distinct cholesterol concentrations in subcellular compartments remains to be elucidated.
In the present study, we generated deletion mutants of all ORP family members in C. elegans (obr-1, -2, -3, and obr-4). We also performed an RNAi modifier screen using obr quadruple mutants and found that a group of MVB-related genes including ESCRT complex genes show strong genetic interactions with obr genes.
A database search revealed the presence of four ORP family members in C. elegans, which are classified into ORP subfamilies I, II, IV and V based on the homology of the sterol-binding domains. We named these ORP genes obr-1, obr-2, obr-3, and obr-4, respectively [obr: Oxysterol Binding protein (OSBP) Related (Figure S1, S3, S4, S5, S6) [19]]. To address the functions of ORP members, we generated deletion mutants of all four ORP genes in C. elegans by PCR-based screening of TMP/UV-mutagenized libraries (Figure S2) [26]. All of these mutations appear to be null or strong loss-of-function alleles because inhibition of each obr gene by RNAi failed to enhance the obr quadruple mutant phenotypes, such as embryonic lethality and slow growth as described below.
Single mutant worms with deletions in obr-1, obr-2, obr-3, or obr-4 were viable and fertile, and displayed an essentially normal phenotype under a dissection microscope (Table 1). The obr-1;obr-2;obr-3;obr-4 quadruple mutants that lacked all obr genes exhibited embryonic lethality (∼11%) and slow growth during larval development (∼18%) (Table 1). Hatched obr quadruple mutants were able to develop to adults and produce subsequent progeny, although they had a reduced brood size (60% of that of wild-type worms) and showed abnormal cuticle structure (Figure S7B and S7D). These data indicate that four C. elegans ORP proteins act redundantly during embryonic and larval development. This is similar to the case in yeast where any one of the 7 ORPs is sufficient for viability [27].
To gain insights into the molecular mechanisms of embryonic lethality in obr quadruple mutants, we conducted a synthetic lethal screen. We used feeding RNAi clones on chromosomes I and III in the Ahringer library to identify RNAi clones that cause embryonic lethality in the obr quadruple mutant background, but not in the wild-type background (see Materials and Methods, and Table S1). As a result, we identified 28 genes that showed synthetic lethality in obr quadruple mutants (Table S2, hereafter, we refer to obr-1;obr-2;obr-3;obr-4 quadruple mutants as the “obrs mutants”). These enhancer genes included genes encoding vesicular transport-related proteins, signaling proteins, and nuclear proteins. Interestingly, among the 28 enhancer genes, 6 genes (hgrs-1, vps-28, vps-2, vps-20, vps-4, and vps-34) have been reported to function in the formation of multivesicular bodies (MVBs), the machinery for degrading membrane proteins (Figure 1). Knockdown of vps-4 caused complete embryonic lethality in the obrs mutants as compared to 13% embryonic lethality in wild-type animals (Figure 1). RNAi against other enhancer genes (hgrs-1, vps-28, vps-2, vps-20, and vps-34) also showed remarkably increased embryonic lethality (50–80%) in the obrs mutants as compared to wild-type worms (0–10%) under the present feeding RNAi conditions (Figure 1). In the eri-1(mg366); lin-15B(n744) background, which is hypersensitive to RNAi, knockdown of vps-4, hgrs-1, vps-28 or vps-32.2 (a component of ESCRT-III) resulted in embryonic lethality with high penetrance in wild-type worms, indicating that ESCRT components are essential for embryonic development in C. elegans [28] (data not shown).
Formation of MVBs requires the components of four complexes that include Vps27 (sometimes referred to as ESCRT-0), ESCRT-I, ESCRT-II, and ESCRT-III [10]. These complexes are recruited sequentially to endosomal membranes where they function in sorting cargo and generating intralumenal vesicles. The 6 obr enhancer genes encode C. elegans homologues of the ESCRT components or their regulatory molecules. These include hgrs-1, a homologue of yeast Vps27, vps-28, a component of ESCRT-I, vps-2 and vps-20, components of ESCRT-III, vps-4, a homologue of yeast Vps4/AAA ATPase that is recruited by ESCRT-III to disassemble and recycle the ESCRT machinery, and vps-34, a class III phosphoinositide 3 (PI3) kinase required for recruitment of ESCRT-0 to early endosomal membranes. ZK930.1, a homologue of mammalian p150 that encodes the PI3 kinase regulatory subunit, was also identified as an obr enhancer gene (Figure 1; Table S2).
A strong genetic interaction between obr genes and MVB-related genes led us to hypothesize that late endocytic compartments (late endosomes/lysosomes) are affected in obrs mutants. In S. cerevisiae, disruption of an MVB-related gene, such as vps-4 or vps-28, causes enlargement of aberrant late endocytic compartments and disturbance of membrane protein degradation [29]. To assess the morphology of late endocytic compartments in obrs mutant embryos, we first used the fluorescent probe LysoSensor Green, which accumulates in acidic compartments because of protonation [30]. In wild-type embryos, the probe localized to small punctate vesicles throughout embryogenesis (Figure 2A). In contrast, in obrs mutant embryos, the number of large fluorescent vesicles increased, indicating that late endocytic compartments were enlarged in obrs mutant embryos (Figure 2B). Knockdown of vps-4 in wild-type embryos also caused the appearance of similar large fluorescent vesicles as observed in obrs mutant embryos (Figure 2C), and these enlarged vesicles were synergistically increased in the obrs mutant background (Figure 2D, Figure S12A). These data indicate that in obrs mutant embryos, late endocytic compartments were enlarged and these morphological defects were enhanced by knockdown of the MVB-related genes.
We next examined the expression of LET-23, a C. elegans homologue of the epidermal growth factor (EGF) receptor which is known to be degraded via the MVB pathway [1]. In wild-type embryos, LET-23::GFP was observed mostly in small punctate vesicles (<0.7 µm) during embryonic morphogenesis (Figure 2I and 2Q). In obrs mutants, LET-23::GFP vesicles were enlarged and GFP intensity was stronger than that in wild-type embryos (Figure 2J). A portion of LET-23::GFP-positive vesicles colocalized with the LysoTracker-labeled endosomes/lysosomes (Figure S12G, S12H, S12I). Knockdown of vps-4 in wild-type embryos also caused the appearance of large vesicles similar to those observed in obrs mutant embryos (Figure 2K and 2Q). These observations indicate that LET-23::GFP was partly localized to late endocytic compartments in obrs mutants, although it is possible that some of the LET-23::GFP was localized to compartments other than endosomes/lysosomes. In the obrs mutant background, the GFP level was synergistically increased by knockdown of an MVB-related gene such as vps-4, hgrs-1 or vps-28 (Figure 2L and 2Q). These results suggest that degradation of the EGF receptor LET-23 was retarded in obrs mutants and this defect was synergistically enhanced by knockdown of MVB-related genes.
We then investigated intracellular transport of soluble proteins endocytosed from the extracellular fluid (the body cavity) to late endocytic compartments in coelomocytes, scavenger cells that are highly active in endocytosis [31]. We first examined the morphology of endosomes and lysosomes in coelomocytes, and found that RME-8-labeled late endosomes (Figure 3A–3C) [32] and LMP-1-labeled lysosomes (Figure 3D–3F) [33] were significantly enlarged in obrs mutants. Enlargement of late endosomes/lysosomes was also observed in vps-4, vps-2 or hgrs-1 RNAi worms (Figure 3C, 3F–3H, and data not shown). These data are in agreement with the previous results showing enlargement of LysoSensor Green-positive vesicles in obrs mutant embryos (Figure 2B–2D and Figure S12A). In contrast, there appeared to be no differences in fluorescence patterns of early endosomes (2x FYVE::GFP), Golgi (AMAN-2::GFP), endoplasmic reticulum (GFP::TRAM) between wild-type and obrs mutants (Figure S9A, S9B, S9C, S9D, S9E, S9F, S9G).
Next, we investigated the fluid-phase endocytosis in obr mutants using a transgenic strain that secretes GFP from the muscle into the body cavity (myo-3p::ssGFP) [31]. In wild-type animals, the secreted soluble GFP (ssGFP) was rapidly endocytosed by the coelomocytes and degraded (Figure 3I and 3J) [31]. However, mutants defective in endocytosis or intracellular transport of endocytosed soluble proteins in coelomocytes showed increased levels of ssGFP in the body cavity [31]. In the obrs mutants, ssGFP appeared to be efficiently endocytosed by coelomocytes, producing animals with bright green coelomocytes as observed in wild-type worms (Figure 3K and 3L). To obtain higher temporal resolution, we microinjected Texas-red BSA into the body cavity of the worms [31], [34]. In wild-type worms, 20 min after injection, the marker started accumulating in the late endosomes of coelomocytes as indicated by the RME-8::GFP-positive compartments (Figure 3M). By 60 min, it was observed increasingly in lysosomes but was absent from RME-8::GFP-positive late endosomes (Figure 3N and 3O). In obrs mutants, the fluid-phase endocytosis and postendocytic trafficking proceeded with the same kinetics as observed in wild-type worms (Figure 3P–3R). We also checked receptor-mediated endocytosis of a yolk protein VIT-2 in oocytes (Rme) [35], and found that VIT-2 was efficiently incorporated into oocytes in the obrs mutants in a similar manner to that in wild-type worms (Figure S8A and S8C). Taken together, these data indicate that endocytic trafficking of soluble proteins to lysosomes is not affected in obrs mutant coelomocytes.
We next examined internalization and subsequent degradation of cell surface membrane proteins in the obrs mutants. To this end, we used a transgenic worm expressing a member of the caveolin protein family, CAV-1, that has been reported to be degraded via the MVB pathway during the oocyte-to-embryo transition [36]. In control oocytes prior to fertilization, CAV-1::GFP was concentrated in intracellular vesicles (Figure 4A, an oocyte indicated by “−1”) [37]. Immediately after oocytes passed through the spermatheca and were fertilized, the CAV-1::GFP signal of intracellular vesicles was lost and the CAV-1::GFP signal on plasma membrane rapidly increased (Figure 4A, an embryo indicated by “+1”). Most of CAV-1::GFP was internalized and degraded in the one-cell stage embryo and was not observed beyond the two-cell stage (Figure 4A, Figure S10A and S10B, embryos indicated by “+2” to “+4”). The post-fertilization increase in the amount of CAV-1::GFP on the cell surface and its subsequent re-internalization were not affected either in the obrs mutants or vps-4 RNAi worms (“+1” and “+2” embryos in Figure 4A–4C). Consistent with previous results [36], knockdown of an MVB-related gene, such as vps-4, hgrs-1, vps-28, or vps-20, resulted in a substantial delay in the degradation of internalized CAV-1::GFP, which remained on internal membranes even in the “+5” embryo (an embryo at about the 26-cell stage) (Figure 4B and data not shown). The obrs mutants exhibited slightly but significantly retarded degradation of internalized CAV-1::GFP, where significant CAV-1::GFP signal was observed in intracellular membranes of +2 and +3 embryos (Figure 4C). A western blot analysis also revealed that the amount of CAV-1::GFP increased in the obrs mutants (Figure 4D). The milder defects in CAV-1::GFP degradation in the obrs mutants than in vps-4 RNAi worms indicate that obr genes are not essential for the degradation of membrane proteins, but are required for efficient degradation of those proteins in C. elegans embryos.
Because ORPs have been implicated in intracellular cholesterol transport, we tested the possible involvement of cholesterol in MVB formation. C. elegans requires cholesterol for normal development, but does not possess the enzymes necessary for de novo sterol biosynthesis. Therefore C. elegans membrane cholesterol must be supplied by the diet [38]. The first generation of wild-type worms placed on cholesterol-depleted plates develop from eggs to adults without external cholesterol because cholesterol is supplied from mother worms grown on normal plates (Brenner condition; 5 µg/ml of cholesterol). However, 5% of second-generation embryos died (Figure 5A) and the development of all hatched larvae was arrested at the early larval stage (data not shown) [39]. Under these cholesterol-restricted conditions, second-generation obrs mutants exhibited 96% embryonic lethality whereas the mutants showed only 11% embryonic lethality under cholesterol-supplemented conditions (Figure 5A). The hypersensitivity of obrs mutants to cholesterol deprivation suggests that the OBR proteins are involved in the utilization of cholesterol in C. elegans.
We next performed knockdown of MVB-related genes under cholesterol-restricted conditions (see Materials and Methods). Under cholesterol-restricted conditions, knockdown of MVB-related genes, such as hgrs-1 and vps-4, resulted in remarkably reduced viability and high penetrance embryonic lethality (Figure 5A). The reduced viability of hgrs-1(RNAi) and vps-4(RNAi) worms under cholesterol-restricted conditions is similar to that observed in the obrs mutant background (Figure 1). These results suggest that cholesterol content is critical for MVB formation during embryogenesis and that obr molecules regulate cholesterol content in C. elegans.
To examine whether the late endosomal/lysosomal defects observed in obrs mutants occur in wild-type worms under cholesterol-restricted conditions, we again used LysoSensor Green to visualize late endocytic compartments. As observed in obrs mutants (Figure 2B), late endocytic compartments were enlarged under the cholesterol-restricted conditions (Figure 5C). We also found that LET-23::GFP vesicles were enlarged and their GFP intensity was stronger under cholesterol-restricted conditions than under cholesterol-supplemented conditions (Figure 5F and 5G). These data indicate that cholesterol is essential for the normal morphology of late endocytic compartments and for the degradation of membrane proteins via MVB formation.
To examine the cholesterol content of the late endocytic compartments, wild-type and obrs mutants were fed with radioactive cholesterol and homogenized with a Dounce homogenizer device [40]. The crude membrane fraction (20,000×g ppt in Figure S11A) was subjected to density gradient centrifugation by using a Lysosome Isolation Kit (see Materials and Methods). ER and Golgi membranes were found in the high-density fractions (Figure S11A; fractions #1–4, PAF-2 and COGC-3, respectively) and late endosomes/lysosomes were recovered in the low-density fractions (Figure S11A; fractions #7, 8, RAB-7::GFP). In wild-type animals, appreciable amount of radioactive cholesterol was recovered in the late endosomal/lysosomal fractions (fractions #7 and #8), whereas the cholesterol content in the late endosomal/lysosomal fractions of the obrs mutants was approximately 75% of that of wild-type worms (Figure S11A and S11B). The total cholesterol content in obrs mutants was also reduced significantly (to ∼60% of that of wild-type, Figure S11C), indicating that ORPs are also important for determining the cholesterol content of C. elegans.
Finally, we examined whether the functions of C. elegans obr members are conserved across species. We expressed all human ORP family members in HeLa cells and found that only ORP1L localized at lysosomes (data not shown) as reported previously [41]. ORP1L is structurally classified to ORP subfamily II which includes C. elegans obr-2 (Figure S1 and Figure S4). To determine the effects of ORP1L depletion on late endosomal/lysosomal morphology, we analyzed the morphology at the ultrastructural level by electron microscopy. In control cells, late endosomal/lysosomal compartments appeared as relatively dense round structures of 0.2- to 1-µm diameter, in which numerous small vesicles (MVBs) could be seen (Figure 6B and 6C). In contrast, large swollen vacuoles of 0.6- to 1.8- µm diameter appeared in ORP1L siRNA-treated cells (Figure 6A and 6D–6F). These enlarged structures appeared to be MVBs because they still contained some intralumenal vesicles, although significantly less in number compared with the intralumenal vesicles in MVBs of control cells. Furthermore, ORP1L siRNA-treated cells had ∼30% less MVBs than control cells (Figure 6G). We next investigated whether depletion of ORP1L affects EGF receptor degradation (see Text S1). In HeLa cells treated with the control siRNA, the EGF receptor was gradually degraded after 1, 2, and 3 hr of EGF stimulation. siRNA against ORP1L delayed EGF-induced receptor degradation more than the control siRNA (Figure S13A and S13B). In conclusion, these results indicate that ORP1L is required for MVB formation, normal morphology of late endosomes/lysosomes and membrane protein degradation, and these functions are evolutionarily conserved in mammals.
Cholesterol is a structural component of animal membranes that influences fluidity, permeability and formation of lipid microdomains. ORP family members have been implicated in the cholesterol distribution among intracellular organelles [18]–[25]; although their in vivo functions are not fully understood. In the present study, we generated deletion mutants of all ORP family members in C. elegans (obr-1, -2, -3, and obr-4) (Figure S2; Table 1). We also performed an RNAi modifier screen using obr quadruple mutants (obrs mutants) and found that a group of MVB-related genes including ESCRT complex genes show strong genetic interactions with obr genes (Figure 1; Table S2).
In obrs mutants, degradation of membrane proteins, such as an EGF receptor (LET-23::GFP) (Figure 2I–2L) and caveolin (CAV-1::GFP) (Figure 4), is delayed and late-endosomes/lysosomes are enlarged (embryos; Figure 2B, coelomocytes; Figure 3B and 3E). At the ultrastructural level, obrs mutants have enlarged vacuoles which are not observed in wild-type worms (Figure S7A and S7B). Similar defects of endocytic compartment have been reported in ESCRT-depleted S. cerevisiae [29] and mammalian cells [42], [43], in which MVB formation is impaired. These observations indicate that ORP molecules are required for efficient membrane protein degradation via the MVB sorting pathway. On the other hand, endocytosed soluble proteins, such as GFP and Texas-red BSA, are normally delivered to lysosomes and are efficiently degraded in obrs coelomocytes (Figure 3K, 3L, and 3P–3R). This data indicate that, at least in obrs coelomocytes, endocytic trafficking from the plasma membrane to lysosomes is not affected and that fusion of late endosomes and lysosomes occurs normally to generate mature lysosomes. Together, these observations suggest that ORP molecules are selectively involved in the degradation of membrane proteins via the MVB sorting pathway. In this study, we analyzed embryonic epithelial cells (Figure 2I–2Q, Figure 5F and 5G) and fertilized eggs (Figure 4) to examine the degradation of membrane cargos (LET-23::GFP and CAV-1::GFP, respectively), and analyzed coelomocytes (Figure 3I–3R) to examine the degradation of lumenal cargos (GFP and Texas-red BSA). The finding that lumenal cargos are normally degraded while membrane cargos are not may be because of tissue differences rather than differences in the cargo-specific functions of ORPs. Therefore, further analyses will be needed to determine if ORPs are involved in the degradation of lumenal cargos in general.
How are ORP molecules involved in MVB formation? In the present study, we showed that the total cholesterol content in obrs mutants was significantly reduced compared to wild-type worms, indicating that ORPs are important for utilization of cholesterol in C. elegans (Figure S11C). We also demonstrated that the cholesterol content of late endosomes/lysosomes was reduced in obrs mutants (Figure S11A and S11B). How C. elegans ORPs control the intracellular cholesterol level is unclear at this time. As mentioned above, ORPs are implicated in many cellular processes including signal transduction, cholesterol metabolisms, vesicular transport and nonvesicular sterol transport [20]. One possibility is that ORPs is involved in cholesterol transport to late endosomes/lysosomes directly by binding cholesterol or indirectly by regulating other cholesterol-binding proteins. ORPs may also control intracellular signaling and/or vesicular transport that determine the cholesterol content among intracellular organelles.
In obrs mutants, knockdown of MVB-related genes remarkably increased embryonic lethality (Figure 1). Knockdown of MVB-related genes also induces high penetrance embryonic lethality under cholesterol-restricted conditions (Figure 5A). Furthermore, late-endosomes/lysosomes are enlarged in both obrs mutants and cholesterol-restricted worms (Figure 2B and Figure 5C). These observations suggest that in obrs mutants, reduction of late endosomal/lysosomal cholesterol content disturbs MVB formation to some extent, and leads to hypersensitive lethality when the expression of MVB-related genes is knocked down. Another possibility is that the reduced cholesterol content in late endosomes/lysosomes indirectly affects MVB function. For example, the reduced cholesterol content might inhibit Golgi-to-lysosome transport of proteins that are required for MVB formation.
In addition to acting as cholesterol transfer proteins, ORPs have also been proposed to act as a sterol sensor that controls cell signaling [44]. Furthermore, two yeast ORPs (Osh6p and Osh7p) have been shown to interact with Vps4p, which has a role in dissociating the ESCRT-III complex from the endosomal membrane [45], suggesting that ORPs directly regulate ESCRT function in response to the cellular cholesterol content. We found that the localization of an ESCRT-III component (VPS-20) is not affected in obrs mutants (Figure S12J, S12K, S12L) and that the localization of mCherry::OBR-2, which fully restores the lysosomal morphology of obrs mutants (Figure S12B, S12C, S12D, S12E, S12F), is not altered by knockdown of the MVB-related genes (data not shown). Further studies are needed to determine whether ORPs are directly involved in ESCRT function.
The formation of MVBs is unique in that it is directed toward the lumen of the compartment, rather than the cytosol [46]. During MVB formation, curvature-inducing proteins, such as clathrins and coat protein complexes, could not be involved in the inward invagination of the endosomal membrane. It is also unlikely that the ESCRT proteins directly induce the invagination of the endosomal membrane without getting trapped in the lumen of the forming vesicles. Under these circumstances, lipids have been assumed to play an important role in the membrane invagination step by creating local membrane environments [47]. In mammalian cells, cholesterol is concentrated in endosomal/lysosomal compartments, especially in the luminal vesicles of MVBs [48]. C. elegans also has a considerable amount of cholesterol in the endosomal/lysosomal fraction (Figure S11A and S11B). However, the mechanism for accumulation of cholesterol in endosomes/lysosomes is largely unknown, and consequently, the biological significance of cholesterol in endosomal/lysosomal compartments has not been fully elucidated. In this study, we showed that disruption of ORPs reduces the cholesterol content in the endosomal/lysosomal compartments and impairs the MVB formation and function. Although it is not clear at present that the decrease in the cholesterol content is a direct cause of MVB abnormalities, the present study lay a firm basis for further work to more fully elucidate how cholesterol is involved in MVB formation.
In C. elegans, cholesterol depletion induces multiple responses such as embryonic lethality, dauer larva formation, and molting defects [38], [39]. Dauer larva formation is regulated by steroid hormone signaling, in which cholesterol-metabolizing enzymes DAF-36 (Rieske-like oxygenase) and DAF-9 (Cytochrome P450) are thought to convert cholesterol into steroid hormones, such as 4-dafachonic acid, that act on a steroid hormone receptor, DAF-12 [49], [50]. C. elegans molting is also thought to be regulated by cholesterol-derived steroid hormones via a steroid hormone receptor, NHR-25 [51]. We have never observed dauer larva formation or molting defects in obrs mutants, suggesting that obr mutations do not affect signaling by these steroid hormones.
In this study, we demonstrated that human ORP1L is required for MVB formation in mammalian cells. A previous study demonstrated that the GTPase Rab7, when bound to GTP, simultaneously binds to ORP1L and RILP to form a RILP-Rab7-ORP1L complex, which is required for the perinuclear localization of late endosomes/lysosomes [52], [53]. Mammalian ORP1L contains three ankyrin repeats at the amino-terminal end, and the interaction with Rab7 through the ankyrin repeats of ORP1L is essential to specify the perinuclear localization of late endosomes/lysosomes (Figure S1) [41]. In C. elegans and D. melanogaster, the obr gene products lack the amino-terminal ankyrin repeats and the late endosomes/lysosomes are not organized into the characteristic perinuclear cluster observed in mammalian cells (Figure S1). These observations suggest that the fundamental role of ORP1L is to maintain enough cholesterol in late endosomes/lysosomes for normal MVB formation. They also suggest that the perinuclear localization of late endosomes/lysosomes in mammals is the result of the appearance of the amino-terminal ankyrin repeats of ORP1L.
As mentioned above, MVB formation requires the inward invagination of the endosomal membrane. Similar membrane invagination also occurs in exosome formation, cytokinesis and viral budding. There is accumulating evidence that the ESCRT proteins have a role in this type of membrane fission. HIV budding from the plasma membrane also requires ESCRT proteins such as Hrs, a homologue of hgrs-1. Interestingly, it has been reported that HIV envelopes contain a high level of cholesterol and cholesterol depletion impairs HIV-1 budding at the plasma membrane. Further studies are needed to assess the involvement of ORP proteins in this process.
In addition to 6 MVB-related genes (hgrs-1, vps-28, vps-2, vps-20, vps-4, and vps-34), we identified 22 other genes that showed synthetic lethality in obr quadruple mutants (Table S2). At the present time, the reason for the strong interaction between these 22 genes and obr genes is unclear. However, like MVB-related genes, several enhancer genes may require a cholesterol-rich membrane environment for their normal functions. Cholesterol-rich microdomains play important roles in several biological functions, such as raft-dependent cellular signaling and caveolae-mediated endocytosis at the plasma membrane [15]. The present study suggested a novel role of cholesterol-rich microdomains, i.e. providing an adequate membrane environment for MVB formation. Further studies of the enhancer genes should uncover other aspects of intracellular cholesterol functions.
Worm cultures, genetic crosses, and other C. elegans methods were performed according to standard protocols [54] except where otherwise indicated. obr-1(xh16), obr-2(xh17), obr-3(tm1087) and obr-4(tm1567) mutants were isolated by TMP (trimethylpsoralen)/UV method [26] and were backcrossed onto the wild-type background five times before phenotypic analysis. Transgenic strains used for this study are cdIs36[punc-122p::C31E10.7::GFP] for endoplasmic reticulum, cdIs54[pcc1::MANS::GFP] for Golgi, pwIs50[lmp-1::GFP] for lysosomes, cdIs85[pcc1::2xFYVE::GFP] for early endosomes, bIs34[rme-8::GFP] for late endosomes, cdIs39[pcc1::GFP::RME-1] for recycling endosomes, arIs37[myo-3p::ssGFP], pwIs28[pie-1p-cav-1::GFP7] tmIs105[vit-2::GFP], xhIs2501[dpy-7p::let-23::GFP], xhEx2503[obr-2 genome::GFP], xhEx2511[unc122p::mCherry::obr-1], xhEx2512[unc122p::mCherry::obr-2], xhEx2513[unc122p::mCherry::obr-3], and xhEx2514[unc122p::mCherry::obr-4]. Some of the strains used in this work were obtained from Caenorhabditis Genetics Center, University of Minnesota, Minneapolis, MN).
Adult wild-type and mutant worms were allowed to lay eggs for 2–3 hr, and the progeny were scored for embryonic lethality and larval arrest. Unhatched eggs were examined 24 hr after being laid, and hatched but arrested larvae were examined 72 hr after being laid. To perform fluid-phase endocytosis assay, Texas red BSA was injected at 1 mg/ml in water into the body cavity of wild-type or obr quadruple mutants expressing RME-8::GFP. At defined time points, animals were mounted on slides, put on ice to stop endocytosis, and fluid-phase internalization of the dye into the coelomocytes was viewed with a confocal microscope. For the quantification of endosomes and lysosomes sizes, discrete intracellular structures in at least 30 coelomocytes were analyzed for each marker (RME-8::GFP for late endosomes, LMP-1::GFP for lysosomes). Individual sections through coelomocyte were scanned, and the diameter of the largest endosomes or lysosomes was scored. Coelomocyte, endosomes and lysosomes areas were calculated from their diameter. To quantify the size of LET-23::GFP-positive endocytic compartments in embryos, LET-23::GFP-positive endocytic compartments were sorted into three size categories according to their diameter: 0.7µm>(normal), 0.7–1.5 µm (weak enlarged), and 1.5 µm<(strong enlarged).
Feeding RNAi was performed as described previously [55]. To score embryonic lethality, young adult worms were placed on each RNAi plate and allowed to feed for 24 hr. Three worms from the original plate were transferred to a fresh RNAi plate and were allowed to lay eggs for 4–5 hr to score embryonic lethality. In an RNAi screen, we first used feeding RNAi clones on chromosome I and III in the Ahringer library to identify RNAi clones that cause high penetrance embryonic lethality in the obr quadruple mutant background, but not in the wild-type background. As a result, we found 22 RNAi clones that caused synthetic lethality with obr quadruple mutations (Table S2, Group A). These enhancer genes included the genes encoding vesicular transport-related proteins, such as apm-1 (μ subunit of AP-1), arf-1.2 (a homologue of ARF), vps-34 (Class III phosphatidylinositol 3 kinase) and vps-2 (ESCRT III). Therefore we next focused on genes whose homologues are known to regulate intracellular vesicular transport in other species (Table S2) [56] (MVB formation-related genes, small G proteins, components of COG complex, SNARE genes, SEC-1 family genes, coatmer proteins, and components of retromer complex). We tested 113 genes listed in Table S1 and identified another six genes that could enhance embryonic lethality of obr quadruple mutants (Table S2, Group B).
To obtain cholesterol-free conditions, agar was replaced by agarose S (Wako, Japan) and peptone was omitted from plates. An overnight culture of the OP50 strain of E. coli was grown on a LB medium. Bacteria were rinsed with M9 medium before use. Bacterial suspension were spread on cholesterol-free agarose plates. To perform RNAi under cholesterol depleted condition, bacteria were grown at 37°C to an O.D. of 0.5–0.8, induced with 0.4mM IPTG for 4hr, then concentrated and spread onto agarose plates containing 0.4mM IPTG. For feeding P0 animals, L4 hermaphrodites were plated directly on these plates at 20°C and their progeny were analyzed.
Fluorescence images were obtained using an Axio Imager M1 (Carl Zeiss MicroImaging Inc., Japan) microscope equipped with a digital CCD camera. Confocal images were obtained using a Zeiss LSM510 META confocal microscope system (Carl Zeiss MicroImaging).
HeLa cells were grown in DMEM, 10% fetal bovine serum (FBS), 100 U/ml penicillin, 100 mg/ml streptomycin, and 2mM L-glutamine. Cells were transiently transfected for 24–36 hr with cDNA constructs in complete medium using LipofectAMINE 2000 (Invitrogen, San Diego, CA, USA). Transfections were carried out according to the manufacturers' instructions. To perform RNAi, the cells were transfected for 48 hr with 20nM ORP1L-specific (sense strand GGACGAAAGGAGUUGGUAAdTdG) or control siRNA (Nippon EGT, Japan) using Lipofectamine 2000 (Invitrogen, San Diego, CA, USA).
A glutathione S-transferase–ORP1L fusion protein corresponding to amino acids 428–553 in the ORP1L protein was expressed in E. coli BL21 (DE3), purified by affinity chromatography on glutathione-Sepharose 4B (Pharmacia AB, Uppsala, Sweden), and used for immunization of New Zealand White rabbits according to a standard protocol. The ORP1L antiserum were purified by using an Affi-Gel (BIO-RAD, Japan) to which the antigen fragment had been coupled. The antibody were used for immunoblotting in 1 ∶ 10 dilution.
HeLa cells cultured on plastic cover glass (Celldesk LF1, Sumitomo Bakelite inc, Tokyo, Japan) in 24-well culture plates were fixed with 2.5% glutaraldehyde in 0.1 M phosphate buffer (pH 7.4) for 2 hr. Cells were post-fixed in 1% OsO4 in the same buffer for 1 hr, and dehydrated with a series of ethanol and embedded in epon. After the resin hardened, Celldesk was removed from the epon block. Ultra-thin sections were cut horizontally to the bottom of Celldesk, stained with uranyl acetate for 60 minutes, stained with lead citrate solution for 1 min, and observed under a Hitachi H-7600 electron microscope. For quantitative analyses, electron micrographs were taken at a magnification of 12,000. The cytoplasmic area and the number and diameter of MVBs were determined. Ten cell profiles were taken from each Celldesk, and three samples were analyzed (a total of 30 cells). C. elegans were pre-fixed with 4% paraformaldehyde and 1% glutaraldehyde in 0.1 M phosphate buffer (pH 7.4). Samples were then cut into small pieces, fixed again with 2% paraformaldehyde and 2% glutaraldehyde in the same buffer, and post-fixed with 2% osmium tetroxide in phosphate buffer for 4 hrs. Afterwards, fixed specimens were dehydrated in a graded series of ethanol and embedded in Quetol 651 epoxy resin. Ultrathin (80 to 90 nm-thick) sections obtained by ultramicrotomy were stained with uranyl acetate for 15 minutes and with modified Sato's lead solution for 5 mins. TEM observation was performed using a JEOL JEM-1200EX electron microscope.
Synchronized first-stage larvae (40,000 worms) were cultured with 6 µCi of [14C]-cholesterol (54 mCi/mmol; American Radiolabeled Chemicals, Inc. St. Louis, U.S.A.) for 54 hr on cholesterol free agar plates (see above) and were harvested from the plates with M9 medium. Late endosomal/lysosomal fraction was then prepared using the lysosome isolation kit (Sigma). Briefly, worms were homogenized using a Dounce homogenizer device and the lysates were subjected to centrifugation at 1,000×g to remove the nuclei. The post nuclear supernatant was subjected to centrifugation at 20,000×g to pellet the membranes, yielding the crude membrane fraction. The crude membrane fraction was resuspended in extraction buffer and subjected to density gradient ultracentrifugation at 150,000×g on an 8–27% Optiprep gradient for 4 hr (Lysosomal Isolation Kit, Sigma-Aldrich). 250 µl fractions were collected from the bottom of the tube with a peristaltic pump. The resulting fraction was treated with 250 mM calcium chloride to remove residual mitochondria and rough ER. Aliquots were assayed for lipid analysis, and the remaining material was processed for immunoblotting. Lipids were extracted by hexane, and were separated by one-dimensional TLC on silica gel 60 plates (Merck Biosciences) in chloroform-methanol (24∶1). Cholesterol was identified by comigration with known standard. Cholesterol ratio of late endosome/lysosomal fraction (fraction 7 and 8) was expressed as the percentage of radioactivity of 20,000×g ppt.
|
10.1371/journal.pgen.1006222 | KdmB, a Jumonji Histone H3 Demethylase, Regulates Genome-Wide H3K4 Trimethylation and Is Required for Normal Induction of Secondary Metabolism in Aspergillus nidulans | Histone posttranslational modifications (HPTMs) are involved in chromatin-based regulation of fungal secondary metabolite biosynthesis (SMB) in which the corresponding genes—usually physically linked in co-regulated clusters—are silenced under optimal physiological conditions (nutrient-rich) but are activated when nutrients are limiting. The exact molecular mechanisms by which HPTMs influence silencing and activation, however, are still to be better understood. Here we show by a combined approach of quantitative mass spectrometry (LC-MS/MS), genome-wide chromatin immunoprecipitation (ChIP-seq) and transcriptional network analysis (RNA-seq) that the core regions of silent A. nidulans SM clusters generally carry low levels of all tested chromatin modifications and that heterochromatic marks flank most of these SM clusters. During secondary metabolism, histone marks typically associated with transcriptional activity such as H3 trimethylated at lysine-4 (H3K4me3) are established in some, but not all gene clusters even upon full activation. KdmB, a Jarid1-family histone H3 lysine demethylase predicted to comprise a BRIGHT domain, a zinc-finger and two PHD domains in addition to the catalytic Jumonji domain, targets and demethylates H3K4me3 in vivo and mediates transcriptional downregulation. Deletion of kdmB leads to increased transcription of about ~1750 genes across nutrient-rich (primary metabolism) and nutrient-limiting (secondary metabolism) conditions. Unexpectedly, an equally high number of genes exhibited reduced expression in the kdmB deletion strain and notably, this group was significantly enriched for genes with known or predicted functions in secondary metabolite biosynthesis. Taken together, this study extends our general knowledge about multi-domain KDM5 histone demethylases and provides new details on the chromatin-level regulation of fungal secondary metabolite production.
| In this work we monitored by proteomic analysis and ChIP-seq the genome-wide distribution of several key modifications on histone H3 in the model fungus Aspergillus nidulans cultivated either under optimal physiological conditions (active growth) or less favourable conditions which are known to promote the production of secondary metabolites (SM). When we correlated the chromatin status to transcriptional activities in actively growing cells we found that the silenced SM gene clusters are flanked by heterochromatic domains presumably contributing to silencing but that the bodies of the clusters only carry background levels of any of the investigated marks. In nutrient-depleted conditions, activating marks were invading some, but by far not all transcribed clusters, leaving open the question how activation of these regions occurs at the chromatin level. Surprisingly, a large number of these gene clusters actually depend on KdmB for normal activation and it will be interesting to see in future how this protein thought to mainly act as repressor by removing positive H3K4m3 marks switches gears to activate transcription directly or indirectly.
| Chromatin is the natural substrate for all eukaryotic nuclear processes such as transcription, replication, recombination or DNA repair. Chromatin structure is necessarily dynamic and the underlying mechanisms involve remodeling of nucleosomes as well as depositing and removing posttranslational modifications on N-terminal and central residues of histones proteins (HPTMs) present in the nucleosome octamer [1–4]. Some of these histone marks, such as acetyl groups on lysines, profoundly influence the chromatin landscape by neutralizing the positive charge of histones thereby weakening the interaction between nucleosomes and DNA and increasing chromatin accessibility [5]. HPTMs also work indirectly by providing binding sites for chromatin-associated proteins that promote or inhibit specific genomic functions. Notably, many HPTMs recruit additional chromatin-modifying enzymes that add new or remove existing marks, enabling cells to dynamically regulate chromatin structure in response to environmental or developmental cues. Fungi have served as model systems for chromatin studies and in many basic mechanisms they are similar to higher eukaryotes but in some aspects they are quite different and this fact allows evolutionary insights into the development of chromatin regulatory systems (reviewed in [6–8]). For example, there is ground-laying work from the filamentous ascomycete Neurospora crassa, where the molecular machinery relating heterochromatin formation and DNA methylation was deciphered [9–12]. Similar to animals also in N. crassa Heterochromatin Protein 1 (HP1), docks on di- or trimethylated lysine-9 on histone H3 (H3K9me2/3) to promote heterochromatin formation [13, 14] and in addition is important to maintain H3K27me3, another repressive mark, at facultative heterochromatin [15, 16]. This mark was found to span 6.8% of the fungal genome [17] corresponding to over 700 transcriptionally repressed genes, some of which are upregulated upon deletion of the H3K27 methyltransferase [16, 17]. While H3K27 methylation and elements of Polycomb Repressive Complex 2 (PRC2) responsible for depositing this mark are present in Neurospora and the Fusarium group of fungal pathogens (see below) this silencing mechanism has not been detected in Aspergillus species [18]. In addition, DNA methylation has not been found in the Aspergilli although a cytosine methyltransferase is functionally expressed in A. nidulans and has a role in regulating sexual development [19].
Mycotoxins, antibiotics, pigments and other low molecular weight natural products are summarized under the term of secondary metabolites (SMs). The Fusarium and Aspergillus genera are large groups of fungi comprising important plant and animal pathogens and they all produce (SMs) at certain developmental stages or under conditions of growth restriction, nutrient limitation and environmental stress (reviewed in [20–23]). It was shown initially in Aspergillus nidulans by genetic analysis that expression of the corresponding SMs biosynthetic genes, which are usually organized in gene clusters, is under chromatin control (reviewed in [24]). Under conditions of active growth SMs genes are silenced by H3 deacetylation [25, 26] as well as by the H3K9 methylation machinery of ClrD (KMT1/ DIM-5 homolog) and the hpo homolog HepA [27]. Interestingly, H3K4 methylation and a subunit of the COMPASS complex which are usually known to be associated with gene activation, also contribute to silencing although this has only been observed for a small subset of SM genes [28]. Several recent studies in a number of other fungi have implicated heterochromatin as a regulator of secondary metabolism and the production of virulence factors. In the plant pathogens F. graminearum (wheat and maize pathogen) and F. fujikuroi (rice pathogen) as well as in the fungal endophyte Epichloë festucae, H3K9me3 and H3K27me3 regulate expression of specific gene clusters responsible for the production of secondary metabolites [20, 23, 29–32]. H3K9me3 and HP1 were also shown to negatively regulate other virulence factors such as genes encoding small secreted proteins (SSPs) in Leptosphaeria maculans [29].
How HPTM patterns change as SM clusters switch from a repressed state to an active state is not completely understood. The requirement of histone H3 and H4 acetylation for SM gene expression is well documented in Aspergillus species through HDAC inhibitor studies and SAGA- complex mutants [33–35]. Interestingly, co-cultivation of A. nidulans cells with Streptomyces rapamycinicus led to an anomalous activation of several SM genes in the fungus [36] and this process is correlated with increased H3 acetylation of the corresponding genes and strictly dependent on GcnE, the catalytic subunit of the A. nidulans SAGA acetylation complex [37]. Also in F. fujikuroi, activation of the GA, bikaverin and fumonisin clusters was correlated with increased acetylation of H3K9 [38].
In contrast to acetylation, the role of histone methylation in fungal SM gene expression is much less clear. In F. graminearum, silent SM clusters are highly enriched for repressive H3K27me3, whereas trimethylated H3 lysine 4 (H3K4me3), an activating mark, is apparently excluded. Upon deletion of the H3K27 methyltransferase kmt6, the silent fusarin C and carotenoid clusters are activated, but H3K4me3 does not accumulate in these clusters [30]. A similar situation was shown in F. fujikuroi where increases in H3K4me2 were only observed in two genes of the gibberellin (GA) cluster. Similar to the case for H3K4me, expression of SM cluster genes in F. graminearum was not associated with increased H3K36me3 [30]. In contrast, H3K36me3 was gained for the sterigmatocystin (ST) and several other SM gene clusters in A. nidulans during activation [18, 39, 40].
H3K4me3 is an HPTM with important roles in transcription and this mark is generated by the COMPASS (Complex associated with Set1) protein complex containing the Set1 methyltransferase catalytic subunit in addition to several regulatory and scaffold proteins [41]. COMPASS is not essential in A. nidulans although synthetic lethality of Set1 and Swd1 subunits was found with mutations in mitotic regulators [42]. Generally, H3K4me3 has been shown to be recognized by three different domains associated with proteins of various functions. One recognition module is the PHD domain, present for example in the “Inhibitor of Growth” (ING) protein, which recruits histone acetyltransferase (HAT) and deacetylase (HDAC) complexes [43, 44]. H3K4me3 is also recognized by the double TUDOR domain of JMJD2A, a JmjC family demethylase that removes methyl groups from di- or trimethylated H3K9 [45] and by the tandem chromodomain of CHD1, an ATP- dependent nucleosomal remodeler [46] recently shown to be necessary for inhibition of intragenic initiation or initiation from cryptic promoters and thus maintaining normal transcript elongation [47]. Accordingly, H3K4me3 plays a central role in the chromatin regulatory network. Usually, H3K4me3 peaks at the transcription start sites (TSSs) and its occurrence is correlated with gene expression [48]. However, the Set1 protein also displays some moonlighting activities as it recruits deacetylase activity independently from the H3K4me3 mark and subsequently promotes heterochromatin formation and transcriptional repression at distinct loci in the fission yeast genome [49]. This evidently negative role of the COMPASS was also documented for regulation of SMs production in three different Aspergillus species carrying genetically engineered COMPASS mutations [28, 50, 51]. Silencing specific SM gene clusters might be related to previously documented subtelomeric silencing functions of the COMPASS complex [41] and mechanistically similar to the recently identified heterochromatin-promoting role in fission yeast [49].
Dynamic demethylation of lysine residues adds additional complexity to the modulation of transcription by lysine methylation [3, 52]. Recently we showed that KdmA, a JMJD2/JHDM3 family H3K9/36me3 demethylase [53, 54] can, in equal measure, positively and negatively influence gene expression in A. nidulans [18]. Here, we characterize another member of the JmjC demethylase family, KdmB, which acts on H3K4me3 in vivo, thus is assigned to the Jarid group of enzymes. Jarid (JMJ–AT-rich interacting domain-containing protein) subfamily demethylases have been shown to target di- and trimethylated H3K4 and are therefore generally considered to be repressors of gene transcription, though they can also act as activators [55]. For example the function of mammalian RBP2 (retinoblastoma binding protein 2, alias JARID 1A or KDM5A according to the new nomenclature [56]) in transcription regulation is context dependent. RBP2 represses transcription via H3K4me3 demethylation and association with an HDAC complex, however when associated with retinoblastoma protein (pRb), it activates certain genes in the mammalian genome [57]. Similarly, the D. melanogaster ortholog LID can repress transcription via H3K4me3 demethylation, however when associated with the MYC transcription factor, its demethylase activity is inhibited and consequently the LID-MYC complex mediates gene activation [58, 59]. These examples demonstrate that Jarid demethylases can act directly on their target genes in a context dependent positive or negative manner.
In this work we studied the Jarid-type demethylase in A. nidulans by reverse genetics and performed genome-wide HTPM profiling by mass spectrometry of histones, by ChIP analysis of H3K4me3, H3K9me3, H3K36me3 and H3 acetylation on K9 and K14 (H3Ac) modifications in wild type and compared the results with the KdmB mutant. We recorded these HPTM changes in parallel with the transcriptome under optimal physiological conditions promoting active growth (primary metabolism) as well as under stationary-phase conditions that lead to SM production (secondary metabolism). Comparison of ChIP-seq profiles with RNA-seq of the same cultures allowed us to correlate transcriptional changes with changes in chromatin landscapes across different conditions and genetic backgrounds. Histone proteomic analysis in wild type and the KdmB histone H3K4 demethylase mutant provided direct evidence for H3K4me3 as the dominant substrate for KdmB and confirmed that A. nidulans does not feature H3K27me3, the canonical facultative heterochromatic mark in other eukaryotes and responsible for SM gene silencing in a number of other fungi.
Based on the domain composition of the full length KdmB (AN8211) and detailed analysis of the amino acid sequences of the catalytic JmjC domains of histone demethylases from yeast to humans, KdmB was classified as a Jarid1-type histone H3 lysine 4 demethylase (Fig 1). Residues responsible for substrate recognition of Jarid demethylases are not known due to the lack of available crystallographic data, although the conserved amino acids required for substrate recognition in the JMJD2 subfamily of lysine K9 and K36 histone H3 demethylases (marked in green in Fig 1A) are not present in the Jarid group [60]. Domain analysis revealed that KdmB is more similar to the proteins from higher eukaryotes than from budding yeast. Specifically, we found that KdmB contains a putative ARID/Bright domain and a C5-HC2 zinc finger motif and an additional PHD domain at the C-terminus, which are both absent from the budding yeast homolog (Fig 1B).
To investigate the in vitro specificity of KdmB we heterologously expressed KdmB as a GST fusion protein in E. coli. KdmB has predicted molecular weight of 216 kDa but the resulting full size recombinant protein was not sufficiently soluble. Another construct producing a truncated KdmB protein without the second PHD domain, however, was readily soluble under native buffer conditions. This KdmB fusion containing residues 1 to 922 displayed an apparent mass of roughly 130 kDa (S1A Fig). In vitro demethylase assays (DeMt) were subsequently performed with purified GST-KdmB(1–922) and calf thymus histones as a substrate. Products of the DeMt reactions were detected with modification-specific antibodies by Western blot (S1B Fig). Under our assay conditions, we found a decrease in trimethylation signals for all three tested lysine residues (H3K4me3, H3K9me3 and H3K36me3) and the strongest reduction in abundance was seen in H3K9me3. Acetylation was not reduced by the enzyme, as expected. Consistent with KdmB being a JmjC-type demethylase, the activity of the GST-KdmB(1–922), fusion protein was dependent on the presence of the cofactors α- ketoglutarate and Fe2+ (S1B and S1C Fig). In our assay conditions we observed high standard deviations between independent replicates of H3K4me3 and H3K36me3-specific Westerns. This could be due to experimental variation in enzymatic activity of different batches of the purified recombinant enzyme.
The very broad substrate range of KdmB in vitro is unexpected for this Kdm5-family member because so far the identified and tested enzymes target either H3K4me2/3 (Jarid1 group enzymes) or H3K9/36me2/3 (Jmjd2 group). However, it is possible that the absence of PHD-finger 2, interacting proteins or the presence of the GST domain compromises substrate specificity in our assay. Although none of the KdmB orthologs identified so far demonstrated such broad substrate specificity in vitro [57, 63–66], the in vitro demethylase activity found in our assays suggests that this protein possesses histone demethylase activity.
To determine whether KdmB can act as a histone demethylase in vivo, we performed LC- MS/MS on acidic extracted histones from actively growing A. nidulans wildtype and kdmBΔ cells (see Materials and Methods for description of gene deletion procedure). In wildtype, mass spectrometry revealed that 71.3% of H3K4 peptides contain at least one methyl-group at the K4 position. We detected all three forms of methyl-H3K4 peptides and found that H3K4me3 is the most abundant (47.5% of total H3K4 peptides), followed by di-methylated (13.5%) and mono-methylated H3K4 (10.3%) (Fig 2).
Notably, our measurements revealed an almost 20% increase in global H3K4 trimethylation in the kdmBΔ strain (57% H3K4me3). Because the levels of H3K4me2, H3K4me1 and unmodified H3K4 were concomitantly decreased in the mutant in roughly the same range as H3K4me3 increased we concluded that in vivo KdmB primarily acts to demethylate H3K4me3. The MS results also revealed that in vivo KdmB does not target H3K36me3 as these levels remained constant in histones of kdmBΔ cells (S2 Fig). Interestingly, the overall low marking of H3K9 by trimethylation (1.53% of the mapped peptides) was further reduced (to 0.2% of the mapped peptides) in the mutant. This contrasts the in vitro assay results which showed a strong H3K9me3 demethylating activity of recombinant KdmB (S1 Fig). The further reduction of H3K9me3 in kdmBΔ cells might be attributable, however, to an increase in the opposing, positively acting H3K4me3 mark limiting the possibility to deposit or maintain H3K9 trimethyl marks in the target regions. Strikingly, in vivo, global H3 N-terminal lysine acetylation (H3K9ac/K14ac) was increased almost by 20% in the kdmBΔ strain at the expense of unmodified peptides of H3 which are reduced from 22% in the wild type to 6% in the mutant (S2A Fig). This more abundant histone acetylation could be the consequence of both stronger marking by acetylases and/or reduced deacetylation. The latter mechanism has already been reported in connection with KdmB homologs in mammals where RBP2 (Jarid1a) and PLU1 (Jarid1b) recruit the Rpd3S histone deacetylase complex [57, 67]. Altogether, our data demonstrate H3K4me3 demethylation activity of KdmB in A. nidulans cells and lack of this activity in kdmB deletion cells leads to a shift in modification equilibrium with more abundant positive (H3K4me3, H3Ac) and less negative (H3K9me3) marks.
To determine the genomic regions in which KdmB influences H3K4me3 levels we performed genome-wide ChIP analysis (ChIP-seq) in wild type and kdmBΔ strains with antibodies specific to H3K4me3 [28]. As our global histone analysis revealed a crosstalk of this modification to H3K9 trimethylation as well as to H3K9/K14 acetylation, we also included these marks in ChIP-seq. Although no changes occurred for H3K36 trimethylation at the level of bulk histones between WT and the kdmB mutant, we were interested if locus-specific differences occur and thus analyzed also this mark by ChIP-seq.
As previous studies from our lab and by others revealed a crucial function of chromatin structure and histone modifications on the regulation of secondary metabolite biosynthesis (SMB), we performed all subsequent RNA-seq and ChIP-seq experiments not only under the already described standard active growth conditions representing primary metabolism (PM; 17h liquid shake cultures, no nutrient limitation) but also under conditions promoting secondary metabolism (48h liquid shake cultures, nutrient depletion, see S3 Fig).
To monitor the distribution of the tested chromatin modifications along A. nidulans genes, we used chromosome IV as an example and plotted the wild type distribution of H3K4me3, H3K36me3, and H3K9/14ac across the promoters and open reading frames (ORFs) of all genes on this chromosome (Fig 3A). In this analysis, all genes are aligned to the predicted ATG (position 0) and read counts per million of mapped reads (CPM) are analysed in a 2 kb window starting with 500 bp of their 5´UTR and promoter sequences (-500) followed by 1500 bp of their coding region. This revealed that the pattern of modifications reflects the distribution observed in other model organisms including fungi [30, 68–70]. H3K4me3 was enriched in characteristic peaks spanning the first three nucleosomes (around 500 bp) of the coding region, whereas H3K36me3 was enriched near the 3’ regions of genes. Finally, H3 acetylation was enriched in the promoter, with highest levels apparent in the first nucleosome just downstream of the predicted translation start sites.
To explore the general relationship between H3K4me3 and transcription we quantified the average level of H3K4me3 in a 2 kb window around the predicted start codon of each gene (average CPM from -500 to +1500) and related this value to the average expression level (expressed as RPKM, reads per kilobase per million reads) of the corresponding gene in both culture conditions (PM and SM). In the resulting scatterplot (Fig 3B) two groups of genes became apparent, i.e. those that displayed high levels of H3K4me3 (log2 RPKM>5) and a second group that showed low to no H3K4 trimethylation (log2 RPKM≤5). Correlation of H3K4me3 levels with transcription of the corresponding gene revealed an overall positive correlation between H3K4me3 levels and transcript abundance (Fig 3B). This suggests that, similar to other well-studied models, H3K4 trimethylation is a marker for actively transcribed genes.
To better characterize the function of KdmB in the context of transcriptional regulation we next compared by ChIP-seq the distributions of four histone modifications in wild type and kdmBΔ (Fig 4) under active growth conditions (PM) and during SM. The kdmB deletion did not cause any gross phenotypic changes in the mutant strain which was rather similar to the wild type in growth rates and nutrient consumption (S3 Fig). ChIP-seq combined with RNAseq analysis revealed the H3K4me3 enriched domains which coincide with transcriptional activity. In the example shown in Fig 4 we noticed, on the gross genomic scale, an overlap between the positively acting marks H3K4me3, H3K36me3 and H3Ac. In contrast, repressing H3K9me3 marks are enriched mainly in pericentromeric and subtelomeric regions and a few isolated H3K9me3 blocks exist (on the left arm of chromosome IV, for example).
At the gross genomic scale the comparison of the chromatin landscape for H3K4me3 marks in chromosome IV between actively growing (17 h cultures) wild type and kdmBΔ cells did not reveal any obvious changes. Moreover, at this scale, no large domains were visibly changed for the other tested modifications (H3Ac, H3K36me3, H3K9me3). Because our mass spectrometry analyses uncovered increased H3K4me3 and H3Ac in the mutant, we reasoned that changes in the levels of these histone marks must occur at a subset of individual genes. To test this, we analyzed H3K4me3 levels in genes that were differentially expressed between wildtype and kdmBΔ. We first examined genes with low H3K4me3 levels [(log2 (RPKM)≤ 5] and found that 301 genes displayed higher expression levels in the wildtype (WT-up/Group 1, Fig 5A) suggesting that for this group KdmB is required for normal expression levels. In contrast, 501 genes had higher expression in kdmBΔ (kdmBΔ-up/ Group 2) which points to a repressing function of the protein in these loci. In the gene set featuring high H3K4me3 levels [(log2 (RPKM)> 5] we again identified both up- and down-regulated genes; 455 genes were expressed at higher levels in wild type (WT-up/Group 3) and 133 genes were expressed at higher levels in kdmBΔ (kdmBΔ-up/ Group 4).
The analysis showed that KdmB influences transcriptomes in both directions. For around 750 genes KdmB function is necessary for normal transcription, whereas for around 630 genes KdmB has a negative function. The repressive role of KdmB was found in both categories, i.e. on genes carrying low (kdmBΔ-up/G2) or high (kdmBΔ-up/G 4) H3K4me3 levels. Significantly, the group with normally low H3K4me3 (G2) displayed a marked increase in this histone mark in the kdmBΔ mutant concomitantly with increased transcript levels. One representative of this group is shown in Fig 5C for a gene (locus AN6321) which is basically not transcribed in the wild type but which gains both positive marks and transcripts in the kdmBΔ strain. Although we have not tested this directly, the strict correlation between increased H3K4me3 levels and transcription, along with the in vitro K4me3-demethylase activity of KdmB, suggests that at least some of these loci are direct targets of KdmB. A slightly different situation was found for the second gene set highly decorated with H3K4me3. Although a subset of these genes showed increased expression in the kdmBΔ mutant (kdmBΔ-up/G 4), this was not accompanied by an increase in H3K4me3 probably due to the already very high K4 methylation levels in the wild type. Consequently, a further increase would hardly be possible and thus the effect of kdmB deletion on H3K4 trimethylation is more subtle compared to genes generally not heavily marked by H3K4me3.
In contrast to the repressive function, KdmB also seems to have a positive role in transcription. kdmB deletion led to reduced expression of 750 genes belonging to both low (WT-up/G1) or high (WT-up/G3) H3K4me3 groups, accompanied by lower H3K4me3, on average, in the mutant. Based on these correlations we can conclude that KdmB function is required for normal expression of these roughly 750 genes, but whether KdmB directly targets these loci or indirectly affects transcription via the transcriptome network remains to be determined.
We also constructed metaplots of H3K4me3 distributions under SM conditions (S4 Fig). Under these growth conditions a similar correlation was observed, i.e. H3K4me3 levels were reduced in genes that were downregulated in kdmBΔ, whereas the genes upregulated in the mutant showed no drastic change (in the high H3K4me3 group) or somewhat higher H3K4 trimethylation. However, in locus-specific analysis by RNA-seq and ChIP-seq (see below), we also found some transcriptionally silent regions with high H3K4me3 as well as some highly transcribed genes with very low levels of this mark (see analysis below) indicating that specific genomic regions exist in which this general positive correlation between H3K4me3 and transcriptional activity does not apply.
Our initial correlation analysis of H3K4me3 and transcription revealed that among genes requiring KdmB for full transcription, the category of SMB genes was significantly enriched (p < 0.05). In further analysis, PM and SMB genes were separated based on functional categories and this bioinformatic approach created a large group of genes (5676 genes) predicted to be involved in general cellular functions and metabolism (category “cell structure and function” abbreviated CSF) and a smaller group of 149 genes predicted to be involved in SMB (category “SM clusters”). [71, 72]. Fig 6 shows that under PM conditions, approximately 5% of genes involved in CSF and 15% of genes assigned to SMB were affected by the kdmB deletion. The majority of A. nidulans SM cluster genes are not under/ during PM conditions, thus it is not surprising that differential expression of SM genes is largely restricted to the 48h cultures. Interestingly, several genes belonging to a gene cluster with a so far unidentified product were highly upregulated in the mutant at this 17h time point and this transcriptional pattern will certainly facilitate the future identification of the product derived from this predicted SM cluster.
In contrast to the mild effect on SM gene expression during PM conditions, KdmB-deficient cells showed significantly altered patterns of gene expression when cells were collected from cultures under SM conditions. Over 50% of all predicted SM genes were misregulated in the mutant. The majority of these displayed lower expression, while approximately 10% of SM genes showed higher expression in the kdmBΔ strain (Fig 6A, upper panel). In contrast, during the same culture condition only ~10% of genes not involved in SM were differentially transcribed in kdmBΔ. These data demonstrate that KdmB is required for normal induction of the majority of SM clusters in A. nidulans. It is probably relevant to note that the defect in SM cluster activation in the kdmB mutant is not due to a lack of wide-domain activator expression as laeA, veA, velB and velC are normally transcribed in the mutant (changes between WT and kdmBΔ log2 ≤ ± 1,7).
The lower panel of Fig 6A presents the number of deregulated genes within each category and time point. During primary metabolism (17h) KdmB function is required for a relatively small number of genes (143 genes in CSF and 10 genes in SM). In contrast, in the nutrient limited 48h cultures gene expression profiles are changed considerably in the mutant: 598 genes (401 CSF and 97 SMB genes) require KdmB function for normal expression and 569 genes (547 CSF and 22 SMB genes) are negatively influenced by the regulator. These data suggest that KdmB is primarily required during the stationary phase and obviously plays an important role for the expression of the majority (97 out of 149 of genes involved in SMB
We also tested whether transcriptional changes in kdmBΔ were correlated with changes in SMB biosynthesis. For this we performed HPLC-MS/MS analyses of cultures grown in two different media, i.e. in conventional minimal medium used throughout the studies (AMM) and in a specialized SM-promoting ZM medium (see Materials and Methods section). The comparison of WT and mutant culture extracts, grown in AMM medium, revealed a strongly decreased production of sterigmatocystin and emericellamides C and D (Fig 6B, left chromatograms) but other metabolites such as emodin and its derivatives were increased in kdmBΔ (Fig 6B, chromatograms a and c). However, our RNA-seq data showed that genes encoding for enzymes involved in emodin biosynthesis embedded in the mdpL-A monodictyphenon pathway are not differentially expressed between WT and the kdmB mutant (S13 Fig). To accommodate these differences, we speculate that the decreased transcription of other secondary metabolite clusters, such as the sterigmatocystin cluster, may lead to higher levels of available emodin precursors, such as acetyl-CoA and malonyl-CoA, and thereby to an increased synthesis of emodin derivatives. ZM culture extracts revealed reduced levels of orsellinic acid in kdmBΔ (Fig 6B, right chromatograms), consistent with our RNA-seq data showing a decreased expression from the orsellinic acid gene cluster in the kdmB deletion (S10 Fig). The complete list of identified metabolites together with LC-MS and LC-MS2 data are shown in the S3 Table.
We also carried out correlation analyses between H3 acetylation and H3K4 methylation in genes which are differentially regulated in the kdmB mutant (S5 Fig). For those genes where KdmB is required for full expression and which are consequently higher transcribed in the wild type (categories WT-up/G1 and G3) H3 acetylation levels are also higher, independently of H3K4 trimethylation. The same is true for genes which are negatively influenced by KdmB (kdmBΔ-up/G2) but only if H3K4me3 levels are low. On the contrary, genes with high H3K4me3 levels under negative KdmB influence (kdmBΔ-up/G4), acetylation levels are lower than in the wild type despite higher expression of the corresponding genes in this group. The molecular basis of this effect has not been investigated further in this study but it would certainly be interesting to determine if KdmB impacts acetylation indirectly or directly through protein interactions with HDACs or HATs.
We also examined a possible influence of KdmB on the distribution of H3K36me3 in genes expressed under primary metabolic conditions (S6 Fig). We have previously shown that this mark is associated with active transcription and that, at some tested loci, the trimethylated H3K36 state is removed by KdmA, another A. nidulans JmjC-containing protein belonging to the KDM4 family [18]. The vast majority of A. nidulans genes are highly decorated by this mark under PM conditions (around 9,100 genes). We did not observe significant differences in the levels or in the distribution of this mark in the kdmBΔ strain neither in this group nor in the group carrying low H3K36me3 levels (1249 genes. This indicates that KdmB is not a demethylase of trimethyl-H3K36 in vivo. Around 13% of the 9,100 genes are de-regulated in the kdmB mutant strain but despite this differential expression there are no significant differences in the associated H3K36 trimethylation levels. This means that, at least for the gene set in which KdmB influences transcription, it does not do this via manipulating H3K36me3 levels.
The genome-wide distribution pattern of H3K9me3 supports the previously reported low levels of H3K9 trimethylation in A. nidulans wild type cells where we found approximately 1.5% of peptides carrying this mark. [18]. In ChIP-seq, the H3K9me3 pattern correlates with AT-rich domains flanking the subtelomeric regions but also includes sites along the chromosome arms, as shown on the left arm of chromosome IV (Fig 4). Inspection of H3K9me3-associated regions revealed that many SMB gene clusters such as the penicillin (S7 Fig), sterigmatocystin (S8 Fig), austinol (S9 Fig), orsellinic acid (S10 Fig) and terrequinone A (S11 Fig) are flanked by H3K9me3 domains at either one (e.g. the ST and TDI clusters) or at both sides (e.g. the PEN cluster) of the cluster. Whether these structures are functionally relevant for the regulation of SM gene clusters remains obscure but possible since deletion of the H3K9 methyltransferase gene clrD or of hepA, the gene coding for the protein recognizing H3K9me3, lead to up-regulation of genes within these clusters [27]. The observation that many H3K9me3 blocks are found in close proximity to SMB gene clusters raises the possibility that higher order chromatin structures or a as yet unstudied set of modifications may be important for normal regulation of SM gene expression, consistent with prior genetic analyses [27, 29, 31, 32].
However, we have also found several SM clusters such as the asperthecin (S12 Fig) and monodictyphenone (S13 Fig) cluster without such H3K9me3 borders. Interestingly, these clusters are not activated under the standard SM growth conditions used here (48 h cultures and nutrient deprivation). Instead, the MDP cluster is only expressed to detectable levels in a strain lacking the CclA regulatory subunit of the COMPASS complex which is responsible for H3K4 di- and tri methylation [28] and APT is highly expressed only in an A. nidulans mutant lacking SUMO, the small ubiquitin-related modifier protein known to profoundly regulate chromatin structure and function [73, 74]. Hence, absence of the H3K9me3 blocks might be correlated with special requirements for activation whereas SMB gene clusters activated under standard SMB conditions feature H3K9me3-flanking domains.
Correlation of H3K4me3 with transcriptional activity suggested that SMB gene clusters carry low levels of this mark even when they are strongly transcribed (see Fig 3). Inspection of ChIP-seq data from these regions confirmed that H3K4me3 is underrepresented in such clusters, as shown in the example of the well-studied sterigmatocystin gene cluster (Fig 7).
When in conditions of primary metabolism, cluster genes are silent and are not associated with H3K4me3 but surprisingly, this mark is not established at most genes even when the cluster is fully activated (Fig 7). Eventually, a single strong H3K4me3 peak occurred around the 5´end of stcD, a gene coding for an unknown function but co-regulated with the sterigmatocystin biosynthesis cluster [75]. Qualitatively, the two other tested activating marks H3K9/K14 acetylation and H3K36 methylation seem to increase around 5´ and 3´ends of the ORFs, respectively, in the activated cluster. A very similar picture emerged from the analysis of other clusters (S9–S13 Figs) and in each case, as expected, no major differences in the H3K4me3 profiles became apparent between the kdmBΔ mutant and the wild type.
To test our qualitative impression for significance we performed statistical analysis of our ChIP-seq data for differences in H3K4me3, H3Ac and H3K36me3 marks in PM and SMB conditions in the wild type and in the kdmBΔ mutant. The bioinformatic separation into “Cell structure and function” and “SM clusters” categories applied for the transcriptome was also kept for ChIP-seq data analysis. The statistical analysis of ChIP data revealed a striking difference in H3K4me3 levels between the two categories. As seen in the box blot in Fig 8 genes involved in SM production are significantly less decorated by H3K4me3, regardless of the culture condition or the presence of KdmB.
Moreover, the pattern is not significantly changed in the kdmBΔ strain, suggesting that KdmB does not promote SM gene expression by directly regulating H3K4me3 within SM clusters. SM cluster activation leads to subtle increases in the level of H3K4me3, H3Ac or H3K36me3 associated with SM cluster genes, and this increase is not visible in kdmBΔ (Fig 5A). In summary, our ChIP-seq data revealed that A. nidulans SM clusters in comparison to genes involved in the cell structure and function have relatively low levels of activating histone marks, especially H3K4me3 and H3K36me3.
Di- and tri-methylation of histone H3K4 is associated with transcriptionally active chromatin. Removal of this modification is accomplished by members of the KDM5-family demethylases, typically resulting in repression of the targeted locus. In fact, the first characterized H3K4 demethylases LID2 [59] and RBP2 [57] were identified as transcriptional repressors. However, these proteins and all KDM5 members are composed of multiple domains which are necessary for the diverse functions these regulators play. For example, demethylase activity of KDM5 is only one of the important functions required for Drosophila development [76] [77]. In addition, some domains have been associated with gene activation, for example, mammalian Jarid1a is recruited to the Per2 circadian gene promoter where it inhibits HDACs function and promotes transcription [78]. We also found in our study that deletion of KdmB has both activating and repressing effects. We found a genome-wide 20% increase in acetylated H3 N-termini and increased transcription of around 630 genes under standard growth conditions (nutrient sufficiency, primary metabolism) in strains lacking KdmB. These results provide evidence that the protein functions as a repressor that is able to remove H3K4me3 and perhaps recruit HDACs. Our data also demonstrate that KdmB is an H3K4me3 demethylase. The protein removed this modification in vitro and genes that are overexpressed in a KdmB-deficient mutant show increased H3K4me3. Unfortunately, from our data, we cannot deduce which part of this gene set is directly targeted by KdmB and which may be indirectly silenced through transcriptome network effects. ChIP analysis of KdmB tagged versions will be able to clarify this point in future.
On the other hand, for around 750 A. nidulans genes KdmB is required for full transcription. It is possible that KdmB mediates activation directly via one or more domains such as the potentially DNA-binding Zn-finger or ARID domains or the methylated histone binding PHD domain. However, deciphering how KdmB promotes transcription requires further investigation. Strikingly, the majority of the genes under positive KdmB control are related to the production of secondary metabolites. These small natural products are defense and signaling molecules of fungi produced during development, under stressful or nutrient-limiting conditions [23] and it is interesting that a general chromatin regulator such as KdmB takes up this specialized function in metabolism. We have shown that KdmB regulates (directly or indirectly) almost 5% of the genome during PM and over 10% during SM. The activation signal for the induction of genes involved in SMs production is transmitted via the so-called velvet activation complex containing also a protein termed LaeA that influences chromatin structure [27]. It will be interesting to determine if KdmB functions in this pathway. It is possible that KdmB could regulate SM gene expression by demethylation of SM regulatory proteins. Recently it was shown that several JmjC family demethylases can target non-histone substrates; however this function, to our best knowledge, has not been demonstrated for Jarid family demethylases [77, 79].
This role of KdmB in SMB gene activation appears to be independent of its histone demethylase enzymatic activity. In kdmBΔ, low levels of H3K4me3 and H3 lysine acetylation in SM gene clusters under activating conditions are likely the consequence of lower transcription at these loci.
One of the most striking features of silent A. nidulans SM clusters is a very low abundance or virtual absence of the four investigated histone marks within the borders of these gene clusters. At the moment we cannot exclude the possibility that other histone marks define the chromatin landscape within and around SM gene clusters. A large number of SM clusters, as exemplified here for sterigmatocystin (ST), penicillin (PEN), orsellinic acid (ORS), teraquinone (TDI), derivative of benzaldehyde 1 (DBA), austinol (AUS) and asperthecin (APT) are located in regions for which H3K4me3, H3K36m3 or H3Ac can hardly be detected. Even the monodictiphenone (MDP, S13 Fig) and asparthecin clusters (APT, S12 Fig), which are located within euchromatic regions, display a low-abundance of HPTMs. A distinguishing feature of these two clusters, which are not activated by the conventional SMB culture conditions applied here, is the lack of flanking by H3K9me3 domains which are characteristic for the majority of the analyzed SMB gene clusters (S7–S13 Figs). Although truncated KdmB demethylates H3K9me3 in vitro (S1 Fig) we did not see increased levels of this mark in the kdmBΔ mutant neither at specific loci nor at the genomic scale. This strongly suggests that in vivo H3K9me3 is not a target of KdmB. Moreover, no KDM5-type H3K9 demethylases have been described in other ascomycete fungi including S. pombe or N. crassa. In A. nidulans we even see genome-wide reduced levels of this mark in the mutant and it is likely that this reduced H3K9me3 is an indirect consequence of increased H3K4me3 or increased H3 acetylation. In addition to H3K9me3, the chromatin landscape changes slightly also for the other tested marks when the silent SM gene clusters are activated. The majority of the genes in these clusters gain H3K36me3 at their 3’ region and H3Ac at their 5’ region. Marking by H3K4me3, however, only occurs for a limited number of genes within these clusters such as some selected genes within the ST cluster (Fig 8) or the orsD gene positioned within the ORS cluster (S10 Fig).
A similar situation was recently reported in two different Fusarium species in which the H3K4 dimethylation level (H3K4me2) was compared to SMB gene transcription. In the rice pathogen F. fujikuroi, only two out of seven highly transcribed genes in the gibberellin cluster were significantly decorated with H3K4me2 [38] and also in F. graminearum, a pathogen of wheat and maize, genes in the fusarin C or the carotenoid biosynthesis clusters carried only background levels of this mark [30]. It is still remarkable that Liu and colleagues found an essential function of the H3K4 methyltransferase Set1 for the expression of the TRI gene cluster coding for deoxynivalenol biosynthesis [80] and in this latter study, H3K4me2 was clearly enriched over the background level and positively correlated with active transcription. Additionally, in contrast to our study, Connolly et al. [30] found H3K36me3 enrichment across the whole chromosome independent of transcriptional activity. These comparisons already highlight the high diversity of chromatin-based regulation in SMB gene expression within one single organism and even more between different organisms and which histone modifications are determining whether the SMB signal is transmitted to the transcriptional machinery or not. Surprisingly, the H3K4 demethylase KdmB plays an essential role in the activation process although this histone mark is not present in the targeted regions.
A. nidulans strains used in this study are listed in S1 Table. Experimental strains were obtained by transformation into an nkuAΔ strain, which reduces the frequency of non-homologous integration [81], or by sexual crosses. Genetic analysis was carried out using techniques as described by Todd et al. [82]. DNA transformation of A. nidulans was performed according to [83]. KdmB deletion cassettes were constructed using DJ PCR [84] with the Aspergillus fumigatus riboB gene as selectable marker, riboB+ transformants were recovered after transformation into nkuA strains. Southern analysis confirming the deletion of kdmB was performed as described elsewhere [18, 40]. AMM minimal media, complete medium, supplements and growth conditions were as described by Todd et al. [82]. ZM1/2 medium (molasses 0.5%, oatmeal 0.5%, sucrose 0.4%, mannite 0.4%, D-glucose 0.15%, CaCO3 0.15%, edamine 0.05%, (NH4)2SO4 0.05%) was used for promoting SM biosynthesis in the experiments analyzing metabolites by HPLC-MS/MS [85]. For LC-MS/MS, RNA-seq, ChIP-seq, DeMt assay, HPLC-MS/MS spores in concentration 4*106/mL were inoculated into 200 mL liquid AMM and incubated at 180 rpm, 37°C for 17 h and 48 h. For SM cluster gene expression, ChIP and HPLC- MS/ MS analysis 10 mM sodium nitrate otherwise ammonium tartrate at 10 mM was added as nitrogen source.
kdmB cDNA was amplified using RevertAid Premium Reverse Transcriptase (Thermo Scientific, EP0732) and specific primers. Full length (1717 aa) and truncated versions (residues 1–922) cDNAs were cloned into pGEX-4T1 expression vector, sequenced, transformed and expressed in Rosetta cells. GST-KdmB (1–922) was purified using glutathione Sepharose 4B (GE Healthcare). Demethylation assay was performed as previously described [66]. Purified KdmB was incubated with calf thymus histones (Sigma, H9250) in demethylase reaction buffer (20mM Tris-HCl pH 7.2, 150 mM KCl, FeSO4 20 μM, α-ketoglutarate 500 μM, ascorbic acid 500μM, ZnCl2 1μM) for 3 to 10 h at 37°. Reaction was stopped by boiling for 5 minutes with 100 mM DTT Laemmli buffer; changes in lysine methylation were measured by Western blot with the specific antibodies (see ChIP section). The demethylation in vitro assay and Western blot were performed three times; negative controls were incubated without the cofactors for JmjC proteins (Fe2+ and α-ketoglutarate).
Mycelia from o/n liquid submerged cultures were harvested by filtration and frozen in liquid nitrogen. Histones were acid extracted as previously described [86], suspended in Laemmli’s SDS sample buffer and quantified with Pierce BCA Protein Assay (Thermo). 15 μg of purified histones, 1 μg of recombinant Xenopus laevis H3 as a negative control (Milipore, 14–441) and 2 μg calf thymus histones (Sigma, H9250) as a positive control were separated on 15% SDS-PAGE gel and subsequently transferred to nitrocellulose membrane (GE Healthcare) by electroblotting. Relevant histone modifications were detected with primary antibodies specific to H3K4me3 (Abcam, 8580), H3K9me3 (Active Motif, 39161), H3K36me3 (Abcam, 9050), histone H3 C-terminus (Abcam, 1791), H3Ac (pan-acetyl) (Millipore, 06–599) and anti- rabbit (Sigma, A0545) and anti- mouse (Sigma, A9044) HRP conjugated secondary antibodies. Chemiluminescence was detected with Clarity ECL Western Substrate and ChemiDoc XRS (Bio- Rad). Densitometric quantification of Western blot signals from demethylase reactions were performed with the ImagJ software. In total three independent blots of demethylase and the control reaction (without the cofactors) were quantified. Signal of respective HPTM were normalize to histone H3 C-term. Subsequently the signal of the control reaction was set to a value 1, consequently the presented results are the fold change to the control reaction. For MS analysis relevant histone H3 protein bands were cut out and digested in gel. The proteins were S-alkylated with iodoacetamide and digested with ArgC (Roche). The peptide mixture was analysed using a Dionex Ultimate 3000 system directly linked to a Q-TOF MS (Bruker maXis 4G ETD) equipped with the standard ESI source in the positive ion, DDA mode (= switching to MSMS mode for eluting peaks). MS-scans were recorded (range: 150–2200 Da) and the 6 highest peaks were selected for fragmentation. Instrument calibration was performed using ESI calibration mixture (Agilent). For separation of the peptides a Thermo BioBasic C18 separation column (5 μm particle size, 150*0.360 mm) was used. A gradient from 95% solvent A and 5% solvent B (Solvent A: 0.1% FA in water, 0.1% FA in ACCN) to 32% B in 45 min was applied, followed by a 15min gradient from 32% B to 75% B that facilitates elution of large peptides, at a flow rate of 6 μL/min.
The fungal cultures were incubated in triplicates, RNA from each technical replicate was pooled and each experiment was performed twice to obtain two biologically independent sets with two technical replicates for each strain and condition. Illumina sequencing libraries were made from RNA samples according to TruSeq RNA Sample prep kit v2 (Illumina) following the manufacturers protocol with 1μg total RNA input. 50 bp single end sequencing was performed using a HiSeq Illumina sequencer. Obtained sequences were de-multiplexed, quality controlled and mapped on the Aspergillus nidulans genome assembly (A_nidulans_FGSC_A4_version_s10-m03-r07). Mapping was performed using Novoalign (NovoCraft) and reverse transcripts were counted using python script HTSeq [87]. Normalization and statistics were done using R/Bioconductor and the limma and edgeR packages, using mean-variance weighting (voom) and TMM normalisation [88]. A significance cut-off of p < 0.01 (adjusted for multiple testing by the false discovery rate method) was applied for analysis. R plots used the ggplot2 package [89]. Transcription levels are log2 read counts per kilobase of exon per million library reads (RPKM). For trace graphs as shown in S7–S13 Figs transcript coverage was calculated as explained for the ChIP-seq experiments to obtain counts per million reads (CPM). SM clusters are annotated as described by Inglis and coworkers [72]. All data are available at NCBI GEO under the accession number GSE72126.
Chromatin immunoprecipitation was performed as described in [18] Chromatin was incubated with antibodies specific to H3K4me3 (Abcam, 8580), H3K9me3 (Active Motif, 39161), H3K36me3 (Abcam, 9050), H3Ac (Millipore, 06–599) or Histone H3 C-terminus (Abcam, 1791) and Dynabeads Protein A (Invitrogen). Precipitated DNA from two biological and two technical replicates was quantified by real-time PCR according to protocol (Bio-Rad) using iQ SYBR Green Supermix and normalized to input DNA or sequenced. Primers used in quantitative PCR were HPLC purified and are shown in S2 Table. For Illumina sequencing, ChIP-seq libraries were prepared using 10 ng of immunoprecipitated DNA following the instructions supplied with Illumina Tru-seq ChIP-seq kits (Illumina Cat# FC-121-2002). Illumina sequencing was performed using an Illumina NextSeq500 Instrument at the University of Georgia Genomics Facility. Short reads were mapped using Novoalign (NovoCraft) to the current Aspergillus genome annotation, obtained from the Aspergillus Genome Database [90]. Read numbers were counted for 10 base pair bins using sam tools and R, and the read density was normalized for total read number and visualized using the Integrated Genome Viewer or Integrated Genome Browser [91–93]. In detail: Metaplots (Figs 3A, 5A, 5B and S4–S6) were calculated from bam files using bedtools genomecov [94]. The sequencing coverage per base pair (bp) was calculated for the whole genome, normalized using a scaling factor (1000000/total mapped sequence read counts) that accounts for the different counts of mapped reads per sample to obtain counts per million mapped reads (CPM) to allow comparison between samples represented as trace files in sgr format. Using R scripts the sgr files were smoothed by averaging a window of 100 bp length that was slided by 10bp, thereby reducing computation demand (10bp binning). Gene start/stop codon position, length and strand were retrieved from gff file provided by Aspergillus Genome Database [90] and 2kb of each gene (500bp upstream + 1500bp downstream of ATG in case of H3Ac and H3K4me3 or 1500bp upstream + 500bp downstream of stop codon for H3K36me) taken and averaged for the specified group (e.g. transcription strength). Detailed R scripts can be obtained from the authors.
Per gene levels of ChIP-seq data were calculated as transcript sequences counting reads per gene and were normalized to the ORF length (in comparison to the exon length) to obtain reads per million reads per kb ORF length (RPKM).All data are available at NCBI GEO under the accession number GSE72126.
Small pieces of kdmBΔ and the wildtype strain, grown on YMG agar, were used to inoculate 100 mL of AMM and ZM1/2 medium in 500 mL Erlenmeyer flasks. The flasks were kept on a rotary shaker at 37°C and 160 rpm until the glucose was consumed (64 hours for AMM and 9 days for ZM1/2). pH value and glucose content of the culture fluid were monitored as described previously [95]. After harvesting, mycelium and culture fluid were separated by filtration. The culture fluid was extracted with the same volume of ethyl acetate twice, the combined organic layers were dried over sodium sulfate and the solvent was evaporated in vacuo (40°C) to provide the crude extract. The wet mycelium was extracted with 100 ml of acetone for 30 min in an ultrasonic bath (25°C) and the organic solvent was evaporated in vacuo (40°C). The remaining aqueous residue was diluted with 20 ml of H2O and extracted with the same volume of ethyl acetate twice. After drying over sodium sulfate, the organic solvent was removed in vacuo (40°C) to yield the crude mycelial extract. The extracts were dissolved in methanol, filtered through SPE C18 cartridges and subjected to mass spectrometric analyses.
All analyses were performed on Agilent 1260 Infinity Systems with diode array detector and C18 Acquity UPLC BEH column (2.1 × 50 mm, 1.7 μm) from Waters. Solvent A: H2O + 0.1% formic acid, solvent B: AcCN + 0.1% formic acid, gradient system: 5% B for 0.5 min increasing to 100% B in 19.5 min, maintaining 100% B for 5 min, flowrate = 0.6 mL min−1, UV detection 200–600 nm. LC-MS and MS/MS spectra were recorded on an ion trap MS (amaZon speed, Bruker) with an electrospray ionization source. Experiments were performed using positive and negative ionization modes. The capillary voltage of the ion source was 4500V and the nebulizer gas was set to 4 bar with drying gas flow of 12 L/min. For MS2 experiments SmartFrag was used with CID voltage of 1V and amplitude ramping of 60–180% (fragmentation time 40 ms, Cutoff 17%). HR-MS spectra were recorded on a time-of-flight (TOF) MS (MaXis, Bruker) with electrospray ionization source using positive ionization mode.
For data analysis and calculation of molecular formulas, including the isotopic pattern, dataAnalysis (Version 4.2) from Bruker was used. Compound search was performed using Dictionary of Natural Products (CRC Press) and Antibase 2010 (Wiley-VCH). ESI-MS, ESI-MS2 and UV-Vis absorption spectra of identified metabolites were compared to corresponding literature data [96–104].
|
10.1371/journal.ppat.1005065 | IL-27 Signaling Is Crucial for Survival of Mice Infected with African Trypanosomes via Preventing Lethal Effects of CD4+ T Cells and IFN-γ | African trypanosomes are extracellular protozoan parasites causing a chronic debilitating disease associated with a persistent inflammatory response. Maintaining the balance of the inflammatory response via downregulation of activation of M1-type myeloid cells was previously shown to be crucial to allow prolonged survival. Here we demonstrate that infection with African trypanosomes of IL-27 receptor-deficient (IL-27R-/-) mice results in severe liver immunopathology and dramatically reduced survival as compared to wild-type mice. This coincides with the development of an exacerbated Th1-mediated immune response with overactivation of CD4+ T cells and strongly enhanced production of inflammatory cytokines including IFN-γ. What is important is that IL-10 production was not impaired in infected IL-27R-/- mice. Depletion of CD4+ T cells in infected IL-27R-/- mice resulted in a dramatically reduced production of IFN-γ, preventing the early mortality of infected IL-27R-/- mice. This was accompanied by a significantly reduced inflammatory response and a major amelioration of liver pathology. These results could be mimicked by treating IL-27R-/- mice with a neutralizing anti-IFN-γ antibody. Thus, our data identify IL-27 signaling as a novel pathway to prevent early mortality via inhibiting hyperactivation of CD4+ Th1 cells and their excessive secretion of IFN-γ during infection with African trypanosomes. These data are the first to demonstrate the essential role of IL-27 signaling in regulating immune responses to extracellular protozoan infections.
| Infection with extracellular protozoan parasites, African trypanosomes, is characterized by a persistent inflammatory immune response. It has been recently shown that maintaining the balance of the inflammatory responses via dampening M1-type myeloid cell activation is critical to guarantee control of the parasites and survival of the host. In this study, we demonstrated that IL-27 receptor-deficient (IL-27R-/-) mice infected with African trypanosomes developed an excessive inflammatory response and severe liver immunopathology, resulting in dramatically reduced survival, as compared to infected wild-type mice. The early mortality of infected IL-27R-/- mice was correlated with significantly elevated secretions of inflammatory cytokines, particularly IFN-γ, and enhanced activation of CD4+ Th1 cells. Importantly, IL-10 production was not impaired in infected IL-27R-/- mice. Either depletion of CD4+ T cells, resulting in a dramatically reduced secretion of IFN-γ, or neutralization of IFN-γ, prevented the early mortality of infected IL-27R-/- mice with a significantly reduced inflammatory response and a major amelioration of the liver pathology. Thus, our data identify IL-27 signaling as a novel pathway to prevent the early mortality via inhibiting hyperactivation of CD4+ Th1 cells and their excessive secretions of IFN-γ during experimental infection with extracellular protozoan parasites African trypanosomes.
| African trypanosomiasis is a vector-borne parasitic disease of medical and veterinary importance. It is estimated that 170,000 people contract the disease every year, and that approximately 70 million people mainly in sub-Saharan Africa are at the risk of contracting the disease [1,2]. In addition, this disease severely limits the agricultural development by affecting domestic animals in the area [2]. The causative agents of this disease are various species of genus of Trypanosoma, which are extracellular protozoan parasites equipped with a flagellum that emerges from the flagellar pocket and provides the parasite with its motility [2]. Upon the bite of the mammalian host by a trypanosome-infected tsetse fly, the parasites enter the blood circulation via lymph vessels and can multiply in the bloodstream and interstitial fluids of the host [3,4]. The parasites have evolved very sophisticated evasion mechanisms to survive in the chronically infected host [3–5], causing a serious disease that is often fatal without treatment [1,2].
Due to practical and ethical reasons, mouse models have become an alternative and proven to be a cornerstone for studying African trypanosomiasis of humans and domestic animals [6]. Most of studies have been performed with T. brucei and T. congolense parasites [3,6]. Based on mouse models, although the parasites circulate in the blood stream, the liver is the major place for clearance of the parasites [7–9]. Recent studies demonstrated that Kupffer cells efficiently engulf trypanosomes, which is mediated by both IgM and IgG antibodies specific to the parasites [10–12]. IFN-γ, mainly secreted by VSG-specific CD4+ T cells [13–15] following activation by dendritic cells [16,17], has been shown to mediate protection during African trypanosomiasis [13,15,18–20]. Proinflammatory cytokines such as IL-12, TNF-α, as well as iNOS produced by M1-type myeloid cells are also critical for host resistance to African trypanosomes [15,21–25]. However, excessive secretions of these inflammatory cytokines by hyperactivated myeloid cells and T cells lead to liver pathology and shorten the survival of infected mice [11,22,26–29]. In this respect, IL-10 has been found to be essential for maintenance of the immunological balance between protective and pathological immune responses during African trypanosomiasis [11,20,22,26,27]. Importantly, the role of IL-10 as an anti-inflammatory agent has been more recently confirmed in cattle, primate and human infections with African trypanosomes [30–32]. It remains unknown whether, in addition to IL-10 signaling, another pathway that maintains this immunological balance exists.
IL-27, a recently identified cytokine produced primarily by macrophages and dendritic cells, is a member of the IL-12 super-family [33]. The IL-27 receptor (IL-27R) complex consists of the specific IL-27Rα subunit (WSX-1) and the IL-6R subunit (gp130), and is expressed on numerous subsets of leukocytes including CD4+ T cells, CD8+ T cells, NK cells, monocytes, Langerhans cells, and dendritic cells [34]. Earlier studies have demonstrated that IL-27, as a proinflammatory cytokine, drives naïve T cells to differentiate into Th1 cells [35–37]. More recent studies have suggested that IL-27 also has the function to inhibit immunopathology via downregulation of active CD4+ T cells during infections, particularly with intracellular protozoan parasites [38–42]. However, the precise mechanism of CD4+ T cell-mediated immunopathogenesis in the absence of IL-27 signaling still remains incompletely understood. In addition, it is not clear so far whether IL-27 plays an important role in regulation of the immune responses during infections with extracellular protozoan parasites such as African trypanosomes. Based on previous data showing that a subset of highly activated pathological CD4+ T cells produces excessive IFN-γ, and leads to immunopathology and early mortality of mice infected with T. congolense [11,28,29], we formulate a hypothesis that IL-27 signaling is, besides IL-10 signaling, another novel pathway that prevents the immunopathology and early mortality via down-regulation of the hyperactivity of CD4+ T cells and their excessive secretion of IFN-γ during experimental Africa trypanosomiasis. With this in mind, we examine in this study how IL-27 signaling regulates the immune responses in mice infected with African trypanosomes.
To evaluate the role of IL-27 signaling during African trypanosomiasis, we first determined whether infection led to increased expression of this cytokine or its receptor. Wild-type C57BL/6 mice were infected with T. congolense, a species of African trypanosomes which are unable to leave the circulation and only live in blood vessels, causing fatal disease in cattle [4]. The mice were euthanized at day 0, 7, and 10 after infection, as parasitemia usually peaked on day 6–7 after infection [15,29]. As the liver is the major organ for clearance of the parasites [7–9,11], the liver was collected for measurement of mRNA levels of IL-27 and its receptor using real-time quantitative RT-PCR. mRNA levels of both subunits of IL-27 (IL-27p28 and EBI3) were upregulated in the liver of mice at day 7 and 10 after infection, compared to uninfected mice (Fig 1A). In contrast, mRNA levels of IL-27 receptor (WSX-1) were not affected by the infection (Fig 1A).
Next, we infected IL-27R-/- (WSX-1-/-) and wild-type mice with T. congolense to assess whether IL-27 signaling affected the disease progression. Similar to infected wild-type mice, infected IL-27R-/- mice could control the first wave of parasitemia (Fig 1B). However, IL-27R-/- mice succumbed to the infection on day 12 to 20 after infection with a mean survival time of 14.5 days (Fig 1C). In contrast, infected wild-type mice survived until day 67 to 138 days after infection with a mean survival time of 123 days (Fig 1C). Compared to infected wild-type mice, the infected IL-27R-/- mice survived significantly shorter (p<0.01). These data demonstrated that IL-27 signaling is required for survival of mice infected with T. congolense.
The above results demonstrated that absence of IL-27 signaling led to earlier mortality of mice infected with African trypanosomes. As uncontrolled inflammation causes early mortality of mice infected with African trypanosomes [3,4], we next examined the plasma levels of inflammatory cytokines and their secretions by cultured spleen cells. As shown in Fig 2A, significantly higher amounts of IFN-γ, IL-12p40, and TNF-α were detected in the plasma of IL-27R-/- mice infected with T. congolense, compared to infected wild-type mice, on day 7 and 10 after infection (p<0.01). Although the plasma level of IFN-γ in IL-27R-/- mice decreased on day 10 after infection probably due to clearance of the first wave of parasitemia, it was still significantly higher than that of the infected wild-type mice (p<0.01, Fig 2A).
To evaluate the secretions of cytokines by spleen cells, spleen cells were collected from IL-27R-/- and wild-type mice on day 7 and 10 after infection with T. congolense, and cultured in vitro for 48 h. The production of IFN-γ, IL-12p40, and TNF-α by spleen cells were significantly elevated in infected IL-27R-/- mice, compared to infected wild-type mice (p<0.01 or <0.05, Fig 2B). As recent studies have shown that IL-27 mainly regulates CD4+ T cell activation during infection with intracellular pathogens [38–42], we further evaluated IFN-γ-producing CD4+ T cells in the spleen cultures using flow cytometry. A limited and similar percentage and absolute number of CD4+ T cells from uninfected wild-type and IL-27R-/- mice produced IFN-γ after 12 h stimulation with Cell Stimulation Cocktail (containing PMA, ionomycin, and protein transport inhibitors). However, by 7 and 10 days post infection both the percentage and the absolute number of IFN-producing CD4+ T cells were significantly enhanced in IL-27R-/- mice when compared to wild-type cohorts (Fig 2C).
We and others have previously shown that excessive systemic inflammatory responses of mice infected with African trypanosomes are associated with severe liver damage [11,22,43,44]. In addition, the liver is the primary organ of trypanosome clearance [7,9,11]. Therefore, we next evaluated effects of IL-27 signaling on liver pathology during the course of infection with the parasites. IL-27R-/- mice, but not wild-type mice, showed extensive pale geographic areas highly suggestive of necrosis on day 10 after infection with T. congolense (Fig 3A). Microscopic examination of the liver of infected IL-27R-/- mice revealed many large areas with loss of hepatocyte cellular architecture and an infiltration of inflammatory cells (Fig 3B). By contrast, these pathological changes were not observed in the liver of infected wild-type mice (Fig 3B). To further characterize the liver pathology, we measured the serum activities of alanine aminotransferase (ALT) of mice during T. congolense infection. As shown in Fig 3C, IL-27R-/- mice had significantly higher serum activities of ALT than wild-type mice on both day 7 and day 10 after infection (p<0.05), indicating death of hepatocytes and release of cytosolic enzymes. These results demonstrated that IL-27 signaling played a major role in prevention of the liver pathology that was associated with enhanced systemic inflammatory responses.
It has been shown that IL-10 is crucial for survival of mice infected with African trypanosomes through limiting inflammation [11,20]. In particular, failure to control inflammatory responses in mice infected with African trypanosomes in the absence of IL-10 signaling is associated with severe liver pathology [11,22,27]. In this regard, IL-27 has been shown to drive CD4+ T cells to produce IL-10 for downregulation of inflammation [45–47]. The similarity of the cytokine profile and liver pathology of infected mice in the absence of IL-27 signaling and IL-10 signaling [11,20] prompted us to examine whether IL-27 signaling prevented early mortality of mice infected with African trypanosomes via IL-10. We first compared the disease progression in the absence of IL-27 signaling with that in the absence of IL-10 signaling. T. congolense-infected IL-27R-/- mice and wild-type mice showed similar parasitemia and a significantly reduced survival after administration of anti-IL-10 receptor (IL-10R) mAb (p<0.01, Fig 4A). Strikingly, infected wild-type mice treated with anti-IL-10R mAb survived significantly shorter than infected IL-27R-/- mice (p<0.01, Fig 4A), suggesting that IL-27 and IL-10 may independently regulate inflammatory responses during African trypanosomiasis. Next we compared the IL-10 levels in plasma, and supernatant fluids of cultured spleen cells or liver leukocytes between IL-27R-/- and wild-type mice infected with T. congolense. There was no significant difference in IL-10 production in plasma and supernatant fluids of the cultures between IL-27R-/- and wild-type mice on day 7 after infection (Fig 4B). Surprisingly, IL-27R-/- mice even showed significantly higher amounts of IL-10 in both plasma (up to 14 folds) and supernatant fluids of cultured spleen cells or liver leukocytes on day 10 after infection (p<0.01 or <0.05, Fig 4B), demonstrating that secretion of IL-10 was strengthened, rather than impaired in IL-27R-/- mice infected with African trypanosomes, probably due to deficiency of the immune regulation mediated by IL-27 signaling in those infected IL-27R-/- mice. Taken together, these data suggested that early mortality of IL-27R-/- mice infected with African trypanosomes was not due to impaired IL-10 production.
Because early mortality of IL-27R-/- mice infected with African trypanosomes was associated with severe liver pathology without impaired secretion of IL-10 as shown above and because IL-27 has been shown to mainly regulate T cell, particularly CD4+ T cell activation during infection with intracellular pathogens [38–42], we next characterized CD4+ T cell responses in the liver of IL-27R-/- mice during infection with T. congolense. We found that the frequency and the absolute number of activated hepatic CD4+ T cells (CD44hiCD62Llow) were significantly higher in IL-27R-/- mice infected with T. congolense, compared to infected wild-type mice (p<0.01, Fig 5A). The production of IFN-γ, IL-12p40, and TNF-α by cultured liver leukocytes from infected IL-27R-/- mice was significantly higher than production of these cytokines by liver leukocytes from infected wild-type mice (p<0.001, <0.01 or <0.05, Fig 5B). In particular, the production of IFN-γ was enhanced by 4–8 folds in the liver leukocyte cultures of infected IL-27R-/- mice (Fig 5B). Thus, we further evaluated the activation of liver CD4+ T cells by examining their secretions of IFN-γ using single cell analysis. A small and similar percentage and absolute number of CD4+ T cells from uninfected wild-type and IL-27R-/- mice secreted IFN-γ after 12 h stimulation with Cell Stimulation Cocktail (containing PMA, ionomycin, and protein transport inhibitors). In contrast, by day 7 and 10 post infection significantly higher percentage and absolute number of IFN-γ-producing CD4+ T cells were detected in IL-27R-/- mice as compared to wild-type cohorts (Fig 5C). Collectively, these data suggested that the early mortality of IL-27R-/- mice infected with African trypanosomes was associated with exacerbated Th1-mediated immune responses with overactivation of CD4+ T cells.
As shown above, CD4+ T cells were excessively activated in the liver of IL-27R-/- mice infected with African trypanosomes, raising the possibility that the early mortality of infected IL-27R-/- mice was a consequence of a CD4+ T cell-dependent immune-mediated pathology. To test this, IL-27R-/- mice infected with T. congolense were treated with depleting anti-mouse CD4 mAb, anti-mouse CD8 mAb, or rat IgG as control; and the course of infection, immune responses, and severity of liver damage were assessed. As shown in S1 Fig, administration of the antibodies efficiently depleted CD4+ T cells or CD8+ T cells in the spleen and liver of the infected mice. Infected mice from all three groups could effectively control the first wave of parasitemia, although depletion of CD4+ T cells resulted in a significantly higher parasitemia at some time points of infection (p<0.01 or <0.05, Fig 6A). Strikingly, infected IL-27R-/- mice treated with anti-CD4 mAb had two fold increase of survival compared to infected IL-27R-/- mice treated with rat IgG (p<0.01, Fig 6A). In contrast, depletion of CD8+ T cells did not affect the survival of infected IL-27R-/- mice (Fig 6A). These results demonstrated that IL-27 signaling had a crucial role in dampening CD4+ T cell activation in experimental T. congolense infection in mice, allowing for prolonged survival.
We next evaluated the effect of CD4+ T cells on weight loss and liver pathology of IL-27R-/- mice infected with T. congolense. Infected IL-27R-/- mice treated with anti-CD4 mAb had significantly less weight loss at the later stage of infection, compared to infected IL-27R-/- mice treated with rat IgG or anti-CD8 mAb (p<0.01; S2A Fig). Importantly, infected IL-27R-/- mice treated with rat IgG or anti-CD8 mAb exhibited many large areas with loss of hepatocyte cellular architecture in the liver, whereas these pathological changes were hardly seen in the liver of infected IL-27R-/- mice treated with anti-CD4 mAb (S2B Fig). In addition, depletion of CD4+, but not CD8+, T cells significantly reduced the serum activities of ALT in IL-27R-/- mice infected with T. congolense (p<0.05, Fig 6B). These data suggested that CD4+ T cells played a central role in the development of liver pathology in experimental T. congolense infection, and that IL-27 was crucial for dampening this CD4+ T cell-mediated pathology.
We further characterized the contributions of CD4+ T cells to secretion of cytokines in IL-27R-/- mice infected with T. congolense. Depletion of CD4+, but not CD8+, T cells significantly reduced plasma levels of IFN-γ and TNF-α in infected IL-27R-/- mice (p<0.001 or <0.05), although the reduction of IL-12p40 did not reach statistical significance (Fig 6C). In addition, depletion of CD4+, but not CD8+, T cells also resulted in significantly less secretion of IFN-γ by spleen cells from infected IL-27R-/- mice (p<0.05, S2C Fig). Interestingly, depletion of CD4+ T cells almost abrogated the production of IL-10 by spleen cells in infected IL-27R-/- mice (p<0.01, S2C Fig), suggesting that IL-10 was predominantly produced by CD4+ T cells. Importantly, the observation that the enhanced survival of infected IL-27R-/- mice treated with anti-CD4 mAb was correlated with very little secretion of IL-10 further suggested that IL-27 signaling inhibited hyperactivation of Th1 cells in an IL-10 independent manner as shown above in Fig 4.
Having demonstrating that IL-27 is crucial for dampening trypanosomiasis-associated CD4+ T cell activation, needed for prolonged survival, we next addressed the mechanism of CD4+ T cell-mediated mortality of infected IL-27R-/- mice. Because the production of IFN-γ, and the frequency and the absolute number of IFN-γ-producing cells were enhanced in infected IL-27R-/- mice compared to infected wild-type mice (Fig 2 and Fig 5), and also because depletion of CD4+ T cells dramatically reduced the IFN-γ production (Fig 6; S2 Fig), we examined whether the early mortality of infected IL-27R-/- mice was directly attributed to the overproduction of IFN-γ. IL-27R-/- mice infected with T. congolense were treated with neutralizing anti-IFN-γ mAb or rat IgG as a control. Although administration of anti-IFN-γ mAb led to doubled parasitemia in infected IL-27R-/- mice at the peak on day 7 after infection (P<0.05), the infected IL-27R-/- mice treated with anti-IFN-γ mAb efficiently controlled the first wave of parasitemia as infected control mice did (Fig 7A). Importantly, administration of anti-IFN-γ mAb significantly enhanced the survival of infected IL-27R-/- mice (p<0.01; Fig 7A), demonstrating that high levels of IFN-γ accelerated the mortality of IL-27R-/- mice infected with African trypanosomes.
We next assessed the effects of IFN-γ neutralization on weight loss and liver pathology of IL-27R-/- mice infected with T. congolense. Infected IL-27R-/- mice treated with anti IFN-γ mAb had significantly less weight loss than infected IL-27R-/- mice treated with rat-IgG on the late stage of infection (p<0.01, S3A Fig). Importantly, infected IL-27R-/- mice treated with anti-IFN-γ did not exhibit areas with loss of hepatocyte cellular architecture in the liver whereas these pathological changes were observed in the liver of infected IL-27R-/- mice treated with rat IgG (S3B Fig). Moreover, neutralization of IFN-γ significantly reduced the serum activities of ALT in infected IL-27R-/- mice (p<0.01, Fig 7B). These data suggested that IFN-γ played a critical role in the development of liver pathology in IL-27R-/- mice infected with African trypanosomes.
We finally examined cytokine responses of infected IL-27R-/- mice treated with anti-IFN-γ mAb. IFN-γ was almost undetectable in the plasma of IL-27R-/- mice treated with anti-IFN-γ, suggesting the neutralization was successful (p<0.01, Fig 7C). Plasma levels of IL-12p40 and TNF-α were dramatically reduced in infected IL-27R-/- mice treated with anti-IFN-γ mAb, compared to infected IL-27R-/- mice treated with rat IgG (p<0.01, Fig 7C). Neutralization of IFN-γ also significantly reduced the production of IL-12p40 and TNF-α by cultured spleen cells (p<0.01, or <0.05, S3C Fig). Thus, the results indicated that IFN-γ was critically involved in the enhanced inflammatory responses in IL-27R-/- mice infected with African trypanosomes.
We finally characterized the role of IL-27 signaling in regulation of immune responses during T. brucei infection. In contrast to T. congolense, T. brucei species have the ability to penetrate the walls of capillaries, invade interstitial tissues, including the brain tissues, thus serving as a model of human African trypanosomiasis [48,49]. T. brucei infection also upregulated the mRNA expressions of IL-27p28 and EBI3, but not IL-27R-/- in the liver of mice (Fig 8A). IL-27R-/- mice infected with T. brucei efficiently controlled the first wave of parasitemia as infected wild-type did, but survived significantly shorter than infected wild-type mice (15 days vs. 32 days, p<0.01, Fig 8B), demonstrating an essential role of IL-27 signaling in prevention of the early mortality of mice infected with T. brucei. IL-27R-/- mice infected with T. brucei also showed enhanced IFN-γ production in plasma and supernatant fluids of spleen cultures, as well as enhanced serum activities of ALT, compared to infected wild-type mice (p<0.01 or <0.05, Fig 8C). Importantly, depletion of CD4+, but not CD8+, T cells enhanced the survival of IL-27R-/- mice infected with T. brucei by 3 folds (p<0.01, Fig 8D). Thus, IL-27 signaling is also required for survival of mice via preventing excessive Th1 immune responses during T. brucei infection.
Successful clearance of African trypanosomes in the bloodstream requires induction of inflammatory immune responses; however, failure to control this inflammation leads to immune-mediated pathology [4,50]. IL-10 signaling has been previously suggested to be involved in maintaining this immunological balance in African trypanosomiasis [11,20]. In the current study, we have identified IL-27 signaling as a novel pathway to maintain this immunological balance in African trypanosomiasis. Our data are the first to demonstrate the essential role of IL-27 signaling in regulating immune responses to extracellular protozoan infections. More importantly, we provided direct evidence, that infection-associated IL-27 signaling served to extend the survival of the infected host by dampening CD4+ T cell activation and their secretion of IFN-γ.
Indeed, the early mortality of infected mice lacking IL-27 signaling (IL-27R-/-mice) was correlated with exaggerated inflammatory responses and liver immunopathology. The disease similarity of infected mice lacking IL-27 and IL-10 signaling raised the possibility that regulatory function of IL-27 is mediated via the induction of IL-10 secretion, as IL-27 has the capability of promoting CD4+ T cells to secret IL-10 [45–47]. However, the fact that blocking IL-10R further shortened the survival of infected IL-27R-/- mice and the fact that infected mice lacking IL-10 signaling and infected mice lacking IL-27 signaling had distinct survival suggested that IL-27 functions through a mechanism independent of IL-10. In addition, compared to infected wild-type mice, infected IL-27R-/- mice produced similar or even higher amounts of IL-10, depending on the time points examined. Furthermore, the enhanced survival of infected IL-27R-/- mice following depletion of CD4+ T cells was correlated with dramatically reduced secretion of IL-10. These data suggested that a defect of IL-10 signaling is unlikely to contribute to the early mortality of IL-27R-/- mice. Thus, we suggest that IL-27 suppresses the liver pathology and prevents the early mortality of mice infected with African trypanosomes through IL-10-independent mechanisms, possibly by direct modulation of T cell function.
It has been previously demonstrated that IL-10 inhibits accumulation and activation of M1-type myeloid cells, in particular, TIP-DCs (CD11b+Ly6C+CD11c+TNF and iNOS producing DCs) in the liver during infection with African trypanosomes [22,26,27]. Accordingly, African trypanosomes-infected CCR2 deficient mice and MIF (macrophage migrating inhibitory factor) deficient mice exhibited significantly reduced accumulation of TIP-DCs, which was correlated with remarked diminished liver pathology, and significantly prolonged survival [26,44]. Thus, IL-10 signaling suppresses liver pathology, mainly through downregulation of M1-type myeloid cells [3,50]. In contrast, IL-27R-/- mice infected with African trypanosomes displayed more activation of T cells, in particular, CD4+ T cells. Moreover, depletion of CD4+ T cells prevented liver pathology and early mortality of infected IL-27R-/- mice. Obviously, IL-27 signaling functions through limiting activation of CD4+ T cells in African trypanosomiasis. Thus, although both IL-10 signaling and IL-27 signaling are crucial for limiting the inflammatory complications associated to African trypanosome in particular in preventing liver pathology, the two signal pathways involve distinct mechanisms.
Dampening accumulation of highly activated CD4+ T cells by IL-27 signaling has also been recently observed in infection with other microorganisms, particularly intracellular protozoan and bacterial pathogens [38,40–42,51]. Our data demonstrate that the same mechanism exists during infections with extracellular protozoan parasites such as African trypanosomes. However, the precise mechanism of CD4+ T cell-mediated early mortality in previous models was not fully elucidated [38,42]. One of the most important properties of CD4+ T cells is that they secret a large amount of IFN-γ upon activation. IFN-γ is required to eliminate intracellular parasites, but also has potential to induce immunopathology [52,53]. Indeed, early mortality of IL-27R-/- mice infected with Toxoplasma gondii, or Plasmodium berghei is associated with significantly enhanced production of IFN-γ [38,42], suggesting that IFN-γ might be a critical molecule for CD4+ T cell-mediated mortality in the absence of IL-27 signaling. Surprisingly, neutralization of IFN-γ did not prolong the survival, and had no effect on the liver pathology of IL-27R-/- mice infected with T. gondii or P. berghei at all [38,54]. Thus, although CD4+ T cell-mediated mortality coincides with significantly elevated secretion of IFN-γ, it still remains inconclusive whether IFN-γ is the direct mediator of CD4+ T cell-dependent mortality in these infections. In contrast, neutralization of IFN-γ significantly enhanced the survival IL-27R-/- mice infected with African trypanosomes accompanied by a major amelioration of liver pathology, providing direct evidence that IFN-γ directly mediated the mortality of infected IL-27R-/- mice. In addition, enhanced survival of infected IL-27R-/- mice depleted of CD4+ T cells was correlated with a dramatically reduced production of IFN-γ. Obviously, either removing of CD4+ T cells or neutralization of IFN-γ got rid of the lethal effect of IFN-γ, leading to the prolonged survival of infected IL-27R-/- mice. Thus, another important finding of this study is that, in the absence of IL-27 signaling, CD4+ T cells mediated mortality directly through their secretion of IFN-γ, at least, during infection with extracellular protozoan parasites African trypanosomes.
It is important to point out that our results in no way exclude the protective role of CD4+ T cells and IFN-γ during infection with the parasites. Indeed, early studies have shown that there was a correlation between high IFN-γ levels in serum, low parasitemia, and host resistance during infection with African trypanosomes [18]. Subsequent studies demonstrated that VSG-specific CD4+ T cells mediated protection via secretion of IFN-γ [13,55]; and splenic DCs were the primary cells responsible for activating naïve VSG-specific CD4+ T cell responses [16,17]. The protective role of CD4+ T cells and IFN-γ in African trypanosomiasis has been recently confirmed by independent groups [14,15,19]. In support of previous findings, we showed that either depletion of CD4+ T cells or neutralization of IFN-γ resulted in a significantly elevated peak parasitemia level in IL-27R-/- mice infected with T. congolense, confirming the protective role of CD4+ T cells and IFN-γ during the infection. It is likely that IFN-γ promotes M1-type myeloid cells to produce IL-12, TNF-α and iNOS, which has been shown to be critically involved in lysis or damage of African trypanosomes [15,21,23,25,56]. On the other hand, excessive production of IL-12, TNF-α and iNOS driven by IFN-γ could also mediate immunopathology of mice infected with African trypanosomes [22,24,26,27,57]. Further, IL-12 and TNF-α could stimulate T cells to produce more IFN-γ [4,21]. Thus, IL-10 is required to down-regulate the production of IL-12, TNF-α and iNOS possibly by direct modulation of M1-type myeloid cells [11,22,26,27]. In the present study, we identified IL-27 signaling as a novel pathway to down-regulate the secretion of IFN-γ by direct modulation of CD4+ T cells. Obviously, in the absence of IL-27 signaling, excessive secretions of IFN-γ by CD4+ T cells also mediate liver pathology and mortality, although IL-10 signaling still fully functions and the infected mice produce even more IL-10, in African trypanosomiasis. Thus, both IL-10 signaling and IL-27 signaling are required for survival of mice infected with the parasites via preventing aberrant inflammatory responses, although they function in a distinct manner in African trypanosomiasis.
In conclusion, we have described an essential role for IL-27 signaling in preventing early mortality of mice infected with African trypanosomes through dampening IFN-γ secretion by CD4+ T cells, thus identifying, in addition to previously described IL-10 signaling, a novel pathway for maintenance of immunological balance during infection with extracellular protozoan parasites African trypanosomes. These data contribute significantly to our understanding of both immunopathogenesis of African trypanosomiasis and mechanisms underlying IL-27 immunoregulation during infection with extracellular protozoan and bacterial pathogens.
This study was performed in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The animal protocols involving mice were approved by the University of Maryland Institutional Animal Care and Use Committee (IACUC) under protocol R-12-60.
Eight- to teen-week-old C57BL/6NCrJ (C57BL/6) mice and five- to six-week-old outbred Swiss white mice (CD1) were purchased from the National Cancer Institute (Frederick, MD). B6N.129P2-Il27ratm1Mak (IL-27R-/-, or WSX-1-/-) mice were purchased from the Jackson Laboratory and bred in-house. All animal experiments were performed in accordance with the guidelines of the Institutional Animal Care and Use Committee and Institutional Bio-safety Committee of the University of Maryland, College Park.
T. congolense, Trans Mara strain, variant antigenic type (VAT) TC13 was used in this study. The origin of this parasite strain has been previously described [58]. T. brucei AnTat1.1E was obtained from the Institute of Tropical Medicine (Antwerp, Belgium). Frozen stabilates of parasites were used for infecting CD1 mice immunosuppressed with cyclophosphamide, and passages were made every third day as described previously [58]. The parasites were purified from the blood of infected CD1 mice by DEAE-cellulose chromatography [59] and used for infecting mice.
Purified rat anti-mouse IL-10 receptor (IL-10R) mAb (Clone 1B1.3a), purified rat anti-mouse CD4 mAb (Clone GK1.5), purified rat anti-mouse CD8 (Clone 53–6.72), and purified rat anti-mouse IFN-γ mAb (Clone XMG1.2) were purchased from BioXCell (West Lebanon, NH). Purified anti-mouse CD16/CD32 (FcγIII/IIR, Clone 2.4G2) were purchased from BD Biosciences. APC-Cy7 anti-mouse CD3e (145-2C11), PE-anti-mouse IFN-γ (XMG1.2), PE-Cy7-anti-mouse CD4 (GK1.5), PE-Cy7-anti-mouse CD4 (RM 4–4), FITC-anti-mouse CD8 (53–6.72), FITC-anti-mouse CD8 (YTS156.7.7), APC-anti-mouse CD44 (IM7), PE-anti-mouse CD62L (MEL-14), and matching controls were purchased from eBioscience or Biolegend.
Mice were infected i.p. with 103 T. congolense TC13 [11] or 5×103 T. brucei AnTat1.1E [44]. Some groups of infected mice were injected i.p. with rat anti-mouse IL-10R mAb (1B1.3a; 0.4 mg on day 0, 2, 4, and 6 after infection, respectively), anti-mouse CD4 mAb (GK1.5; 0.5 mg on day 0, 2, 4, and 6 after infection, respectively), anti-mouse CD8 mAb (53–6.72; 0.5 mg on day 0, 2, 4, and 6 after infection, respectively), anti-mouse IFN-γ mAb (XMG1.2; 0.4 mg on day 0, 2, 4, 6, 8, 10, 12, and 14 after infection, respectively), or rat IgG (as a control). Parasitemia was counted at ×40 magnification by phase-contrast microscopy. The survival time was defined as the number of days after infection that the infected mice remained alive.
For analysis of mRNA expression, total RNA was extracted from the homogenates of the liver of uninfected wild-type C57BL/6 mice or mice infected with T. congolense or T. brucei, following the manufacturer’s recommendation (Life Technologies). IL-27p28, EBI3, and WSX-1 mRNA levels were quantified by real-time quantitative RT-PCR. The cDNA expression for each sample was standardized using the house keeping gene β-actin. Cycling conditions were as follows: initialization 2 min at 50°C and 10 min at 95°C, followed by 40 cycles of 15 s at 95°C and 1 min at 60°C. Primer pair used were: IL-27p28: 5’-CTGGTACAAGCTGGTTCCTG-3’, 5’-CTCCAGGGAGTGAAGGAGCT-3; EBI3: 5’-CAGAGTGCAATGCCATGCTTCTC-3’, 5’-CTGTGAGGTCCTGAGCTGAC-3’; WSX-1: 5’-CAAGAAGAGGTCCCGTGCTG-3’, 5’-TTGAGCCCAGTCCACCACAT-3’.
Splenocytes were collected from mice. Cells were cultured at a concentration of 5 × 106 cells/ml (200 μl/well) in 96-well tissue culture plates in a humidified incubator containing 5% CO2. The culture supernatant fluids were collected after 48 h and centrifuged at 1,500g for 10 min, and the supernatant fluids were stored for cytokine assays at -20°C until used.
Liver leukocytes were isolated as described previously [60]. Briefly, the liver was perfused with PBS until it became pale. Thereafter, the gallbladder was removed and the liver excised carefully from the abdomen. The liver was minced into small pieces with surgical scissors and forced gently through a 70 um cell strainer using a sterile syringe plunger. The preparation obtained was suspended in 50 ml RPMI-1640 medium containing 10% FCS. The cell suspension was centrifuged at 30g with the off-brake setting for 10 min at 4°C. The obtained supernatant was centrifuged at 300g with the high-brake setting for 10 min at 4°C. The pellet was resuspended in 10 ml 37.5% Percoll in HBSS containing 100 U/ml heparin and then centrifuged at 850g with the off-brake setting for 30 min at 23°C. This new pellet was resuspended in 2 ml ACK buffer (erythrocyte lysing buffer), and incubated at room temperature for 5 min, then supplemented with 8 ml RPMI-1640 medium containing 10% FCS, followed by centrifugation at 300g with the high-brake setting for 10 min at 8°C. Cells were collected and cultured at a concentration of 5 × 106 cells/ml (200 μl/well) in 96-well tissue culture plates in a humidified incubator containing 5% CO2. The culture supernatant fluids were collected after 48 h and centrifuged at 1,500g for 10 min, and the supernatant fluids were stored for cytokine assays at -20°C until used.
Recombinant murine cytokines and Abs to these cytokines for use in ELISA were purchased from BD Biosciences or R&D Systems. The levels of cytokines in culture supernatant fluids or plasma were determined by routine sandwich ELISA using Immuno-4 plates (Dynax Technologies), according to the manufacturer’s protocols.
To assess the activation of T cells, intrahepatic leukocytes were isolated as described above. The cells were incubated (15 min, 4°C) with purified anti-mouse CD16/CD32 ([FcγIII/II Receptor], clone: 2.4G2) to block nonspecific binding of Abs to FcRs, washed with staining buffer (eBioscience), resuspended in staining buffer, and stained with mAbs specific for various cell surface markers, or the relevant isotype-matched control Abs. For intracellular IFN-γ staining, spleen cells or intrahepatic leukocytes were diluted to 5 × 106 cells/ml and cultured (200 μl/well) in a 96-well plate in the presence of 1x Cell Stimulation Cocktail (containing PMA, ionomycin, and protein transport inhibitors, eBioscience) for 12 h. The cells were then harvested and washed twice in staining buffer. The cells were incubated (15 min, 4°C) with purified anti-mouse CD16/CD32, washed with staining buffer, followed by staining with mAbs specific for cell surface markers. The cells were fixed and permeabilized using Intracellular Fixation & Permeabilization Buffer Set (eBiosciences). Intracellular staining was then performed using mAbs specific for IFN-γ. Samples were resuspended in staining buffer, tested by FACSAria II, and analyzed using FlowJo software.
Liver alanine transaminase (ALT) activities were determined using EnzyChrom Alanine Transaminase Assay Kit (BioAssay Systems) according to the manufacturer’s instructions. For histopathological examination, the liver was taken from mice on day 10 after infection and fixed with 10% formalin in PBS. Sections were stained with Hematoxylin and Eosin.
Data are represented as the mean ± SEM. Significance of differences was determined by ANOVA or a log-rank test for curve comparison using the GraphPad Prism 5.0 software. Values of p≤0.05 are considered statistically significant.
|
10.1371/journal.pgen.1000868 | Derepression of the Plant Chromovirus LORE1 Induces Germline Transposition in Regenerated Plants | Transposable elements represent a large proportion of the eukaryotic genomes. Long Terminal Repeat (LTR) retrotransposons are very abundant and constitute the predominant family of transposable elements in plants. Recent studies have identified chromoviruses to be a widely distributed lineage of Gypsy elements. These elements contain chromodomains in their integrases, which suggests a preference for insertion into heterochromatin. In turn, this preference might have contributed to the patterning of heterochromatin observed in host genomes. Despite their potential importance for our understanding of plant genome dynamics and evolution, the regulatory mechanisms governing the behavior of chromoviruses and their activities remain largely uncharacterized. Here, we report a detailed analysis of the spatio-temporal activity of a plant chromovirus in the endogenous host. We examined LORE1a, a member of the endogenous chromovirus LORE1 family from the model legume Lotus japonicus. We found that this chromovirus is stochastically de-repressed in plant populations regenerated from de-differentiated cells and that LORE1a transposes in the male germline. Bisulfite sequencing of the 5′ LTR and its surrounding region suggests that tissue culture induces a loss of epigenetic silencing of LORE1a. Since LTR promoter activity is pollen specific, as shown by the analysis of transgenic plants containing an LTR::GUS fusion, we conclude that male germline-specific LORE1a transposition in pollen grains is controlled transcriptionally by its own cis-elements. New insertion sites of LORE1a copies were frequently found in genic regions and show no strong insertional preferences. These distinctive novel features of LORE1 indicate that this chromovirus has considerable potential for generating genetic and epigenetic diversity in the host plant population. Our results also define conditions for the use of LORE1a as a genetic tool.
| In contrast to animals, where germline differentiation initiates early in embryogenesis, germline differentiation in plants starts in the adult phase during reproductive development. Transpositions of transposable elements in both somatic and gametic cells can be transmitted to the next generation. As a result, plant genomes may contain transposable elements exhibiting a variety of tissue-specific activities. Thus far, the spatio-temporal activity of LTR retrotransposons, the most abundant class of transposable elements in plants, has not been well characterized. Here, we report a detailed analysis of the spatio-temporal transposition pattern of a plant LTR retrotransposon in the endogenous system. Using the model legume Lotus japonicus, we found that LORE1a, a member of the chromovirus LORE1 family that belongs to the Gypsy superfamily, was epigenetically de-repressed via tissue culture. Activation was stochastic and derepression was maintained in regenerated plants. This feature made it possible to trace the original spatio-temporal activity of the retrotransposon in the intact plants. We determined that the plant chromovirus retrotransposes mainly in the male germline, without obvious insertional preferences for chromosomal regions. This finding suggests that the tissue specificity of transposable elements should be taken into account when considering their impact on the host genome dynamics and evolution.
| A large proportion of the eukaryotic genome is composed of transposable elements (TEs). In flowering plants, Long Terminal Repeat (LTR) retrotransposons have been regarded as the largest order of TEs [1],[2] and it has been suggested that the ratio between propagation and exclusion of LTR retrotransposons may have affected the size of host genomes [3],[4]. In line with this notion, large plant genomes usually contain substantially more LTR retrotransposons than small plant genomes [5],[6]. However, data from a wide range of flowering plants strongly suggest that LTR retrotransposons are not distributed evenly in genomes. Biased accumulation has led to the formation of LTR retrotransposon-rich, gene-poor heterochromatic blocks, which separate gene-rich euchromatic regions [7]. Thus, the activity of LTR retrotransposons has contributed remarkably towards generating the basic structure of current plant genomes.
In flowering plants, the LTR retrotransposons have been classified into two superfamilies, Gypsy and Copia, according to their structural features [8]. In many plants, Gypsy outnumbers Copia [9]–[13]. An exception is grapevine, in which the number of Copia elements exceeds that of Gypsy [14]. Chromovirus is a most widely-distributed lineage of Gypsy, characterized by a chromodomain at the carboxyl terminal of the ORF [15],[16]. It has been proposed that the insertion site preference of chromoviruses is controlled by the chromodomain [15],[16], and this suggestion has been supported by functional characterization of MAGGY, identified in the rice blast fungus Magnaporta grisea [17],[18]. The MAGGY chromodomain was shown to interact with histone H3 di- and tri-methyl K9, which are hallmarks of heterochromatin [18]. When it was fused to the integrase of Tf1 retrotransposon, the modified Tf1 preferentially transposed into heterochromatic regions in Schizosaccharomyces pombe genome [18]. In flowering plants, chromoviruses are phylogenetically distinct from the lineage containing MAGGY and they are classified into four clades, Reina, Tekay, Galadriel and CRM [16],[19]. Members of CRM were originally known as Gypsy elements which accumulate in centromeric and pericentromeric regions in plant genomes [20]–[23]. Since all four clades have been identified in both dicots and monocots, and Reina and CRM elements have been found in angiosperms and gymnosperms, these elements are likely to have an ancient origin within the seed plants [16],[19]. In order to complement these evolutionary studies, a precise characterization of retrotransposon transpositional activity is now being pursued by experimental analyses, and this activity represents one of the subjects that must be addressed if we are to develop a deeper understanding of plant genome dynamics and evolution.
Previously, most experimental studies of transpositional activity and the regulation of plant LTR retrotransposons were conducted using three Copia elements, Tnt1 and Tto1 in tobacco, and Tos17 in rice. Transpositions of these elements were observed only in cultured cells, where their transcriptional up-regulation occurs [24]–[26]. Since transpositional activity is immediately repressed in regenerated plants due to a decrease in transcription, transpositions in intact plants have not been well characterized. Thus far, transposition of Tos17 has been observed in intact transgenic plants in which the transcriptional level of a gene encoding histone H3K9 specific methylase was downregulated by RNA interference [27], but the spatio-temporal pattern of transposition remained unclear. Furthermore, little is known about the transpositional activity of plant Gypsy elements, including chromoviruses, despite their high abundance in plant genomes.
In more than a decade of studies, the model legume Lotus japonicus has facilitated dissection of the molecular mechanisms governing symbiotic nitrogen fixation with rhizobia. The L. japonicus genome has been sequenced and sequence data covering 67% of the genome (472 Mb), corresponding to 91.3% of the gene space, is now available [13]. From this model legume, we have identified two transpositionally active LTR retrotransposon families designated as LORE1 and LORE2 (Lotus Retrotransposon 1 and 2) [28],[29]. Both belong to the Gypsy superfamily and were first identified as insertions in symbiotic mutants isolated from a transgenic plant population established by tissue culture- mediated transformation [28]–[31]. However, the machinery underlying their activation remained to be characterized. Both LORE1 and LORE2 encode unique long open reading frames (ORFs) with a chromodomain at the carboxyl terminal ends, which suggests that they are chromoviruses (Figure 1A) [29]. Although this chromodomain was overlooked in the original characterization of LORE1 [28], Novikova et al. re-classified LORE1 as a member of the Reina clade of chromovirus [19].
Previously, we estimated the number of “preexisting copies” (insertions that were already present in a plant accession) of LORE1 in the Gifu accession as ten, and obtained full or partial sequences for nine out of the ten preexisting LORE1 copies [28]. Nucleotide sequence polymorphisms among the nine copies enabled us to distinguish them from each other, and we designated them in alphabetical order as LORE1a, b, c, d, e, f, g, h, and i [28]. In this report, we show that in the Gifu accession, the preexisting LORE1a can be epigenetically de-repressed in standard tissue culture. However, transpositions per se occur primarily in pollen, i.e., male gametophytes, of regenerated intact plants, and so far new insertions generated in cultured cells have not been detected. We assume that the pollen-specific LTR promoter of LORE1a regulates the spatio-temporal pattern of transposition. Although LORE1 is a chromovirus, it does not appear to have a strong insertional preference for heterochromatin. These distinctive features of LORE1 underlie its ability to generate insertional polymorphisms, leading to a wide range of genetic and epigenetic diversity in a population. The results also define conditions for using LORE1 for insertion mutagenesis.
The transpositional activity of LORE1 was first demonstrated by the identification of four symbiotic mutant alleles, nin-7, symrk-1, nup133-3 and nap1-1, in which gene inactivation was caused by the insertion of LORE1 [28],[32]. As all four mutants were isolated from the same Ac/T-DNA tagging population established using the L. japonicus Gifu accession [30],[31], we screened other plants of the same population for LORE1 transpositions. Sequence-specific amplified polymorphism (SSAP) analysis of LORE1 insertion sites detected new transpositions in 32 plants out of a sub-population of 41 plants (Population 1 in Table 1), indicating that LORE1 was widely active in this population. Next, we investigated whether LORE1 transpositions were present in four transgenic or non-transgenic regenerated plant populations created using the Gifu accession. To detect new insertion sites of LORE1, we used SSAP to analyze the T1 and R1 progeny of primary transformants (T0) and of primary non-transgenic regenerated plants (R0). In addition to population 1 (Table 1), transpositions were detected in three of the other four populations. Importantly, transposition was detected in transgenic plants generated using six different constructs, as well as in non-transgenic regenerated plants. These results suggest that the simple process of in vitro tissue culture can activate LORE1 in a stochastic manner that is independent of the presence or absence of transgenes, antibiotic selection, and of the composition and contents of transgene constructs. The newly transposed LORE1 copies observed in R1/T1 plants might have resulted from transpositions in cultured cells and/or in the parental R0/T0 plants. However, LORE1 transposition was absent, infrequent, or below the detection levels of the SSAP method in a total of 27 plants from the initial R0/T0 plants from populations 2 and 3 (Table 1; data not shown). Previously, we observed the absence of obvious transcriptional or transpositional activation of LORE1 in cultured cells [28],[29]. These results suggest that even though LORE1 was apparently de-repressed in tissue culture, the transpositions per se appear to have occurred in regenerated intact plants, rather than in the cultured cells (see details in the next section).
To gain more precise information about transposition of LORE1 in intact plants, eight independent T0 plants were randomly selected from population 3 (Table 1) and investigated together with their T1 progeny. LORE1 transpositions were detected in the T1 progeny from 6 of the 8 T0 plants. A typical result of a genomic Southern blot analysis and SSAP analysis of a T0 and its 10 T1 progeny plants (in this instance plant line no. 30) are shown in Figure 1B and Figure S1, respectively. Notably, the banding pattern in the T0 plants was the same as in the control Gifu, again indicating absent or infrequent LORE1 transposition in the primary regenerated plants (Figure 1B). However, additional bands corresponding to newly transposed LORE1 copies were detected in the T1 progeny (Figure 1B and Figure S1). The highly polymorphic banding pattern indicates the occurrence of frequent independent transpositions of LORE1 in T1 plants (Figure 1B and Figure S1).
Next, we determined whether the new insertions of LORE1 found in the T1 plants were the result of transmission of previous transpositions in somatic cells from T0 forming sectors or of de novo transposition. We analyzed T1 plants originating from two seed pods at the top of the same shoot of the parental T0 plant (Figure 1C). We did not detect any new bands that were shared by the two neighboring pods, or T1 plants originating from the same pod. This result indicates that the majority of LORE1 transpositions occurred at late developmental stages in T0 plants. Reciprocal crosses between plant no. 30 (from the Gifu accession) and plants from the MG20 accession were used to determine if the new transposed copies detected in the T1 plant were transmitted via male or female gametes. Five F1 plants obtained from each reciprocal cross were analyzed for LORE1 copy number (Figure 1D). In total, 21 bands corresponding to new LORE1 transpositions were detected among the 5 F1 plants obtained from the MG20 (female) × no. 30 (male) cross. In contrast, only 1 newly transposed LORE1 copy was detected in 5 F1 plants from no. 30 (female) × MG20 (male) cross. We conclude that although LORE1 is active in both male and female gametophytes, its activity is much higher in male tissues. Next, we used parent-specific single nucleotide polymorphisms (SNPs) in the flanking regions to determine the parental origin of the seven new insertion sites in MG20x30 F1 plants. This analysis showed that all the new transpositions originated from Gifu, the pollen donor. Hence, the majority of LORE1 transpositions detected in the F1 plants seemed to occur before fertilization. Altogether, LORE1 was revealed to be robustly active especially in male gametophytes. Previous reports indicate that activated retrotransposons can be re-silenced again by activities such as copy number-dependent establishment of epigenetic silencing [33],[34]. However, LORE1 was still active in three T1 plants that already possessed an increased number of LORE1 copies (Figure S2A). This finding indicates that once activated, LORE1 was able to transpose over at least two successive generations. On the other hand, we also observed that LORE1 was inactivated in the nup133-3 mutant, in which a single new transposition was detected in the Nup133 gene (Figure S2B).
Since the newly inserted LORE1 copies identified in the three symbiotic mutant alleles (nfr5-2, symrk-2, and nup133-3) were identical to one of the nine preexisting copies, LORE1a, we suspected that LORE1a was preferentially activated [28]. LORE1a-specific SNPs were identified in regions 1 and 2 (Figure 1A) and in all eight of the newly-transposed LORE1 fragments from population 3. This observation is consistent with our suggestion that LORE1a is responsible for the majority of LORE1 transpositions described here.
The transpositional activity of retrotransposons is often controlled at the transcriptional level [1], [2], [24]–[26]. We used RT-PCR to compare the levels of LORE1 transcription in mature flowers containing both male and female gametophytes (where LORE1 transposition presumably occurs). Among the eight T0 plants, including no. 30 from population 3 (Table 1), higher levels of LORE1 transcription were observed in the six T0 plants that possessed active LORE1 elements, compared to the control Gifu plant (Figure 2A). This finding indicated a correlation between the transcriptional and transpositional activities of LORE1. To determine which LORE1 family members were present in the transcript pool, RT-PCR products were TA cloned and sequenced. RT-PCR products spanning regions 1 and 2 were amplified separately from flowers of the control Gifu plant and two T0 plants (nos. 30 and 45) that exhibited LORE1 activity. LORE1a-specific SNPs were present in the region 1 of 7/16, 15/15 and 15/16 clones from the control Gifu, no. 30 and no. 45 plants, respectively. For region 2, LORE1a-specific SNPs were present in 3/12, 16/16 and 16/16 clones from the control Gifu, no. 30 and 45 plants, respectively. These data suggest that transcriptional activation is responsible for the preferential transposition of LORE1a among the family members. This expectation is supported by the following lines of evidence: i) a generally increased level of LORE1 transcripts in flowers of active lines; ii) a clear increase in LORE1a transcripts in two activated plants; and iii) all transposition events detected thus far are of LORE1a origin.
The pattern of LORE1a activation via tissue culture is different from that of other well-characterized retrotransposons such as Tos17, Tto1, and Tnt1, which are activated and transpose during tissue culture, resulting in a copy number increase in the primary regenerated plants (R0) [24]–[26]. We hypothesized that tissue- or cell-specific transcription determines the unique spatio-temporal pattern of LORE1 transposition. To test this hypothesis, we compared LORE1 transcript levels in leaves and flowers among four T0 plants and a control Gifu plant, as well as its transcriptional level in cultured cells (Figure 2B). We found that there were no detectable differences in LORE1 transcript levels in the leaves of the four T0 plants or in the control Gifu plant. In contrast, high levels of LORE1 transcripts accumulated in the flowers of plant nos. 3 and 30 compared to nos. 11 and 42, or the control Gifu plants and cultured cells. Furthermore, high LORE1 transcript levels were detected in pollen from the two T0 plants exhibiting LORE1 activity, compared to the two T0 plants without LORE1 activity or the control Gifu plant (Figure 2C). These observations suggest that LORE1 has transpositional activity in pollen and that tissue specificity is controlled at the transcriptional level.
Since the 5′ LTR is known to function as a promoter for LTR retrotransposons [1], we determined promoter activity of the LORE1a LTR using a transgenic L. japonicus Gifu accession carrying LORE1a LTR fused to a GUS reporter gene. GUS activity was detected in mature pollen grains that were released from anthers and had accumulated at the tip on the inside of the keel (Figure 3A), as well as in isolated pollen grains (Figure 3B). We could not detect LTR-driven GUS activity in any other tissues (data not shown). A similar pattern of GUS activity was observed in three out of six independent transgenic plant lines. These results are in good agreement with the RT-PCR analyses, which indicate up-regulation of LORE1 transcription in pollen grains (Figure 2C). To investigate the LTR promoter activity in a heterologous system, we generated transgenic Arabidopsis plants carrying the same construct. Four out of the seven Arabidopsis transgenic lines showed GUS activity in hydrated pollen grains on stigmas and in pollen tubes (Figure 3F and 3G). Prolonged staining for GUS activity detected weaker expression in developing young anthers (Figure S3A). In the youngest anthers showing activity, GUS was detected primarily in cell layers around the developing pollen, rather than in the developing pollen grains (Figure S3C and S3E). No GUS activity was detected in other tissues. Taken together, these results indicate that the LORE1 LTR specifically promotes transcription in pollen and that the tissue specificity of the cis-elements may be operational in a wide range of flowering plants.
The reported locations of several chromoviruses in the host plant genomes suggest that chromoviruses preferentially accumulate in heterochromatic regions [18], [20]–[23]. Of the nine preexisting LORE1 copies so far identified, the insertion sites of LORE1d, e, f, h, and i were found in genomic clones containing highly repetitive sequences, which were potential heterochromatic regions. However, the remaining four, LORE1a, b, c and g, were found in contigs that did not display any apparent heterochromatic characteristics (S. S. unpublished data). To investigate whether LORE1 exhibits a strong insertion site preference for heterochromatic regions, we used SSAP to obtain flanking sequences located immediately 5′ of new insertions in the T1 and R1 populations. A total of 97 SSAP fragments longer than 40 bp were analyzed by homology search using public databases including the L. japonicus genome sequence data obtained from the MG20 accession [13]. The absence of the 97 LORE1 insertions in the wild-type Gifu accession was confirmed by PCR (data not shown). In this analysis, only sequences showing homology higher than 77%, along stretches longer than 40 bp and with bit scores larger than 58, were considered homologous sequences. For the 75% of the LORE1 flanking sequences (73 out of the 97), homologous sequences including possible identical (allelic) sequences were identified from the published L. japonicus genome sequences (Table 2). The percentage (75%) is close to the coverage of the whole genome reported for the genome sequence project (67%) [13]. Among the 73 sequences, 37 were protein coding cellular genes or expressed sequence tags (ESTs), 11 were homologous to transposable elements (TEs), and the residual 25 did not show homology to genes or TEs and were categorized as unknown (Table 2). On the other hand, among the 24 fragments that did not show significant homology with L. japonicus sequences, 6 were classified as genes or ESTs, one was categorized as a TE, and the remaining 17 were classified as unknown (Table 2). Thus, a total of 43 sequences were assumed to be in genic regions. Among the 43, 31 were predicted to be exonic, since the insertion site was positioned in a region homologous to protein coding sequences and/or deposited ESTs. In contrast, 12% of the 97 LORE1 flanking sequences showed homology with TEs, which is lower than the predicted TE content of the L. japonicus genome (36%) derived from end-sequencing data of randomly selected BAC clones (S. S. unpublished data). Finally, we physically mapped 24 of the 73 SSAP sequences, and 4 of the 9 preexisting LORE1 members whose positions could be uniquely assigned, to the latest version of L. japonicus chromosome pseudo molecule [13] (Figure 4). This mapping indicated that the new insertion sites were distributed across the Lotus genome and no strong preference for LORE1 insertion sites was observed from those data.
Because of the frequent but stochastic derepression of LORE1a in regenerated plant populations (Table 1), we predicted that LORE1a activation accompanying tissue culture was induced epigenetically rather than genetically. We examined the status of cytosine methylation around the 5′ end of LORE1a by Southern blot analysis using two restriction enzymes, Hind III and Alu I, which are sensitive to cytosine methylation at residues inside their recognition site [35]. We examined genomic DNA from five T0 plants (nos. 3, 11, 30, 42 and 45), together with the control Gifu (Figure 5A). When Hind III was used to digest genomic DNA from leaves, we observed distinct bands (approximately 1.5 kb) in all of the five plants, suggesting the absence of cytosine methylation at the two Hind III sites surrounding the region complementary to the DNA probe used in this analysis (Figure 5B). When genomic DNA samples were digested with Alu I, signals corresponding to approximately 300 and 650 bp DNA fragments were detected in each plant (Figure 5A). We assumed that the lower band signals represented a mixture of three Alu I fragments of 263, 284, and 306 bp, resulting from the digestion of the Alu I site 5′ adjacent to LORE1a and one of three Alu I sites in the 5′ LTR (Figure 5B). Thus, detection of the smaller hybridizing bands indicates the presence of hypomethylated Alu I sites in the 5′ LTR. On the other hand, the larger band was assumed to correspond to the 640 bp Alu I fragment, resulting from the absence of hypomethylated cleavable Alu I sites in the 5′ LTR (Figure 5B). Detection of signals from both high and lower sized DNA fragments indicates heterogeneity of the methylation status at the three Alu I sites in the 5′ LTR of each of the six investigated plants. However, the relative signal intensity of these DNA fragments showed variation among the five plants. The intensity of lower bands (corresponding to a hypomethylated status) was predominant in plant nos. 3 and 30, which have active LORE1a. However, the higher band (corresponding to hypermethylated alleles) was more intense in plant no. 11, which did not have active LORE1a. In plants nos. 42 and 45, both higher and lower bands were detected, with intensities similar to that of the control Gifu (Figure 5A). These trends in the relative signal intensity between the large and smaller sized bands were reproducible in independently extracted genomic DNA (Figure S4). The banding patterns observed in flowers, where transcriptional activation of LORE1a was observed, were similar to those observed in leaves (Figure S4). This finding suggests that no obvious changes in cytosine methylation pattern can be correlated to changes in LORE1 transcriptional level between the two tissues. Altogether, it would appear that T0 plants have a variable epigenetic status for LORE1a, and that it is different from Gifu control plants.
An independent determination of the cytosine methylation status was obtained by bisulfite sequencing of the 5′ LTR of LORE1a in the five T0 plants and control Gifu. The same genomic leaf DNA samples used in the Southern blot in Figure 5 were analyzed, and twenty to twenty-four amplicons were sequenced from each plant line. This analysis revealed that cytosine residues in U3, the promoter region of LORE1a containing the three Alu I sites, are frequently methylated in control Gifu DNA, especially at CG and CHG sites (Blue bars in Figure 6A–6E). Graphical representation of the methylation status obtained from twenty amplicons showed some heterogeneity in the cytosine methylation patterns of the control Gifu (Figure S5A). This correlates with data obtained from the Hind III and Alu I digestion patterns (Figure 5 and Figure S4). LORE1a is activated in plant no. 30, and compared with control Gifu, this line showed a dramatic decrease in the cytosine methylation level throughout the investigated region (Figure 6C and Figure S5D). Plant no. 3 possesses activated LORE1a and it showed a general decrease in the methylation level in the U3 region; in three of twenty-three amplicons a complete loss of cytosine methylation in U3 was observed (Figure 6A and Figure S5B). LORE1a remains inactive in plant no. 11, and methylation at CG and CHG sites was maintained, as well as being very evident in the U3 (Figure 6B and Figure S5C). Plant nos. 42 and 45 showed similar methylation patterns when averaged among clones (Figure 6D and 6E). However, two amplicons corresponding to alleles that were completely demethylated in U3 were observed in plant no. 45, in which LORE1 is active, but not in no. 42, in which LORE1 remains inactive (Figure S5E and S5F). Among the T0 plants analyzed, these data support the idea that there may be a correlation between LORE1a activation and the presence of LORE1a alleles that have totally lost cytosine methylation in U3.
To determine whether alteration in the methylation pattern occurs in the same region of other LORE1 loci, we used bisulfite sequencing to determine the methylation status of two LORE1 loci, LORE1b and LORE1f, which contain 5′ LTRs identical to that of LORE1a. This analysis revealed that the cytosine methylation profile of LORE1f is similar to that observed for LORE1a in control Gifu (blue bars in Figure 6K–6O). Specifically, it shows a higher level of methylation in the U3 region compared with the remaining regions in the investigated areas. However, in contrast to LORE1a, the methylation profile of LORE1f was largely unchanged among the five T0 plants investigated (red bars in Figure 6K–6O). LORE1b showed a moderate level of methylation throughout the investigated region in control Gifu, resulting in a flatter profile of methylation compared with LORE1a and LORE1f (blue bars in Figure 6F–6J). The significant decrease in methylation levels in LORE1b was observed in all the 5 T0 plants, even though the level of decrease differed (red bars in Figure 6F–6J). Taken together, the bisulfite sequencing unveiled variation of epigenetic status at LORE1 loci in control Gifu plants and indicated alteration of this status in the five T0 plants investigated. A characteristic observed with LORE1a was the variability of epigenetic changes among the T0 plants, whereas LORE1b and LORE1f exhibited stability or rather similar changes among the five T0 plants.
In this study, we found that transposition of LORE1a, one of the LORE1 elements present in the Lotus japonicus accession Gifu, can be activated in plants regenerated from de-differentiated cells. In addition, we show that LORE1a transposes in the male germline, giving rise to independent insertions in the progeny. The frequency of activation differs between populations (Table 1), but was independent of construct or antibiotics used to select transgenes. Combining all the data, we infer that the phenomenon observed here is a result of a series of processes. The first is a tissue culture step that induces epigenetic changes in LORE1a. This alteration was documented by observing variation in the cytosine methylation patterns among the T0 plants investigated. In turn, this variation leads to the preferential transposition of de-repressed LORE1a in the pollen of intact regenerated plants, since the LORE1a LTR promoter is specifically active in pollen grains. Finally, newly transposed copies in the male germlines are inherited by the following generation. Our data suggests that mechanisms regulating the tissue-specific activity of TEs should be taken into account when considering the biology of TEs and their impact on genome dynamics and evolution. Activation of LORE1a appears to be an attractive system for investigating these mechanisms, as well as for the experimental analysis of plant chromovirus behavior.
Once de-repressed, the transpositional activity of LORE1a was maintained for at least two generations, indicating that the retrotransposon escaped the re-establishment of silencing during this period. One possible explanation for this escape from silencing is the low level of transcription of LORE1 in somatic cells, where de novo transcriptional silencing can be induced by an RNA-directed DNA methylation pathway. It has been shown that transcriptional gene silencing is inducible by artificial RNAi constructs utilizing the 35S promoter to drive the transgenes [36]. This promoter has been shown to be active in somatic tissues but not in pollen [37], and we do not know if de novo transcriptional silencing is inducible in pollen. Addition to that, it has been demonstrated that the RNA-directed transcriptional gene silencing and DNA methylation is less effective when the targets are located in the genic sequences, compared to those in the repetitive sequences [38]. Since LORE1a is located in an intron of a MAP kinase gene [28], efficiency of establishment of transcriptional gene silencing on once activated LORE1a may be low. An alternative possibility is that the increase in LORE1 copy-number was insufficient to induce copy number-dependent silencing [33],[34]. However, re-silencing of LORE1 was observed in the nup133-3 mutant, which also has a low LORE1 copy number. Interestingly, it was recently shown that pollen sperm cells accumulate transcripts of a set of genes involved in small RNA and DNA methylation pathways [39]. Furthermore, small RNAs of TEs originating from vegetative nuclei can transfer to sperm cells [40]. Investigation of the re-silencing of once-activated LORE1a in pollen, together with the steady state silencing of LORE1a, should provide new insights into the significance of epigenetic regulation in plant gametes.
Genetic changes, such as transposition of TEs and nucleotide substitutions and deletions generated during tissue culture, have been regarded as causes of the so-called somaclonal variations often observed as phenotypic changes in regenerated plant populations. Our investigation has unveiled another hidden layer of genetic changes creating phenotypic variation in regenerated plants. Epigenetic derepression of TEs induced via tissue culture can result in TE transpositions not in cultured cells but in regenerated plants. A similar behavior was observed for Karma, a rice LINE retrotransposon [41], and even though the underlying mechanism remains unclear, this observation indicates conservation of the feature. There are most likely other examples, but the temporal and spatial gaps between derepression and transposition of such TEs might have limited their detection. Our observation also indicates the potential use of tissue culture as a breeding method for generating epialleles of a gene of interest, even though these epialleles may not always be epigenetically stable, as demonstrated by recombinant inbred lines with epigenetically mosaic chromosomes consisting of wild-type and CG methylation-depleted segments [42]. Since epigenetic changes can be also generated in animal cells in culture [43], and considering the growing importance of the generative therapy using cultured stem cells, the risk of transposition of TEs after the regeneration of tissues should be given more attention and properly validated.
Even though we observed a good correlation between LORE1a activation and the presence of alleles with complete demethylation in the U3 of T0 plants, the presence of one amplicon of highly hypomethylated U3 in the control Gifu plant (completely hypomethylated except for one CG site, Figure S5A) suggests that demethylation alone might not be sufficient for LORE1a derepression. Therefore, there may be additional factors contributing to loss of LORE1a silencing in regenerated plants, but not in the Gifu plants. Alternatively, the changes in cytosine methylation pattern observed here may represent a by-product accompanying changes in chromatin states, such as histone modifications, which directly trigger LORE1a activation. Since the T0 plants analyzed in this study were selected with antibiotics during tissue culture, they are most likely of unicellular origin. Therefore, we suspect that the epigenetic variation at LORE1a that we observed among regenerated plants might already exist in cultured cells. Corroborating this suggestion is the finding that the epigenetic status of long-term cell cultures of Arabidopsis deviates from that of intact plants [44]. The range of epigenetic variation represented by the cytosine methylation pattern on LORE1a was more pronounced than in the other two LORE1 loci investigated in T0 plants. This suggests that although different members of a TE family may possess over 99% identity, their epigenetic regulation may differ and that tissue culture could influence the silencing variably. Position effects might represent a possible explanation for the different epigenetic changes among the three loci. Potential position effects have been observed in maize, which shows low heritability of silencing of a MuDR element induced by the Muk locus, a MuDR derivative producing a hairpin RNA molecule [45]. The transcriptional regulation of the neighboring MAP kinase gene might also affect expression of LORE1a. Although LORE1b was dramatically demethylated in regenerated plants, we have not yet observed transcriptional or transpositional activation of the copy. This finding indicates that silencing of LORE1b may be achieved by methods other than DNA methylation, such as histone modification, or even non-epigenetically via mechanisms influenced by the surrounding sequence, as demonstrated by Cheng et al. [46].
Chromodomains of chromoviruses are categorized into three groups, according to their structural features. Reina, Tekay, and Galadriel chromodomains are classified into group II, while the MAGGY chromodomain belongs to group I [18],[19]. Group I chromodomains contain three conserved aromatic residues that are necessary for interaction with methylated H3K9. Group II chromodomains only retain the second of these residues. Since CRM chromodomains differ more than those of groups I and II, they are referred to as CR motifs [18]. Even though neither group II chromodomains nor CR motifs interact with histone H3 methyl-K9, the interacting partner of group I chromodomains, they are able to target a YFP fusion to heterochromatic regions when expressed in plant cells, suggesting that they interact with an unknown partner present in plant heterochromatin [18]. Following the standard classification, the LORE1 and LORE2 chromodomains both belong to group II (Figure S6). Although the group II chromodomain in LORE1 appears canonical, we have not observed any strong global preference for insertion of LORE1 into heterochromatin. However, since the Lotus genome project was focused on euchromatic regions [13], we cannot exclude the possibility that LORE1 exhibits an insertional preference for gene-poor regions at a local level. In rice chromosome 1, the distribution pattern of chromoviruses possessing group II chromodomains suggest such a preference [18]. Future characterization of large numbers of new LORE1a insertion sites will, therefore, provide an opportunity to understand the biological function of the group II chromodomains. Gorinsek et al. pointed out that the genome of L. japonicus seems to contain a larger diversity of particular chromoviral clades than other plant species including Medicago truncatula, another model legume [16]. This may suggest that the L. japonicus genome was formed under the influence of the very active chromoviruses. Information on new insertion sites of LORE1a will also be useful for elucidating the survival strategy of these successfully propagated chromoviruses and the impact they have had on the current structure of the L. japonicus genome. From a different perspective, it might be interesting to see if the insertion site preference of LORE1 is affected by the chromatin structure in the pollen where it transposes, since the features of chromatin in plant sperm cells are distinct from somatic cells. Usually chromatin in pollen sperm cells is transcriptionally active at the same time as being highly condensed; it may use sperm-specific variants of histone H3.3, which is a hallmark of active chromatin [47],[48].
It is possible that in bisexual flowering plants, TEs like LORE1, which are active in germlines, could be strong generators of genetic variation over a short evolutionary period. Furthermore, the uniparental activity of these TEs, i.e., showing transposition mainly in male gametophytes, might provide an advantage as a survival strategy. Activity in pollen minimizes the risk of adversely affecting fertility because the number of pollen grains is usually large. Since particular TE families often show distinct biases for one of the two sex chromosomes, uniparentally-active TEs might also be involved in formation of sex chromosomes, which are evolutionarily recent events in flowering plants [49],[50]. On a shorter time-scale, as in the transpositional activity of LORE1, gametophytic transposition, as well as the lack of strong bias for insertion sites and frequent insertions into genes, indicates that this retrotransposon could be an ideal tool for establishing an insertional gene tagging system. We estimate that the population size necessary to obtain at least one insertion allele for all genes at a 95% probability is approximately 200,000 plant lines in L. japonicus. This calculation is based on the following assumptions: the value 2.7, the highest average number of new copies observed here in a T1 plant derived from a T0 plant; 2.9 kb as the average gene size; and 472 Mb as the genome size of L. japonicus [13]. As L. japonicus is a perennial plant and can be propagated by cuttings, harvesting 200,000 seeds from the identified plants possessing active LORE1 is feasible. We have started to establish a small-sized tagging population to test the system. Other transposable elements, activated in the same way as LORE1, might be identified in the course of establishment of this population; LORE2 [29] is one such candidate.
After the submission of this article, Tsukahara et al. reported the identification of a Gypsy element transposed in intact ddm1 mutant plants of Arabidopsis thaliana [51]. Precise characterization of the behavior of the Gypsy element, together with that of LORE1, will facilitate our understanding of the interaction between LTR retrotransposons and plant genomes.
The Gifu accession of Lotus japonicus was used to generate both the transgenic and regenerated populations. The MG20 accession was used in the reciprocal crosspollination experiment with the T0 plant exhibiting LORE1 activity. For promoter analysis of LORE1 LTR using Arabidopsis thaliana, the ecotype Columbia was used to generate transgenic plants.
Transgenic and regenerated plant populations were produced from the Gifu accession using two different protocols. Populations 1 and 2 were generated according to the method described in [52]. Populations 3, 4, and 5 were generated following the method described in [53]. Antibiotic selection was not used when populations 2 and 5 were produced.
The 225 bp LTR of LORE1a, corresponding to the region from 137 bp to 361 bp of the AJ966990 sequence, was cloned into a multi-cloning site upstream of an intron-containing GUS gene in the binary vector pZN-GUS [54]. The resulting plasmid was introduced into Agrobacterium tumefaciens strain EHA105. Arabidopsis thaliana ecotype Colombia was subsequently infected to generate transgenic plants following the method described in [55]. L. japonicus Gifu accession was infected with the same Agrobacterium strain and transgenic plants were generated following the method described in [53].
Genomic Southern blots were carried out following the method described in [29]. Hind III was used to digest genomic DNA in the Southern blot analyses shown in Figure 1 and Figure S2. Hind III and Alu I were used to digest genomic DNA in Southern blot analyses shown in Figure 5 and Figure S4. Washes were performed at high stringency (65°C, 0.1x SSC, 0.1% SDS). The DNA probe used in Figure 1 and Figure S2 was generated by PCR using the primer pair LORE1gagF (5′-GTTGCCAGTATCGCCATGGACG-3′) and LORE1gagR (5′-GGATTGAGGCCTCCAAGATAAC-3′), and BAC DNA containing LORE1a [28]. The DNA probe used in Figure 5 and Figure S4 was generated by PCR using the primer pair 5′FLKF (5′-TTGACCTGCTCTTCAGTGCATG-3′) and 5′FLKR (5′-GAATCCGGGTATAAGGGTTCC-3′). The Megaprime DNA Labeling System (GE Healthcare) was used for labeling the DNA probes with alpha-32P-dCTP.
SSAP analyses to detect new LORE1 insertions were conducted as described in [28]. In brief, genomic DNA was digested with Mse I (New England Biolabs), and ligated with Mse I adapters. The first PCR was conducted using a primer annealing to a internal region of LORE1 and oriented outward, and a primer specific to the Mse I adapters. A nested PCR was conducted using the first PCR reaction as template. The amplified SSAP fragments were electrophoresed on polyacrylamide sequencing gels, and detected by silver staining. Bands for putative new insertions, i.e., absent from control Gifu analyses, were excised using a scalpel, boiled in 1x PCR buffer, and then used as a template to reamplify the fragment using the same primer pairs as in the nested PCR of the SSAP reaction. The reamplified fragments were electrophoresed on 1% agarose gels, excised, and extracted from the gel using Wizard SV Gel and PCR Clean-up System (Promega). Cleaned fragments were sequenced using a BigDye Terminator v3.1 Cycle Sequencing Kit (Applied Biosystems). The reamplified fragments would be expected to contain the junction sequence between LORE1 and its flanking DNA, in which Mse I sites are absent. Sequences that contained Mse I sites were regarded as artifacts and not subjected to further analyses. To amplify junctions between the flanking DNA and LORE1, we designed primers specific to the flanking sequences obtained and oriented toward LORE1. When genome sequences corresponding to flanking DNA were available on the database, they were utilized to design primers. We confirmed that amplifications were successful for plants from which the SSAP fragments were recovered, but not from the parent plant or control Gifu accession.
DNA sequences corresponding to regions 1 and 2 in newly transposed LORE1 elements were obtained by direct sequencing of PCR products. These were amplified by primers specific to the 5′ flanking sequences of each LORE1 element and primer 4 (5′-CAACAGTAGTATCAAATGTAGG-3′), as indicated in Figure 1A, using a BigDye Terminator v3.1 Cycle Sequencing Kit. The primers used for sequencing region 1 were Reg1F (5′-AGTAGCACCTGTAACAGTGGAG-3′) and Reg1R (5′-CATTAAGAGAGACTTTAGGAAC-3′), and those for region 2 were Reg2F (5′-CCTCCAACATTGTCAGTGATAG-3′) and Reg2R (5′-TAGCTGTAAAGCTCCTGTCCAC-3′). In the reciprocal cross analysis shown in Figure 1D, PCR reactions were performed using Primer 1 (5′-GACTAAGTGCCTCTTCAACTGC-3′) and Primer 2 (5′-GACTAAGTGCCTCTTCAACTGC-3′) to amplify LORE1a from Gifu, and Primer 1 and Primer 3 (5′-CACCTGACGATGCTAGCCTTGG-3′) to amplify the region allelic to LORE1a (absence of LORE1) from MG20 (see Figure 1 legend).
Genomic DNA samples were extracted from the leaves of T0 plants. Sodium bisulfite treatment of the DNA was conducted using a BisulFast Methylated DNA Detection Kit (TOYOBO), following the manufacturer's instructions. Briefly, 1 µg of column-purified genomic DNA was digested with Eco RI, treated with Proteinase K, and then subjected to bisulfite modification. Bisulfite-treated DNA (1 µl) was used as template for PCR reactions. Primary and nested PCR reactions were conducted for each LORE1 locus. The following primers were used for the primary PCR reactions: BSF R1 (5′-CTCTRAAACCTTRTTRCTTCARCCAT-3′) in combination with BSFa F (5′-TAAAAGAGAATYTGGGTATAAGGGAA-3′) for LORE1a; BSFb F (5′-TTYAAAGGTGYAGTYTYAATTGTATT-3′) for LORE1b; and BSFf F (5′-AGGGAGAYGAYAGTGATGGTGTTTT-3′) for LORE1f. For nested PCR reactions, 1 µl of the primary PCR reaction was used as template, with the following primers: for LORE1a, BSF R2 (5′- CCATRATTCRCTCCTCCRCTTCAC-3′) and BSFa F; for LORE1b, BSF R2 and BSFb F; and for LORE1f, BSF R2 and BSFf F. PCR reactions (20 µl) were conducted as follows: incubation at 94°C for 2 min as an initial denaturation step, 30 cycles of 30 s at 94°C, 45 s at 55°C, and 45 s at 72°C for amplification, and incubation at 72°C for 5 min. Amplified fragments were TA cloned using the pGEM-T Easy Vector System (Promega). For LORE1a, 6 to 8 TA clones were obtained from each of three PCR reactions and, in total, between 20 and 24 clones were sequenced for each plant analyzed. For LORE1b, 12 clones obtained from a PCR reaction were analyzed for each plant examined. For LORE1f, 11 or 12 clones obtained from a PCR reaction were analyzed for each plant examined.
A method modified from [56] was used for RNA isolation from plant tissues. Ground tissues (∼0.1 g) were incubated with 700 µl of extraction buffer (2% ß-mercaptoethanol, 2% hexadecyltrimethylammonium bromide, 100 mM Tris-HCl [pH 8.0], and 25 mM EDTA) at room temperature for less than 5 min. The recovered RNA was treated with 5 U DNase I at 37°C for 30 min in a 100 µl reaction. DNase-treated RNA was purified and recovered using an RNeasy Mini Kit (QIAGEN), with additional DNase treatment performed on a column, following the manufacturer's instructions. For RT-PCR, cDNA was synthesized by ReverTra Ace α (TOYOBO) using 1 µg of purified total RNA and oligo (dT) 20 primer in a 20 µl reaction. A 5× dilution of the cDNA reaction (2 µl) was used as template for semi-quantitative RT-PCR in a 20 µl PCR reaction using Ex Taq (TaKaRa) and 5 pmoles of each primer. The primers LORE1gagF and LORE1gagR were used for detection of LORE1 transcripts and products were amplified with 28 PCR cycles. As a control, the primers EF1αF (5′-GTGAGGGACATGAGACAGACTG-3′) and EF1αR (5′-AAATAGCAGTGTAGGACAAGTC-3′) were used for detection of transcripts of elongation factor 1 alpha, and these reactions required 24 PCR amplification cycles. To identify the transcription of LORE1 members, RT-PCR amplifications of regions 1 and 2 were conducted using the primer pairs Reg1 F and Reg1 R, or Reg2 F and Reg2 R, respectively.
BLAST searches were used to identify sequences homologous to SSAP fragments. These were conducted using Miyakogusa jp (http://www.kazusa.or.jp/lotus/), NCBI BLAST (http://blast.ncbi.nlm.nih.gov/Blast.cgi), and Phytozome Glycine max (http://www.phytozome.net/soybean). Pfam was accessed at http://pfam.sanger.ac.uk/. Bisulfite sequencing data was analyzed using QUMA [57] and CyMATE [58].
|
10.1371/journal.pgen.1008065 | Floral regulators FLC and SOC1 directly regulate expression of the B3-type transcription factor TARGET OF FLC AND SVP 1 at the Arabidopsis shoot apex via antagonistic chromatin modifications | Integration of environmental and endogenous cues at plant shoot meristems determines the timing of flowering and reproductive development. The MADS box transcription factor FLOWERING LOCUS C (FLC) of Arabidopsis thaliana is an important repressor of floral transition, which blocks flowering until plants are exposed to winter cold. However, the target genes of FLC have not been thoroughly described, and our understanding of the mechanisms by which FLC represses transcription of these targets and how this repression is overcome during floral transition is still fragmentary. Here, we identify and characterize TARGET OF FLC AND SVP1 (TFS1), a novel target gene of FLC and its interacting protein SHORT VEGETATIVE PHASE (SVP). TFS1 encodes a B3-type transcription factor, and we show that tfs1 mutants are later flowering than wild-type, particularly under short days. FLC and SVP repress TFS1 transcription leading to deposition of trimethylation of Iysine 27 of histone 3 (H3K27me3) by the Polycomb Repressive Complex 2 at the TFS1 locus. During floral transition, after downregulation of FLC by cold, TFS1 transcription is promoted by SUPPRESSOR OF OVEREXPRESSION OF CONSTANS1 (SOC1), a MADS box protein encoded by another target of FLC/SVP. SOC1 opposes PRC function at TFS1 through recruitment of the histone demethylase RELATIVE OF EARLY FLOWERING 6 (REF6) and the SWI/SNF chromatin remodeler ATPase BRAHMA (BRM). This recruitment of BRM is also strictly required for SQUAMOSA PROMOTER BINDING PROTEIN-LIKE 9 (SPL9) binding at TFS1 to coordinate RNAPII recruitment through the Mediator complex. Thus, we show that antagonistic chromatin modifications mediated by different MADS box transcription factor complexes play a crucial role in defining the temporal and spatial patterns of transcription of genes within a network of interactions downstream of FLC/SVP during floral transition.
| The initiation of flowering in plants is exquisitely sensitive to environmental signals, ensuring that reproduction occurs at the appropriate time of year. The sensitivity of these responses depends upon strong repression of flowering under inappropriate conditions. FLOWERING LOCUS C (FLC) and SHORT VEGETATIVE PHASE (SVP) are related transcription factors that act in concert to strongly inhibit flowering in crucifer plants through repressing transcription of their target genes. Many direct FLC/ SVP targets have been identified in genome-wide studies, however few of these genes have been characterized for their roles in regulating flowering time or other aspects of reproductive development. Here, we characterize TARGET OF FLC AND SVP1 (TFS1) as a novel target of FLC and SVP, and demonstrate that TFS1 contributes to proper flowering-time control. Moreover, we provide a detailed mechanistic view of how TFS1 transcription is controlled during reproductive development through the repressive activity of FLC/SVP being overcome by the transcriptional activator SUPPRESSOR OF OVEREXPRESSION OF CONSTANS 1. Thus we further elucidate the network of genes repressed by FLC/SVP to block flowering and determine mechanisms by which their repressive activity is overcome during the initiation of flowering.
| The transition from vegetative to reproductive development in plants is controlled by a complex transcriptional network that responds both to environmental cues and endogenous hormonal signals [1, 2]. In Arabidopsis thaliana, the MADS-box transcription factor FLOWERING LOCUS C (FLC) plays a major role in this network as an inhibitor of floral transition [3, 4]. Transcription of FLC is repressed by extended exposure to cold that mimics winter conditions (vernalization) so that flowering can proceed when plants are subsequently exposed to warm in spring. The repression of FLC transcription by accumulation of histone modifications in response to cold has been extensively studied [5] and FLC target genes have been described by whole genome chromatin immunoprecipitation (ChIPseq) [6–8]. Nevertheless, our understanding of how FLC influences the transcriptional network that controls floral transition and how it represses expression of its target genes is still fragmentary. Here, we utilize data derived from genome-wide binding studies of FLC and its partner MADS box transcription factor SHORT VEGETATIVE PHASE (SVP) [7, 9–12] to identify a common target gene that we named TARGET of FLC and SVP 1 (TFS1). We show that this gene acts in the network downstream of FLC and other floral regulators, and has an important role on the flanks of the meristem during the early stages of floral development.
FLC binds to several hundred target genes, but only a small subset of these are conserved between species [8]. Genes involved in flowering control are enriched among the conserved targets, and FLC represses transcription of several of these, including FLOWERING LOCUS T (FT), SQUAMOSA PROMOTER BINDING PROTEIN-LIKE 15 (SPL15), SUPPRESSOR OF OVEREXPRESSION OF CONSTANS 1 (SOC1) and SEPALLATA 3 (SEP3). MADS box transcription factors are proposed to bind DNA as dimers and tetramers [13, 14], and FLC interacts with and binds to a subset of its targets in complexes with the related protein SHORT VEGETATIVE PHASE (SVP) [7, 10]. Mutants at SVP are early flowering and exhibit increased levels of SOC1 and FT mRNAs [9, 10, 15, 16]. How FLC represses transcription of its targets is not completely clear, but appears to involve modification of histones. FLC and its homologue FLOWERING LOCUS M (FLM) recruit EMBRYONIC FLOWER 1 (EMF1), a plant-specific Polycomb Repressive Complex 1 (PRC1) component [17, 18], to FLOWERING LOCUS T (FT) in leaf veins to repress its transcription [19]. Moreover, cooperativity between FLC or FLM and PRC1 contributes to maintenance of PRC2-induced trimethylation of lysine-27 at histone H3 (H3K27me3) at FT chromatin, most probably through the activity of the histone methyltransferase CURLY LEAF (CLF) and the H3K27me3-associated protein LIKE HETEROCHROMATIN PROTEIN1 (LHP1) [19]. Additionally, the JmjC domain-containing trimethyl histone H3 lysine-4 demethylase JUMONJI14/PKDM7B (JMJ14/PKDM7B) also associates with PRC1 to further antagonize the active chromatin state at FT [20, 21].
Other targets of FLC and SVP that are involved in floral induction encode the MADS-box transcription factor SOC1 and the plant-specific transcription factor SPL15 [22–27]. Both SOC1 and SPL15 are expressed in the shoot apical meristem where they cooperate at the promoters of target genes such as FRUITFULL (FUL) and MIR172b to activate the basal floral promotion pathway under non-inductive environmental conditions [12, 22, 28]. Interestingly, SOC1 coordinates the recruitment of the histone demethylase RELATIVE OF EARLY FLOWERING 6 (REF6) to the promoter of FUL and MIR172b to orchestrate the removal of the H3K27me3 mark and activate transcription [22]. Furthermore, SPL15 activity is repressed post-transcriptionally by miR156 and post-translationally through its physical interaction with the gibberellin (GA)-labile DELLA protein REPRESSOR OF GA1-3 (RGA) [22]. By contrast, SPL9, a paralogue of SPL15 that is expressed after floral induction at the periphery of the SAM, requires interaction with DELLA proteins to potentiate its trans-activation activity and contribute to the induction of expression of the floral meristem identity gene APETALA1 (AP1) in the low gibberellin context present in the cells that give rise to the floral primordium [29].
Here, we show that TFS1, which encodes a B3-type transcription factor that is a member of the REPRODUCTIVE MERISTEM (REM) family [30, 31], constitutes a target of FLC and SVP and that its transcription is repressed through cooperation with PRC complexes during vegetative growth. After floral induction, TFS1 expression is induced at the periphery of the SAM by SOC1 and the age-regulated transcription factor SPL9 through coordinated recruitment of the histone demethylase REF6 and the chromatin remodeler BRAHMA (BRM) [32–34]. This analysis deepens our understanding of the mechanism by which FLC represses the floral network and indicates the importance of antagonistic histone modifications mediated by different MADS box complexes on common target genes.
To further define the regulatory network controlling floral transition at the shoot apex, recently published ChIP-Seq and tissue-specific RNA-Seq datasets were examined to identify genes that are expressed specifically at the shoot apex and are bound by the floral repressor transcription factors FLC and SVP [6–8]. Cross-referencing these datasets identified the gene encoding the B3-type transcription factor TARGET OF FLC AND SVP1 (TFS1), which was formerly annotated as REPRODUCTIVE MERISTEM 17 (REM17) [30, 31]. A phylogenetic analysis revealed that TFS1 is part of a gene family including REDUCED VERNALIZATION RESPONSE1 (VRN1), VERDANDI (VDD) and VALKYRIE (VAL), and that REM18 and REM19 are the closest homologs of TFS1 within a sub-branch of the phylogenetic tree (S1A Fig). To verify the binding of FLC and SVP at TFS1, chromatin-immunoprecipitation (ChIP) analysis was performed using chromatin extracted from the aerial parts of 15-day old plants grown under inductive long days (LDs) and using antibodies that were raised against FLC [8] and SVP (S1B Fig, S3 and S4 Tables, Methods). In agreement with the ChIP-seq data, specific enrichment of a fragment that encompasses the putative CArG-boxes, designated as CArG-box II, located at the 3’ end of TFS1 was detected after ChIP of FLC or SVP (Fig 1A and 1B). Moreover, the ChIP-qPCR analyses also demonstrated a mutual co-operation between these floral repressors at TFS1, because binding of FLC and SVP was enhanced in the presence of the other protein (Fig 1A and 1B), as reported previously for several other targets of these transcription factors [7, 10].
To further characterize the regulation of TFS1 by FLC and SVP, the abundance of TFS1 mRNA in leaves and apices was tested by RT-qPCR in Col-FRI flc-3, Col-FRI svp-41 and Col-FRI flc-3 svp-41 mutants as well as Col-FRI wild-type. TFS1 mRNA was exclusively detected at the shoot apex and its abundance was increased throughout a developmental time course in the single mutants compared to wild-type and was strongly increased in the double mutant (Fig 1C, S1C Fig). Furthermore, TFS1 expression was increased at the shoot apex, but not in leaves, after Col-FRI plants were exposed to vernalisation and returned to normal growth temperature (S1D Fig), consistent with the repression of TFS1 transcription by FLC. To determine whether the spatial pattern of expression of TFS1 differs in Col-FRI flc-3 and Col-FRI svp-41 mutants compared to Col-FRI, in-situ hybridisation analysis of the shoot apex was performed during floral transition after transfer of plants from short days (SDs) to inductive LDs (Fig1D, S1E Fig). The Col-FRI and Col-FRI flc-3 plants grown for 2 wks under SD were still vegetative (Fig 1D), but TFS1 expression level was clearly elevated in Col-FRI flc-3 (S1C Fig), indicating that FLC represses TFS1 expression even in the vegetative stage. After transfer to LDs, TFS1 mRNA appeared at the periphery of the SAM in the Col-FRI wild-type. This spatial pattern was not changed in the Col-FRI svp-41, Col-FRI flc-3 or Col-FRI flc-3 svp-41 mutants, but the mRNA appeared more rapidly after transfer to LDs in the mutants than in Col-FRI wild-type (Fig 1D, S1C Fig). Therefore, TFS1 transcription is induced during floral transition, while the timing and amplitude of its expression are modulated by SVP and FLC.
The temporal expression pattern of TFS1 coincided with the transition to flowering, so the flowering time of tfs1 mutants was determined under inductive and non-inductive conditions. Interestingly, tfs1-1 mutants flowered significantly later than wild-type plants under both conditions, suggesting that TFS1 is involved in promoting floral transition (Fig 1E and 1F, S2A–S2F Fig). The role of TFS1 was confirmed by transgenic complementation of tfs1-1 (see later), and by showing that a second allele (tfs1-2) caused a similar late-flowering phenotype under LDs (S2E and S2F Fig). The tfs1-1 mutation also delayed flowering in the Col-FRI flc-3 svp-41 background, supporting the idea that TFS1 acts downstream of FLC and SVP to promote flowering (Fig 1E and 1F). In addition, the Col-FRI flc-3 svp-41 tfs1-1 plants showed impaired flower development, suggesting redundancy between these genes in flower and inflorescence development (S2G and S2H Fig). Overall, these results demonstrate that TFS1 is a direct target of FLC and SVP and is specifically expressed at the periphery of the SAM during floral transition to promote flowering and floral development.
Transcriptional repression by FLC and SVP has been linked to activity of Polycomb Repressive Complex (PRC) 2 (PRC2) [10, 19]. PRC2 catalyzes the methylation of histone 3 (H3) at lysine 27 (H3K27me3) and is associated with the repression of transcription [35]. Therefore, whether TFS1 is subjected to PRC-mediated regulation in a FLC and SVP dependent manner was tested by performing ChIP-qPCR using chromatin extracted from 15-day old plants grown under LDs and antibodies directed against H3K27me3 or H3K4me3 (S3 Table). H3K27me3 was detected in the gene body of TFS1 in Col-FRI plants at much higher levels than in Col-FRI flc-3, Col-FRI svp-41 and Col-FRI flc-3 svp-41 mutants (Fig 2A). However, in Col-FRI flc-3 svp-41 plants an additional peak in H3K27me3 levels was detected close to the transcriptional termination site of TFS1 (Fig 2A, S3A Fig). Furthermore, a commercially available antibody (S3 Table) was used for ChIP-qPCR of LHP1, which is frequently found associated with H3K27me3 marked chromatin [36, 37], and the protein was detected at TFS1 in a similar pattern to H3K27me3 (Fig 2B). Therefore, H3K27me3 and LHP1 are present in the gene body of TFS1 in a FLC and SVP dependent manner correlating with reduced transcription of TFS1.
In contrast to H3K27me3, K4-trimethylated H3 (H3K4me3) is present at genes that are actively transcribed [38]. Enrichment of H3K4me3 was detected close to the transcriptional start site (TSS) of TFS1 in the Col-FRI flc-3, Col-FRI svp-41 and Col-FRI flc-3 svp-41 mutants, whereas it was not present at the gene in Col-FRI plants (S3B Fig). These data are in agreement with previous reports of the dynamic and antagonistic relationship between H3K27me3 and H3K4me3 during development and the floral transition [35, 39]. Therefore, under these conditions the dynamic change in these chromatin marks at TFS1 correlates with the repression of FLC transcription, the induction of TFS1 and the transition to flowering.
The PRC2 mutation curly leaf (clf), which impairs the activity of an enzyme that catalyses H3K27me3 deposition [40], partially suppresses the late-flowering phenotype of Col-FRI plants [41]. Strikingly, early flowering Col-FRI clf-2 mutant plants expressed highly elevated levels of TFS1 mRNA in the apex, while still expressing high levels of FLC and SVP mRNA (S3D–S3F Fig) [41, 42]. These observations supported the idea that PRC2 might contribute to transcriptional repression of FLC target genes such as TFS1. To examine this possibility, ChIP analyses were performed to test for H3K27me3 enrichment across the coding region of TFS1 in Col-FRI clf-2 and Col-FRI flc-3 svp-41 mutant plants. In agreement with the functional role of CLF in the deposition of H3K27me3, a strong reduction in this mark at TFS1 was detected in Col-FRI clf-2 and this was similar to the reduction observed in Col-FRI flc-3 svp-41 mutant plants when compared to wild-type Col-FRI (Fig 2C). In contrast, ChIP analyses for H3K4me3 at TFS1 showed higher enrichment patterns in Col-FRI clf-2 as well as in Col-FRI flc-3, Col-FRI svp-41 and Col-FRI flc-3 svp-41 mutants than in wild-type Col-FRI (S3B and S3C Fig). These data indicate that reduction in H3K27me3 at TFS1 in Col-FRI clf-2 mutants correlates with an increase in H3K4me3, consistent with the antagonistic role of these marks during floral transition. Finally, to determine whether the reduction in H3K27me3 levels at TFS1 in Col-FRI clf-2 plants correlated with reduced FLC binding, ChIP analyses were performed for FLC. FLC binding was strongly compromised in Col-FRI clf-2, indicating that FLC binding requires and is sustained by PRC2 function (Fig 2D).
SVP and FLC interact respectively with LHP1 and the Polycomb Repressive Complex 1 protein EMBRYONIC FLOWER (EMF1) [18, 19, 43, 44], suggesting a link between PRC function and the activity of these floral repressors. Furthermore, a complex including LHP1, EMF1 and the H3K4me3 demethylase JMJ14/PKDM7B, has been described to play roles related to PRC1, including delaying flowering in non-inductive photoperiods [19–21]. These observations, together with the result described above that H3K4me3 levels are lower at TFS1 in the presence of SVP and FLC (S3B Fig), suggested that the H3K4me3 demethylase activity of JMJ14 might be required for FLC and PcG mediated repression of TFS1. To test this idea, ChIP-qPCR analyses of H3K27me3 were performed on TFS1 in Col-FRI jmj14-2 mutants. In these plants the enrichment levels of the repressive mark H3K27me3 were strongly reduced compared to Col-FRI wild-type (Fig 2E). By contrast, the active chromatin mark H3K4me3 was increased at TFS1 in Col-FRI jmj14-2 compared to Col-FRI and was present at a similar level as in Col-FRI flc-3 svp-41 mutants (S3C Fig). Consistent with the increased levels of H3K4me3, Col-FRI jmj14-2 also showed higher mRNA levels of TFS1 in apices, but did not affect the expression of SVP or FLC (S3G–S3I Fig). In support of the notion that FLC binding to TFS1 requires PRC2 activity and higher levels of H3K27me3 (Fig 2D), reduced binding of FLC to CArG-box II at TFS1 was also detected in Col-FRI jmj14-2 plants, although FLC mRNA level was unaffected (Fig 2F, S3H Fig). Collectively, these data suggest that for FLC mediated transcriptional repression of TFS1, the recruitment of PRC2 and deposition of H3K27me3 are required, and that these can also be inhibited by increasing the levels of H3K4me3 through mutation of the JMJ14 demethylase.
The observations that FLC and SVP bind 3’ of the TFS1 stop codon (Figs 1A, 1B, 2D and 2F) and that they interact with a PRC complex and LHP1 [19, 43] that are associated with the gene body of TFS1 (Fig 2B), suggested that a chromosomal loop might form between the 3’ distal region and the gene body. To test for the presence of such a loop, a chromosome conformation capture (3C) assay was performed. Indeed, in Col-FRI plants fragments B, C and D (middle region) were found to interact with the 3’ end (G and H) of TFS1, suggesting that a ‘locked’ DNA loop was formed (Fig 3). The presence of this loop was then tested in different mutants to determine whether it required FLC/SVP and PcG function. In Col-FRI flc-3 svp-41 mutants, the interaction between the middle region and the 3’ end was strongly reduced compared to Col-FRI, indicating that FLC and SVP are required for the formation of the ‘locked’ DNA loop (Fig 3A–3D). To define the contribution of PRC2 in the formation of this loop at TFS1, Col-FRI clf-2 plants were examined. In this genotype, the loop appeared significantly weaker than in Col-FRI plants, indicating that CLF activity is required to support the formation of the ‘locked’ DNA loop at TFS1 (Fig 3E–3G). These data indicate therefore that transcriptional repression of TFS1 by FLC and SVP is associated with the formation of a chromatin loop that requires FLC and SVP binding at the 3’ end of the gene and high levels of H3K27me3 within the gene body.
In a genome-wide study of binding sites of the MADS-box protein SOC1, a site at the 3’ end of TFS1 was detected [28]. To determine whether TFS1 is regulated by SOC1, TFS1 transcript abundance was tested by RT-qPCR using RNA extracted from leaves and apices in soc1-2, soc1-2 svp-41 and Col genotypes. TFS1 mRNA levels were much lower in apices of soc1-2 mutants than Col, but were largely restored to Col levels in soc1-2 svp-41 double mutants (Fig 4A). To determine the spatial pattern of TFS1 expression in soc1-2 and soc1-2 svp-41, in-situ hybridisations were performed on apices during floral transition. The overall spatial expression pattern was similar in Col and soc1-2 svp-41, however, a significant delay in TFS1 expression at the periphery of the SAM was observed in soc1-2 (Fig 4B).
ChIP analyses were then performed with antibodies that were directed against endogenous SOC1 and SVP (S1B Fig, S3 Table, [22]). To verify the results of the previous report on the genome-wide study for SOC1, ChIP-qPCR was performed and detected SOC1 binding in the region of CArG-box I (CArGI), which is located at the 3’ end of TFS1 (Figs 1A and 4C). The ChIP-qPCR experiment with SVP generated a specific enrichment in the region of CArG-box II (CArGII) in Col, and this was enhanced in the soc1-2 mutant, suggesting that SOC1 reduces SVP recruitment to TFS1 (Fig 4D). Similarly, Dexamethasone (DEX)-induced translocation of SOC1:GR into the nucleus in 35S::SOC1:GR plants caused higher TFS1 transcription and increased binding of SOC1 to CArGI as well as reduced binding of SVP to CArGII (S4A–S4C Fig). Therefore, binding of SOC1 to the 3’ end of TFS1 occurs during floral transition, whereas SVP binds during vegetative development, and is in agreement with their observed roles in the transcriptional regulation of TFS1.
To test in vivo whether the CArG-boxes identified within the ChIP-qPCR amplicons are responsible for the regulation of TFS1 by SOC1 and SVP, a TFS1::TFS1:9xAla-Venus (TFS1::TFS1:9AV) gene fusion was constructed that contained the entire intergenic region flanking TFS1 on the 5’ and 3’ sides (Fig 4E). This gene fusion complemented the tfs1-1 mutant phenotype (S5A and S5B Fig) and in the transgenic plants VENUS signal was detected at the periphery of the SAM in a similar pattern as observed by in situ hybridization of TFS1 mRNA (Figs 1D and 4F, S4E Fig). Also, the confocal imaging of the TFS1:9xAla-Venus fusion protein indicated that it predominately localized in the cytosol of slowly dividing meristematic cells (S6A and S6B Fig) while treatment with leptomycin B (LMB), which impairs the activity of nuclear exportin [45], suggested that this was due to active export from the nucleus (S6C Fig). However, in actively dividing cells in young sepals the VENUS signal appeared to be nuclear (S6D to S6G Fig), suggesting the nuclear accumulation and presumably the activity of TFS1 may be closely related to cell division.
The CArG-boxes identified in the ChIP amplicons were then mutated in this gene fusion construct. Two mutant plasmids were generated in which CArGII or both CArGI and CArGII were mutated (Fig 4E). Several independent transformants carrying each construct were analysed (Fig 4F, S4E–S4H Fig). Transformants harbouring the mCArGII construct displayed a stronger and broader VENUS signal than those carrying the wild-type construct, whereas the mCArGI+II construct conferred a VENUS signal that was similar to the wild-type construct (Fig 4F, S4E–S4G Fig). The relative strength of these mutant constructs was supported by RT-qPCR analysis performed on RNA extracted from apices of the transgenic plants (S4H Fig). Furthermore, the strong signal of the mCArGII construct was greatly reduced when it was introduced into the soc1-2 mutant by crossing (Fig 4F). The low level of TFS1:9AV expression detected in the soc1 TFS1::TFS1:9AV mCArGII and TFS1::TFS1:9AV mCArGI+II plants was consistent with their delayed flowering time compared to TFS1::TFS1:9AV plants (S5A–S5H Fig). Taken together, these observations were consistent with the transcriptional profile of TFS1 in Col and soc1-2 svp-41 (Fig 4A), and supported the proposal that SOC1 activates and SVP represses transcription of TFS1 at least partly through binding to CArG-box I and CArG-box II, respectively.
Consistent with the role of SOC1 in promoting TFS1 transcription, increased levels of the repressive mark H3K27me3 were detected in aerial parts of 15-day old soc1-2 mutants across the TFS1 gene body (Fig 4G). Furthermore, in DEX-induced 35S::SOC1:GR plants, H3K27me3 levels were reduced following binding of SOC1 (S4D Fig). Additionally, the presence of the active chromatin mark H3K4me3 and of RNA polymerase II (RNAPII) were monitored along the transcribed region of TFS1 in soc1-2 mutants. ChIP-qPCR analysis revealed that H3K4me3 levels were strongly reduced in soc1-2 mutants, similar to those observed in Col-FRI wild-type (Figs 2A and 4H). Also, the ChIP profile of RNAPII demonstrated enrichment throughout the transcribed region of TFS1 in Col and soc1-2 svp-41. By contrast, in soc1-2 mutants the loss of RNAPII enrichment was most apparent in regions of the gene body (Fig 4I), which is reminiscent of inactive genes that display a poised RNAPII machinery at promoter regions.
Overall, these experiments demonstrate that SOC1-induced activation of TFS1 transcription is invoked by eviction of SVP and reduction in H3K27me3 as well as by releasing RNAPII to transcribe the gene.
The histone demethylase REF6 and the chromatin remodeler BRM physically interact to antagonize PcG proteins at target loci, and both proteins were detected at the 3’ end of TFS1 in a genome-wide study [33, 34, 46]. Furthermore, REF6 and SOC1 co-purified in the same complex, which was required to facilitate transcriptional activation of target genes through the removal of the repressive histone mark H3K27me3 [22, 47]. Thus, to understand the effects of REF6 and BRM on TFS1 transcription, TFS1 transcript abundance was monitored by RT-qPCR in ref6-1 and brm-1 mutants grown for 9 to 17 days in LDs. Throughout the time-course, the transcript profile of TFS1 was not changed in leaves of ref6-1 or brm-1 mutants compared to Col, but a dramatic reduction in TFS1 transcript abundance was detected in apices of both mutants (Fig 5A and 5D). Therefore, REF6 and BRM are required for the activation of TFS1 in apices.
To validate the previously reported binding of REF6 and BRM to TFS1 [46], ChIP-qPCR analyses were performed using REF6::REF6:HA ref6-1 and BRM::BRM:HA brm-1 transgenic lines [47, 48]. Binding of REF6:HA and BRM:HA to sites located at the 3’ end of TFS1 was detected and these sites flank CArGI, to which SOC1 binds (Fig 5B and 5E). To understand whether association of REF6 and BRM with chromatin is dependent on SOC1 (Fig 4C), the soc1-2 mutation was introduced into the REF6::REF6:HA ref6-1 and BRM:BRM:HA brm-1 transgenic lines by genetic crossing. A strong reduction in binding of REF6:HA and BRM:HA was detected by ChIP-qPCR in soc1-2 mutants, indicating that SOC1 supports REF6 and BRM binding to the 3’-end of TFS1 (Fig 5B and 5E). Additionally, the BRM:BRM:HA brm-1 transgene was introduced into the ref6-1 mutant to study binding behaviour of BRM:HA to the 3’ end of TFS1. Chromatin association of BRM:HA was strongly reduced in ref6-1 compared to Col (Fig 5G), which further corroborated the idea that REF6 is required for BRM recruitment.
REF6 is a H3K27me3 demethylase [47] and BRM and REF6 appeared to act as direct activators of TFS1 transcription, so H3K27me3 levels were tested at TFS1 in the respective mutants. Compared to Col, increased H3K27me3 levels were detected by ChIP-qPCR along the TFS1 genomic locus in ref6-1 and brm-1 mutants. The pattern of increase of H3K27me3 was identical in both mutants and consistent with the observed increase in soc1-2 mutants (Figs 4G, 5C and 5F). In summary, these findings suggest that the histone demethylase REF6 and the chromatin remodeler BRM are required to activate transcription of TFS1 in association with SOC1.
To test whether BRM recruitment leads to activation of TFS1 through changes in chromatin accessibility, limited Micrococcal nuclease (MNase) digestion followed by tiled oligo qPCR was employed to identify well-positioned nucleosomes near the SOC1, REF6 and BRM bound site at the 3’ end of TFS1. The MNase-qPCR analysis identified in ref6-1 and brm-1 mutants a nucleosome at a position that encompasses the binding site for SOC1 and this nucleosome was destabilized in Col (Fig 5H and 5I). Therefore, SOC1 appears to increase chromatin accessibility and transcription of TFS1 through recruitment of REF6 and BRM.
The spatial and temporal expression patterns of TFS1 appeared similar to those of SPL9 [22, 49, 50], which encodes a transcription factor that binds to regulatory sequences in the promoter of the floral meristem-identity gene APETALA1 (AP1) to regulate floral fate [29, 50]. Moreover, SPL15 and SOC1 co-operate to regulate floral commitment under non-inductive conditions [22]. Taken together, the data suggested that SPL9 and SOC1 might cooperate to activate TFS1. To test for the molecular effect of SPL9, RNA was extracted from apices of SPL9::GFP:rSPL9 to monitor TFS1 transcript abundance by RT-qPCR. TFS1 transcript levels were strongly increased in SPL9::GFP:rSPL9 plants compared to wild-type (Fig 6A). Next, ChIP-qPCR analysis was employed to test binding of GFP:rSPL9 at the 5’ and 3’ ends of TFS1 (Fig 6B, S7A and S7B Fig). Consistent with direct activation of TFS1 by SPL9, fragments at the 5’ and 3’ ends of the gene were enriched after immunoprecipitation of GFP:rSPL9 (Fig 6B, S7A and S7B Fig).
The presence of markers for transcriptional activity at TFS1, particularly the Mediator head-module component Med18, RNAPII and H3K4me3, was scored by ChIP-qPCR in different genotypes. Higher enrichment levels of these markers were detected at TFS1 in SPL9::GFP:rSPL9 than in wild-type, supporting that SPL9 activates TFS1 (S7C–S7F Fig). In contrast, reduced TFS1 transcript abundance was detected in spl9-1, spl15-1 and spl9-1 spl15-1 mutants (S8A–S8E Fig) and this was accompanied with a reduction in H3K4me3 and an increase in H3K27me3 at the TFS1 locus (S8F–8H Fig).
Whether the identified binding sites for SPL9 are responsible for in vivo regulation of TFS1 was then examined. To this end, a reporter gene cassette was constructed, mGTACa1/2, in which the two GTAC motifs overlapping with the ChIP-qPCR peak of GFP:rSPL9 at the 5’ end of TFS1 were mutated (Fig 6C). Transformants harbouring the mGTACa1/2 mutated form were generated and compared by confocal microscopy with plants harbouring a wild-type construct. VENUS fluorescent signal detected at the periphery of the SAM in wild-type was missing in the mGTACa1/2 plants, supporting the idea that SPL9 binds to these sites to activate transcription (Fig 6C, S7G Fig). Surprisingly, however, VENUS fluorescent signal was retained in the epidermis of mGTACa1/2 plants, indicating that expression in these cells likely takes place independently of SPL9 and other SPL transcription factors (Fig 6C, S7G Fig). Additionally, in-situ hybridisation indicated that TFS1 mRNA appeared more rapidly on the flanks of the meristem after transferring SPL9::GFP:rSPL9 plants from SDs to LDs than after transferring Col wild-type (S7F Fig). These studies are consistent with SPL9 binding to the 5’ end of TFS1 to activate transcription at the periphery of the SAM.
In support of the notion of functional cooperativity between SPL9 and SOC1, co-immunoprecipitation of GFP:rSPL9 and SOC1:MYC(9x) was detected in protein extracts from shoot apical tissue of SPL9::GFP:rSPL9 35S::SOC1:MYC(9x) transgenic lines (Fig 6D). The cooperativity between SPL9 binding at the 5’ end of TFS1 and SOC1 binding at the 3’ end suggested that DNA loop formation might occur between their binding regions. Therefore, chromosome conformation capture (3C) was employed to test for DNA loop formation in Col and 35S::miR156b, in which several redundant SPL transcription factors are reduced in expression [51] leading to reduced transcription of TFS1 (Fig 6E). The 3C analyses suggested that interaction between SPL-binding sites located at the 5’ end and the CArG-box predicted to bind SOC1 that is located at the 3’ end of TFS1 occurred in a SPL9 dependent manner (Figs 4C, 6F and 6G). Together with results described above, the data suggest that SPL9 cooperates with SOC1 to form an ‘active’ DNA-loop that is required for active TFS1 transcription.
The chromatin remodeler BRM is recruited to TFS1 in a SOC1 dependent-manner to increase chromatin accessibility and TFS1 transcription (Fig 5E and 5I). In addition, SOC1 is required for increased H3K4me3 at TFS1 (Fig 4H). This chromatin mark is also supported by the COMPASS-like (Complex Proteins Associated with Set1) histone H3 lysine-4 methyltransferase complex component WD40 REPEAT HOMOLOG 5 (WDR5), which associates with the active elongating RNAPII [52]. Therefore, ChIP analyses were performed using commercial antibody (S3 Table) that recognises WDR5a and WDR5b [53] to test WDR5 enrichment at TFS1 in different genotypes. Consistently, the presence of WDR5 at TFS1 was decreased across the gene body in soc1-2, brm-1 as well as spl9-1, spl15-1 and spl9-1 spl15-1 mutants (S9A–S9C Fig). Additionally, using commercial antibody (S3 Table) ChIP analyses for the histone variant H2A.Z, which marks both transcriptionally active and inactive genes [54, 55], detected colocalization of H2A.Z with WDR5 at TFS1 in Col and decreased enrichment in soc1-2 and brm-1 (S9D–S9F Fig). In contrast, no difference in H2A.Z enrichment was detected between spl mutants and Col, further corroborating the idea that SPL functions to orchestrate transcriptional machinery rather than influencing nucleosomal composition (S9D–S9F Fig).
The data described so far suggested that SOC1 mediated recruitment of BRM might enable association of SPL9 to chromatin. To test this idea, SPL9::GFP:rSPL9 was introduced into brm-1 mutants by genetic crossing. Unexpectedly, in SPL9::GFP:rSPL9 brm-1 most of the floral structures were converted into carpelloid structures at the primary inflorescence, a more severe phenotype than either parental line (S10A and S10B Fig). To further characterize the molecular effect, TFS1 transcript abundance was examined by RT-qPCR using RNA extracted from leaves and apices. The enhanced apex specific expression of TFS1 in SPL9::GFP:rSPL9 was strongly suppressed by brm-1, supporting the idea that BRM is required to support SPL9 activity (Fig 7A). Similarly, expression of other floral marker genes such as SOC1, FUL, LEAFY (LFY), AP1 and SEPALLATA3 (SEP3) was also reduced in this genotype, although SPL9 protein level was not affected (S10C–S10E Fig).
Consistent with the idea that BRM alters nucleosomal positioning leading to changes in the exposure of a critical SPL-binding site located at the 3’ end of TFS1 (Fig 5I), reduced binding of GFP:rSPL9 was detected to the 5’ and 3’ end of TFS1 in brm-1 (Fig 7B). This result suggested that BRM facilitates binding of SPL9 to its cognate binding sites. SPL15 recruits RNAPII through Mediator [22], and consistent with SPL9 having a similar role at TFS1, a strong reduction in the recruitment of MED18, RNAPII and markers of active transcription such as WDR5 and H3K4me3 was detected in brm-1 (Fig 7C–7F). Taken together, these data indicate that SOC1-dependent recruitment of BRM is required to allow SPL9 to bind to TFS1 and that Mediator conveys regulatory information from SPL9 to the basal RNAPII transcriptional machinery that is coupled with the COMPASS-like complex to activate TFS1 transcription.
The MADS box transcription factors FLC and SVP are well-established negative regulators of floral induction in Arabidopsis, however only fragmentary information is available on the roles of their direct targets in floral transition and the architecture of the regulatory network they control. Here, we addressed these issues by characterizing TFS1, an immediate target gene of FLC/SVP that encodes a B3-type transcription factor, which is expressed specifically on the flanks of the shoot apical meristem and promotes floral transition under LDs and SDs. TFS1 transcription is repressed by FLC/SVP and promoted by SOC1, another MADS box transcription factor that is also encoded by a primary target of FLC/SVP. We show that FLC/SVP and SOC1 have opposing effects on transcription through mediating antagonistic histone modifications at TFS1. These data provide insight into the complexity of the regulatory network controlling floral transition downstream of FLC/SVP and define mechanisms by which MADS box transcription factors antagonistically regulate transcription of their direct targets.
TFS1 is a member of the B3-type transcription factor superfamily that is specific to the Viridiplantae [56]. Within this superfamily, TFS1 falls in the REM family, several of which have established or proposed roles in reproduction of Arabidopsis [30, 31]. Loss of function alleles of two members of this family, VERNALIZATION 1 (VRN1) and VERDANDI (VDD), provided genetic support for roles in reproductive development [57, 58]. VRN1 is required for stable transcriptional repression of FLC during induction of flowering by vernalization [57, 59], and appears to bind DNA non-specifically [57], while VDD is involved in ovule development [58]. In addition, several other members of this family are specifically expressed in the inflorescence meristem or developing flowers [31, 60–62]. REM transcription factors and MADS box proteins, another family of transcription factors with multiple roles in reproductive development, appear to often regulate one another’s expression. For example, VRN1 regulates FLC, VDD transcription is controlled by SEEDSTICK, TFS1 is repressed by FLC/SVP and genome-wide studies of binding sites of MADS box factors AG, AP3, PI and AGL15 identified several REM genes as direct targets [31]. Both families of transcription factors are amplified in higher plants [30, 63], and they may have co-evolved to act in common pathways during the evolution of reproductive development.
The mechanism of action of REM proteins is not known, although they are believed to bind DNA via their B3 domains. A GFP-tagged form of VRN1 was found to associate widely with Arabidopsis chromosomes, and this association persisted through mitosis but was lost at meiosis [59]. Interestingly, TFS1 was also previously identified in a targeted proteomics approach as interacting with PCNA, a component of the DNA replication complex [64]. Also, our confocal imaging suggested that the nuclear localization and activity of TFS1 is closely related to cell division. The molecular functions of REM proteins such as TFS1 and how they are related to chromatin structure and cell division are interesting areas for future experimentation.
Genome-wide studies demonstrated that binding of FLC and SVP is predominately associated with transcriptional repression of target genes [6–8]. One of these targets is the flowering-time gene FT, whose expression in the vascular tissue of leaves is repressed by FLC and SVP [16, 65]. The capacity of FLC-like transcription factors to repress FT transcription has been reported to be associated with their ability to recruit PRC components and maintain H3K27me3 levels at the gene [19]. We also found that transcriptional repression of TFS1 by FLC at the shoot meristem is associated with H3K27me3 accumulation, and that this involves formation of a chromatin loop between the 3’-end and intragenic regions of TFS1 that requires PRC complexes. The JMJ14 H3K4 demethylase also associates with EMF1 and LHP1 [19–21], and we found that in jmj14 mutants H3K27me3 levels as well as binding of both FLC and SVP were significantly reduced at TFS1, although the expression levels of FLC and SVP were not compromised. Collectively, these data suggest a model whereby PRC complexes involving EMF1, LHP1 and JMJ14 are recruited by FLC and SVP to TFS1 to sustain H3K27me3 levels and binding of these transcription factors, thereby stably repressing TFS1 transcription (Fig 8A). In this model, how the PRC1-like complexes reinforce binding of FLC and SVP and whether binding of these transcription factors is a prerequisite for PcG recruitment and PcG-mediated gene silencing remain to be resolved. Another possibility is that a co-factor for FLC binding, perhaps another MADS box transcription factor such as AGL16 [66], is reduced in expression in circumstances in which H3K27me3 levels are reduced. In this case, reduction in H3K27me3 would indirectly lower FLC binding.
In contrast to FLC/SVP, the MADS box factor SOC1 activates TFS1 transcription during floral transition. Furthermore, SOC1 binds directly to TFS1 as defined in genome-wide [12, 28] and targeted ChIP-qPCR experiments performed here. Induction of SOC1:GR was sufficient to activate TFS1 transcription in the presence of SVP, demonstrating that SOC1 activation is epistatic to the repression mediated by SVP and that after SOC1:GR activation SVP binding to TFS1 was strongly reduced. This reduction of SVP at TFS1 could be due to its displacement by SOC1 binding to an adjacent CArG box or to the transcriptional repression of SVP by SOC1 [12, 15, 28]. Furthermore, our ChIP data indicate that the repressive state imposed by FLC/SVP is overcome by SOC1 through recruitment of the H3K27me3 demethylase REF6 and the chromatin remodeler BRM to the TFS1 locus. Similarly, REF6 was recently shown to be recruited to targets by other MADS box transcription factors [67]. Our observations suggest that SOC1 displays characteristics associated with pioneer transcription factors, as it resolves condensed chromatin structures and opens chromatin through the combinatorial activity of REF6 and BRM. Similarly, a recent report in Caenorhabditis elegans demonstrated that the pioneer factor PHA-4 binds to promoters required for foregut development to recruit RNAPII and promote chromatin opening [68]. PHA-4 was proposed to facilitate chromatin opening by depositing RNAPII at target gene promoters. Similarly, the Drosophila maternal pioneer factor ZELDA (Zld) recruits poised RNAPII to Dorsal (Dl) target genes, facilitating chromatin accessibility for Dl which then mediates their zygotic activation [69, 70]. Accordingly, we found that the SOC1-REF6-BRM complex relaxes and opens chromatin at TFS1 to facilitate binding of SPL9 and to activate poised RNAPII, resulting in a reduction in H3K27me3 levels across the TFS1 genomic locus. Many genes directly repressed by FLC or SVP to maintain vegetative development are likely to be subsequently bound and activated by other MADS box transcription factors during reproductive development. Thus the mechanisms defined here by which SOC1 antagonises the repression of TFS1 transcription imposed by FLC/SVP are likely to be more broadly relevant during the transition to flowering.
SOC1 functionally co-operates with SPL15 to form a chromatin loop associated with activation of FUL transcription [22]. Similarly, we showed by co-immunoprecipitation a physical interaction between SPL9 and SOC1 at TFS1. Similarly, we detected looping at the TFS1 locus between the SPL9-binding region close to the TSS and the SOC1 binding region at the 3’-end of TFS1 that might enable a higher turn-over rate of RNAPII to yield higher transcriptional activity. These observations suggest that the formation of an active chromatin loop could enable SOC1 and SPL9 to recruit respectively REF6 and RNAPII to the TSS, and then the active elongating RNAPII could cause the gene body of TFS1 to change its position relative to the stable SOC1-SPL9 complex enabling REF6 to track along the gene with RNAPII progressively removing H3K27me3 (Fig 8B). This model predicts that the SPL9-SOC1 interaction induces dynamic chromatin folding that facilitates movement of the RNAPII along the gene body, rather than that RNAPII separates from the pre-initiation complex and tracks along the TFS1 gene body. It will be interesting to determine in a genome-wide context whether other targets of SPL9 and SOC1 display similar features.
The analysis presented here incorporates TFS1 into a network of interactions among FLC target genes. At the shoot meristem, FLC directly binds to and represses transcription of SOC1 and TFS1 [6, 7, 27, 43]. Furthermore, SOC1 directly activates the transcription of TFS1. Thus repression of TFS1 by FLC involves both direct repression of expression of its positive activator SOC1 as well as direct repression of TFS1, a relationship characterized as a coherent feed forward loop type II [71]. The temporal and spatial patterns of TFS1 expression on the flanks of the inflorescence meristem are overlapping with and partially conferred by SPL9, and may indicate an important role for TFS1 in modulating the expression of genes in cells that will give rise to floral primordia. This suggestion is strengthened by the observation that the Col-FRI flc-3 svp-41 tfs1-1 triple mutant shows a floral morphology defect not shown by any of the single mutants. Previously, the soc1-2 agl24-1 svp-41 combination was also demonstrated to have a synergistic effect on floral development due to redundancy among these transcription factors in the repression of genes involved in floral organ development [43]. Our data suggest that there may also be redundancy among FLC, SVP and TFS1 in the regulation of downstream genes, which could be characterized in a future analysis of TFS1 targets. More generally, our work emphasises that defining the network of genes negatively regulated by FLC/SVP, and understanding how these then interact during the progression to flowering when FLC expression is repressed or lost by mutation, is proving to be a productive approach in defining critical mechanisms controlling floral transition.
All seed stocks are in the Columbia-0 (Col-0) genetic background and were obtained from the Nottingham Arabidopsis Stock Centre (NASC; S1 Table) except for 35S::SOC1:GR soc1-1 (Hyun et al., 2016), which is in a Landsberg erecta (Ler-0) genetic background. Seeds were sown on soil or on full-strength Murashige and Skoog (MS) medium containing 1% sucrose, stratified for 3 days at 4°C, and grown at 22°C under either long-days (16hrs light/8hrs dark; 150μmol.m-1.s-1) or short-days (8hrs light/16hrs dark; 150μmol.m-1.s-1). Plant age was measured when seeds were transferred from stratifying to ambient growth conditions.
Full-length TFS1 genomic region was cloned by PCR with Phusion Enzyme (New England Biolabs) according to the manufacturer’s recommendations and used to generate TFS1::TFS1::9xAla-Venus. To introduce 9xAla-Venus coding sequence, we employed Polymerase Incomplete Primer Extension (PIPE) cloning method [72] and plasmids were then introduced into Agrobacterium to transform Col plants by floral dip [73]. The sequences of the primers used for PIPE cloning are listed in S2 Table.
Total RNA of indicated genotypes at different days after sowing from leaves and apices was isolated with NucleoSpin RNA plant kit (Macherey-Nagel). DNA was removed by an on-column treatment with rDNase and 2 μg RNA was reverse transcribed with an oligo(dT) primer, RNAseOUT Recombinant Ribonuclease Inhibitor (Thermo Fisher Scientific) and SuperScript II Reverse Transcriptase (Thermo Fisher Scientific). The cDNA equivalent of 20ng of total RNA was used in a 12 μL qPCR reaction on a Roche Light Cycler 480 instrument (Roche) with either iQ SYBR Green Supermix (BioRad) or GoTaq qPCR Master Mix (Promega) and quantified using the UBC21 (AT5G25760) as a reference gene to which data was normalized [74]. The mean of three biological replicates with standard deviation is shown and list of primers used for expression analyses can be found in S2 Table.
ChIP was performed as previously described with minor modifications [22]. In brief, above-ground tissue of 15LD-grown plants was collected at ZT8 and fixed in PBS solution containing 1.5% formaldehyde. ChIP-assays in which indirect binding of the protein of interest to chromatin was studied, Di(N-succinimidyl) glutarate (DSG; Synchem) at a final concentration of 1 μM was used to introduce protein-protein crosslinks prior to formaldehyde-assisted protein-chromatin crosslinking. To determine fold enrichment levels, ChIP-DNA was quantified on a Roche Light Cycler 480 instrument (Roche) with iQ SYBR Green Supermix (BioRad) and normalized against ACT8 (AT1G49240). In ChIP assays in which histone modifications were tested, the values of the tested histone marks were normalized against histone H3. The average of three biological replicates is shown and list of primers used for fold enrichment analyses can be found in S2 Table.
3C assay was performed as described previously with minor modifications. A total of 2g of above-ground tissue of 15 day-old LD-grown plants was used for 3C study. Chromatin DNA was digested for 16 hrs at 37°C with 400U Sau3AI (New England Biolabs, S3 Table) while agitating at 900 r.p.m. For intramolecular ligation, digested nuclei were incubated for 5 hrs at 16°C in 500 U T4 DNA Ligase (Promega, S3 Table). In parallel, the cloned TFS1:9xAla-Venus construct was digested and ligated. The 3C DNA ligation products were quantified by RT-qPCR and normalised to the TFS1:9xAla-Venus control using the delta-delta Ct method. The sequences of the primers used in the 3C-assay are listed in S2 Table.
Micrococcal nuclease-assay was performed as described previously with minor modifications [75]. For nuclear extraction, above-ground tissue of 15 day-old LD-grown plants was harvested, ground in liquid nitrogen and resuspended in lysis buffer (LB) [50 mM HEPES pH7.5, 150 mM NaCl, 1 mM EDTA, 1% Triton X-100, 10% glycerol, 5 mM ß-mercaptoethanol and protease inhibitor cocktail (Roche)]. After 1 hr of lysis, the lysis mixture was filtered twice through 2 layers of Miracloth (Calbiochem) and protocol was followed as previously described. For MNase treatment, nuclei were treated with 5U MNase (Thermo Fisher Scientific) for 15 min and digest was stopped by adding 16 μL 250 mM EDTA, and then treated with RNase A and Proteinase K (Sigma Aldrich), each for 1 hr.
Total protein extraction and in vivo co-immunoprecipitation were performed as described previously with minor modifications [22, 76]. For SVP Western-analysis, roughly 50 apices of 15 day-old LD-grown plants were harvested. Protein concentration was determined by Bradford series and a total of 50μg for crude extract and 1mg for immunoprecipitation was used. The amino acid sequences of the epitopes for generating SVP antibody are presented in S4 Table.
In-situ hybridisation was performed according to the method described previously [77]. The sequences of the primers used for the in-situ hybridisation experiments are listed in S2 Table. For confocal microscopy, shoot apices at different developmental stages were collected and fixed with 4% paraformaldehyde (PFA) prepared in phosphate-buffered saline (PBS) at pH7.0. Samples were then vacuum infiltrated for 20 min on ice, transferred to fresh 4% PFA, and stored at 4°C overnight. The fixed samples were washed twice for 1 min in PBS, then cleared with ClearSee [10% (w/v) xylitol, 15% (w/v) sodium deoxycholate and 25% (w/v) urea][78] for 3 to 8 days at room temperature. After clearing, the shoot meristems were imaged by confocal laser scanning microscopy (Zeiss LSM780), as described previously [79].
All image processing and figure construction was performed with Photoshop (www.adobe.com).
Mutant and transgenic lines used in this study, including references for their origin and description in literature, and their respective AGI identifiers are listed in S1 Table.
|
10.1371/journal.ppat.1006736 | Progesterone impairs antigen-non-specific immune protection by CD8 T memory cells via interferon-γ gene hypermethylation | Pregnant women and animals have increased susceptibility to a variety of intracellular pathogens including Listeria monocytogenes (LM), which has been associated with significantly increased level of sex hormones such as progesterone. CD8 T memory(Tm) cell-mediated antigen-non-specific IFN-γ responses are critically required in the host defense against LM. However, whether and how increased progesterone during pregnancy modulates CD8 Tm cell-mediated antigen-non-specific IFN-γ production and immune protection against LM remain poorly understood. Here we show in pregnant women that increased serum progesterone levels are associated with DNA hypermethylation of IFN-γ gene promoter region and decreased IFN-γ production in CD8 Tm cells upon antigen-non-specific stimulation ex vivo. Moreover, IFN-γ gene hypermethylation and significantly reduced IFN-γ production post LM infection in antigen-non-specific CD8 Tm cells are also observed in pregnant mice or progesterone treated non-pregnant female mice, which is a reversible phenotype following demethylation treatment. Importantly, antigen-non-specific CD8 Tm cells from progesterone treated mice have impaired anti-LM protection when adoptive transferred in either pregnant wild type mice or IFN-γ-deficient mice, and demethylation treatment rescues the adoptive protection of such CD8 Tm cells. These data demonstrate that increased progesterone impairs immune protective functions of antigen-non-specific CD8 Tm cells via inducing IFN-γ gene hypermethylation. Our findings thus provide insights into a new mechanism through which increased female sex hormone regulate CD8 Tm cell functions during pregnancy.
| Increased female sex hormones during pregnancy generate a temporary immune suppression status in the pregnant that protect the developing fetus from maternal rejection but renders the pregnant highly susceptible to various pathogens. However, molecular mechanisms underlying such an increased maternal susceptibility to pathogens during pregnancy remain to be further understood. Here we show in pregnant women that increased progesterone levels are associated with IFN-γ gene hypermethylation and reduced IFN-γ production in peripheral CD8 Tm cells. By using murine models of LM infection, for the first time we show a causal relationship between increased level of progesterone, a characteristic female sex hormone of pregnancy, and increased susceptibility to Listeria monocytogenes, an intracellular bacterium that endangers both the pregnant and the fetus. Such an impact on anti-listeria host defense is mediated through progesterone-induced IFN-γ gene hypermethylation in CD8 Tm cells, resulting in impaired IFN-γ production and reduced immune protection by antigen-non-specific CD8 Tm cells. This study provides new insights into molecular mechanisms underlying the increased susceptibility to intracellular pathogens during pregnancy.
| Increased susceptibility to a variety of pathogens during pregnancy has been related to a temporary status of immune suppression induced by increased female sex hormones such as progesterone and estrogen[1–5]. Indeed, previous studies showed that female sex hormones play regulatory roles in various human immune cells ex vivo[6,7]. In animal models, progesterone and estrogen have been shown to exert immune regulatory roles that facilitate maternal-fetal tolerance and protect animals from autoimmune diseases such as experimental autoimmune encephalomyelitis[7–10]. It has also been shown in animal models of infections including influenza virus infections that progesterone reduces anti-virus cellular immune responses while at the same time limits immunopathology[11–14]. Despite of these understandings, roles of female sex hormone in increased susceptibility to infections during pregnancy and the underlying cellular and molecular mechanisms remain to be further defined[3,4,7].
Pregnant women and animals are at higher risks of infection with Listeria monocytogenes(LM), an Gram positive intracellular bacterium[15–19]. In Europe, the incidence rate of listeriosis was estimated to vary between 0.1 and 11.3 per million population, with approximately 20% neonatal infections[20]. In the USA, there were 758 reported cases of listeriosis during 2004–2007, with 16.9% pregnant associated[21]. Although mostly asymptomatic, LM infection during pregnancy can be dangerous not only to the maternal body but also fatal to the developing fetus[17,18]. Innate immunity is critical to optimal control of LM[15,22–25]. Early studies carried out in mouse models showed that IFN-γ produced by NK cells triggered by IL-12 and IL-18 activates bactericidal functions of macrophages against phagocytized LM and is thus critically required in innate bacterial control early after infection[22–24]. Moreover, innate immune responses against LM are important to the establishment of subsequent adaptive immune responses and facilitate bacterial clearance by T cells[15,25]. Conventionally, innate immune responses against LM are restricted to innate immune cells such as NK cells and macrophages[15]. However, recent studies showed that CD8 T memory(Tm) cells provide antigen-non-specific and innate-like immune protection early after LM infection via IFN-γ production in a cognate antigen-independent but innate cytokine dependent manner[26,27]. Notably, such an antigen-non-specific immune protection by CD8 Tm cells is more prominent than that of NK cells due to preferential co-localization of CD8 Tm cells with LM and macrophages[27]. Thus, IFN-γ produced by antigen-non-specific CD8 Tm cells is indispensable to host defense against LM.
The risk of LM infection in pregnant women is highest in the third trimester of pregnancy, when maternal serum progesterone levels are 5–10 folds higher than that before pregnancy[4,28]. Indeed, it has been reported that progesterone modulates functions of various immune cells, regardless of known progesterone receptor expression in these cells[7]. In human CD8 T cells, progesterone has been shown to reduce IFN-γ production upon stimulation ex vivo, although the underlying mechanisms remain unknown[6]. IFN-γ production by CD8 T cells is tightly regulated by a series of consequential epigenetic modulating mechanisms[29–33]. It has been shown in CD8 Tm cells that DNA methylation at the CpG sites of IFN-γ gene promoter is a key mechanism through which IFN-γ production by CD8 Tm cells is regulated[33]. Upon recognition of cognate antigens via TCR, CD8 Tm cells have rapidly IFN-γ gene demethylation, facilitating the transcription of IFN-γ gene[31]. Despite of these intriguing findings, it remains unknown whether and how increased progesterone during pregnancy inhibits IFN-γ production by antigen-non-specific CD8 Tm cells that is crucial to host defense against LM.
Here we show in pregnant women that increased serum progesterone levels are associated with DNA hypermethylation of IFN-γ gene and decreased IFN-γ production in CD8 Tm cells upon antigen-non-specific stimulation. In both pregnant mice and progesterone treated non-pregnant female mice, hypermethylation of IFN-γ gene and significantly reduced IFN-γ production by antigen-non-specific CD8 Tm cells upon LM infection are also observed. And such a reduction in IFN-γ production by CD8 Tm cells is reversed following treatment with demethylating agent. More importantly, antigen-non-specific CD8 Tm cells from progesterone treated mice have impaired protection against LM when adoptive transferred in pregnant mice or IFN-γ-deficient mice, and demethylating agent rescues the impaired adoptive protection of antigen-non-specific CD8 Tm cells induced by progesterone. These data demonstrate that increased levels of progesterone impair immune protection of antigen-non-specific CD8 Tm cells against LM, via inducing DNA hypermethylation of IFN-γ gene. Our findings thus reveal a novel mechanism through which increased female sex hormone regulates CD8 Tm cell functions that result in increased susceptibility to intracellular pathogens during pregnancy.
Pregnant women have significantly reduced host defense against various intracellular pathogens, particularly during the third trimester of pregnancy when serum levels of progesterone reaches the highest levels[18,19,28]. Immune responses mediated by CD8 Tm cells are critically required in host defense against intracellular pathogens[34]. IFN-γ, a key molecule in CD8 T cell functions that is subject to epigenetic regulation pathways including DNA methylation, is reduced by progesterone, a characteristic female sex hormone that is significantly increased during pregnancy[6]. This prompts us to ask whether serum progesterone levels are related to the methylation level at IFN-γ gene promoter region that controls IFN-γ production by CD8 Tm cells from pregnant women. To test this, we purified peripheral blood CD8 Tm cells from 10 women at before, weeks14 and 28 of pregnancy, and approximately 1 year after delivery. Serum progesterone levels were also determined at all the four time points. Methylation levels at six known CpG sites in the promoter region of IFN-γ gene was determined by using bisulfate sequencing. Some CD8 Tm cells were stimulated ex vivo with PHA, followed by intracellular staining of IFN-γ. Before pregnancy, median percentage of IFN-γ gene methylation at the six CpG sites was less than 25% (Fig 1A). At weeks 14 and 28 of pregnancy, the percentages of IFN-γ gene methylation were around 40% and 50%, respectively, with that of week 28 significantly higher than before pregnancy (Fig 1A). One year after delivery, the percentage of IFN-γ gene methylation was reduced to a comparable level to that before pregnancy, being significantly lower than that at week 28 (Fig 1A). Correlation analysis data showed that increased serum progesterone level was correlated to hypermethylation of IFN-γ gene promoter CpG sites (Fig 1B). Consistent to the IFN-γ gene methylation levels, relative expression of IFN-γ mRNA in CD8 Tm cells upon ex vivo stimulation was reduced during pregnancy but not at one year after delivery (S1 Fig). And frequency of IFN-γ-producing CD8 Tm cells upon ex vivo stimulation was significantly reduced at weeks 14 and 28 of pregnancy as compared to that before pregnancy (Fig 1C). One year after delivery, frequency of IFN-γ-producing CD8 Tm cells recovered to a comparable level with that before pregnancy (Fig 1C). Not unexpectedly, correlation analysis data showed that frequency of IFN-γ-producing CD8 Tm cells was negatively related to IFN-γ gene methylation levels (Fig 1D). Our data thus suggest that increased serum progesterone levels during pregnancy are related to IFN-γ gene hypermethylation and reduced IFN-γ production in CD8 Tm cells.
To address the causal relationship between IFN-γ gene hypermethylation and reduced IFN-γ production by CD8 Tm cells during pregnancy, we treated CD8 Tm cells from pregnant women at 28 week of pregnancy ex vivo with demethylating agent decitabine, followed by stimulation of CD8 Tm cells. Demethylation treatment significantly reduced IFN-γ gene methylation level in CD8 Tm cells from pregnant women at 28 weeks of pregnancy (Fig 2A and 2B). Moreover, pre-treatment with demethylating agent significantly increased the frequency of IFN-γ-producing CD8 Tm cells from pregnant women following both TCR-independent stimulation by PHA and TCR-dependent stimulation by CMVpp65 peptide ex vivo (Fig 2C and 2D). Thus our findings suggest that reduced IFN-γ production by CD8 Tm cells during pregnancy is dependent on IFN-γ gene hypermethylation, which is related to increased progesterone level (Fig 1B).
In pregnant women, there is a large number of circulating CD8 Tm cells that are non-specific for LM antigens[35,36]. And it is speculative, based on murine studies, that LM antigen-non-specific CD8 Tm cells provide immune protection against LM via IFN-γ production[26,27]. To better mimic this immunological scenario in human, we next use a murine model of LM infection during pregnancy to further determine the functional significance of IFN-γ gene hypermethylation and reduced IFN-γ production by antigen-non-specific CD8 Tm cells during pregnancy. We immunized naïve female mice with rAdHuOVA to generate LM non-specific CD8 Tm cells[37]. These immunized female mice were then mated to males to generate pregnant mice with LM non-specific CD8 Tm cells (S2A and S2B Fig).
Compared to non-pregnant female mice, pregnant mice had over 3 logs higher bacterial CFU number at 72h post LM infection (Fig 3A). At 24h post infection, both frequency and absolute number of IFN-γ-producing OVA-specific CD8 Tm cells in peripheral blood, spleen and mesenteric lymph nodes (MLN) were significantly lower in pregnant mice compared to non-pregnant females (Fig 3B, 3C and 3D). As these CD8 Tm cells are not LM antigen-specific, our data suggest that IFN-γ production by antigen-non-specific CD8 Tm cells early after LM infection is impaired during pregnancy. We further analyzed IFN-γ gene methylation levels at various time points before and after infection. Although demethylation of IFN-γ gene occurred in both pregnant and non-pregnant mice post infection, the IFN-γ gene methylation level was significantly higher in pregnant mice at all the time points (Fig 3E and S3A Fig). Further methylation analysis at distal regulatory elements of IFN-γ gene before LM infection further support a hypermethylation status of IFN-γ gene in pregnant mice (S3B Fig). Similar to that observed in pregnant women, our data demonstrate IFN-γ gene hypermethylation and reduced IFN-γ production in antigen-non-specific CD8 Tm cells, which is associated with increased susceptibility to LM in pregnant mice. We also determined NKG2D expression on OVA-specific CD8 Tm cells in various organs in pregnant mice, as NKG2D has been associated with bystander activation of antigen-non-specific CD8 Tm cells. Our data showed that NKG2D expression was not altered in pregnant mice compared to non-pregnant female mice (S4 Fig).
To further determine the impact of increased progesterone levels on IFN-γ gene methylation and IFN-γ production by antigen-non-specific CD8 Tm cells, rAdHuOVA immunized female mice were injected with progesterone for 14 consecutive days. Serum progesterone concentration following such an exogenous progesterone supplementation strategy reflected that in pregnant mice (S5 Fig)[38,39]. Some progesterone-treated mice were administered with demethylating agent decitabine. Progesterone treated mice had nearly 3 logs higher bacteria CFU numbers in the spleen and the liver at 72h post LM infection compared to control mice (Fig 4A). Demethylation treatment significantly reduced bacteria burden in both the spleen and liver in progesterone treated mice with around 1.5 logs reduction in bacteria CFU numbers (Fig 4A). At 24h post infection, both frequency and absolute number of IFN-γ-producing OVA-specific CD8 Tm cells in peripheral blood, spleen and MLN were reduced in progesterone treated mice (Fig 4B and 4C). Consistent with the reduced IFN-γ production by OVA-specific CD8 Tm cells, both frequency and absolute number of IFN-γ-producing H2KbOVA tetramer- CD8 T cells were reduced in progesterone-treated mice (S6A and S6B Fig). These data suggest that progesterone treatment reduces IFN-γ production in both OVA-specific CD8 Tm cells and other CD8 Tm cells such as endogenous antigen-inexperienced CD8 Tm cells[40]. In contrast to CD8 Tm cells, there was a moderate but not significant reduction of IFN-γ-producing NK cells in progesterone-treated mice following LM infection (S6C and S6D Fig), suggesting that the impact of progesterone on IFN-γ production is a cell type-specific phenotype. Notably, demethylation treatment significantly increased the frequency and absolute number of IFN-γ-producing OVA-specific CD8 Tm cells in peripheral blood, the spleen and MLN (Fig 4B and 4C).
Similar to that in pregnant mice, CD8 Tm cells in progesterone treated mice had higher IFN-γ gene methylation levels at various time points before and after infection (Fig 4D and S7 Fig). Such an effect of progesterone on IFN-γ gene methylation was reversed by demethylation treatment (Fig 4D and S7 Fig). We also determined IFN-γ production by CD8 Tm cells after ex vivo stimulation with innate cytokines IL-12 and IL-18 for 24h without cognate antigen stimulation. As shown in Fig 4E and 4F, frequency and absolute number of IFN-γ-producing OVA-specific CD8 Tm cells in peripheral blood, spleen and MLN were reduced in progesterone treated mice. Demethylation treatment rescued the reduced frequency and absolute number of IFN-γ-producing antigen-non-specific CD8 Tm cells induced by progesterone (Fig 4E and 4F). These findings demonstrate that progesterone reduces IFN-γ production by antigen-non-specific CD8 Tm cells and impairs host defense against LM via IFN-γ gene hypermethylation.
We next determined whether increased susceptibility of pregnant mice to LM is dependent on progesterone induced impairment of IFN-γ production in antigen-non-specific CD8 Tm cells. To do this, we adoptive transferred OVA-specific CD8 Tm cells or IVA NP366-374-specific CD8 Tm cells. As shown in Fig 5A, 5B, 5C and 5D, adoptive transfer of either OVA-specific CD8 Tm cells or IVA NP366-374-specific CD8 Tm cells significantly reduced LM bacterial burden in LM-naïve host pregnant mice. We also adoptive transferred OVA-specific CD8 Tm cells from mice treated with progesterone alone or in combination with decitabine, into pregnant mice. At 24 hours post LM infection, adoptively transferred OVA-specific CD8 Tm cells from progesterone-treated donor mice had significantly reduced IFN-γ-producing capacity, as compared to those without progesterone treatment or those treated with progesterone and demethylating agent decitabine (S8A and S8B Fig). More importantly, adoptive transfer of OVA-specific CD8 Tm cells from progesterone treated mice failed to reduce bacterial burdens in the spleen and liver (Fig 5A and 5B). Whereas CD8 Tm cells from either un-treated or progesterone- and decitabine-treated mice significantly reduced bacterial burdens (Fig 5A and 5B). These data thus demonstrate that progesterone impairs anti-LM protection by antigen-non-specific CD8 Tm cells via hypermethylation-dependent mechanisms.
As one of the most severe clinical outcomes from gestational LM infection is fetal loss, we also determined pregnancy outcomes in LM-infected pregnant mice following antigen-non-specific CD8 Tm cell transfer. OVA-specific CD8 Tm cells from donor mice without, but not with progesterone treatment, moderately increased number of viable fetus per mouse and moderately decreased abortion rate in LM-infected pregnant mice (S9A and S9B Fig), although the differences were not statistically significant.
To further show that impaired protective functions of antigen-non-specific CD8 Tm cells are dependent on reduced IFN-γ production due to DNA hypermethylation, we adoptive transferred OVA-specific CD8 Tm cells from wild type mice treated with progesterone alone or in combination with decitabine, into naïve IFN-γ-deficient (IFNG-/-) mice, followed by LM infection. IFNG-/- mice without CD8 Tm cell transfer were highly susceptible to LM as demonstrated by over 108 CFUs in both the spleen and the liver at 72h post infection (Fig 6A and 6B). Adoptive transfer of CD8 Tm cells significantly reduced spleen and liver bacterial CFU numbers by around 2 logs (Fig 6A and 6B). Adoptive transferred CD8 Tm cells from progesterone treated mice, however, reduced spleen and liver bacterial CFU numbers by around only 1 log (Fig 6A and 6B). CD8 Tm cells from progesterone and decitabine treated mice were nearly as protective as those from progesterone untreated mice (Fig 6A and 6B). Moreover, the protection of CD8 Tm cells from progesterone and decitabine treated mice was abolished by in vivo administration of IFN-γ neutralizing antibody (Fig 6A and 6B). Consistent to the phenotype observed in IFN-γ neutralized mice, adoptive transferred IFN-γ-deficient CD8 Tm cells generated in IFNG-/- mice (IFN-γ-/- Tm) failed to reduce LM bacterial burden (Fig 6C and 6D), further suggesting that such a protection is dependent on IFN-γ produced by antigen-non-specific CD8 Tm cells. These data demonstrate that progesterone impairs IFN-γ-mediated immune protective functions of antigen-non-specific CD8 Tm cells via DNA hypermethylation.
In this study we show that in pregnant women, increased serum progesterone levels are associated with decreased IFN-γ production of CD8 Tm cells which is dependent on IFN-γ gene hypermethylation. Pregnant mice are highly susceptible to LM infection. And there are IFN-γ gene hypermethylation and reduced IFN-γ production in antigen-non-specific CD8 Tm cells in both pregnant mice and progesterone-treated non-pregnant female mice early after LM infection. Moreover, LM antigen-non-specific CD8 Tm cells from progesterone-treated mice have reduced protection against LM after adoptive transfer to pregnant mice or IFNG-/- mice, which is dependent on progesterone-induced IFN-γ gene hypermethylation and reduced IFN-γ production early after LM infection.
Host defense against LM infection depends primarily on cellular immune responses[15,41]. The requirement of T cells, in particular CD8 T cells, in adaptive immune protection against intracellular pathogens such as LM has been well established in previous studies[41]. In mouse models, it has been shown that primary LM infection induces potent antigen-specific CD8 T cell immune responses that subsequently generates long-lasting antigen-specific CD8 Tm cells with augmented protective functions during secondary LM infection[42,43]. However, key molecules that are required for antigen-specific CD8 T cell-mediated immune protection against LM vary in different experimental settings[44,45]. Naïve IFNG-/- mice are highly susceptible to LM infection, suggesting critical dependence of IFN-γ in bacterial clearance [45]. However, IFNG-/- LM antigen-specific CD8 Tm cells adoptive transferred into naïve wild type mice provide equal protection as IFNG+/- CD8 Tm cells, suggesting an IFN-γ-independent mechanism through which antigen-specific CD8 Tm cells exert bacterial clearance[44]. More recent studies showed that antigen-non-specific CD8 Tm cells provide immune protection against LM via an IFN-γ dependent but cognate antigen independent mechanism[26,27]. Furthermore, IFN-γ-mediated protection by antigen-non-specific CD8 Tm cells is superior to that by NK cells due to the preferential co-localization of CD8 Tm cells with LM and macrophages in target organs post infection[27]. Importantly, such an antigen-non-specific CD8 Tm cell-mediated IFN-γ-production better reflect immune responses to LM in adult human with much more LM antigen-non-specific than antigen-specific CD8 Tm cells due to a highly diversified TCR repertoire[35]. Thus, IFN-γ-production by antigen-non-specific CD8 Tm cells plays indispensable roles in optimized host defense against LM infection[15,26]. Here in this study we generate a LM infection model in pregnant mice that have pre-established LM antigen-non-specific CD8 Tm cells, in order to determine the impact of increased progesterone on antigen-non-specific immune protective functions of CD8 Tm cells. We believe our current model reflect a critical aspect of the immune scenario of LM infection in pregnant women, which is not reflected in previous LM infection models in either antigen-inexperienced pregnant mice or LM antigen-specific CD8 Tm models in which LM antigen-specific CD8 Tm cells dominant the CD8 Tm cell repertoire[19,44].
Pregnant women and animals have significantly increased susceptibility to a variety of intracellular pathogens including LM, to which CD8 T cell-mediated immune responses are critically required[5,15,18,46]. Such an increased susceptibility during pregnancy has been associated with significantly increased female sex hormones such as progesterone and estrogen[1,7,18,46]. Indeed, most cases of LM infection during pregnancy were reported during the third trimester, when serum progesterone and estrogen levels reach the highest levels[18,28]. Progesterone plays sophisticated roles in immune cell functions presumably via both directly binding to cognate receptors or potentially undefined receptors, and indirectly through intermediate cells or molecules[7]. Functions of essentially all major immune cells including CD8 T cells are subject to the modulation of progesterone[7]. It has been reported that progesterone reduces IFN-γ-production by human CD8 T cells that express progesterone receptor[6], although the intracellular and molecular mechanisms remain unknown. By using human CD8 Tm cells from pregnant women, we show here that reduced IFN-γ-production by human CD8 Tm cells during pregnancy is dependent on IFN-γ gene hypermethylation. By using pregnant mice and female mice administered with exogenous progesterone at a dose level that reflect progesterone level in pregnant mice[38,39], we identify the causal relationship between increased progesterone and IFN-γ gene hypermethylation which is required to the reduced IFN-γ-production by antigen-non-specific CD8 Tm cells post LM infection. More importantly, the functional significance of progesterone-induced IFN-γ reduction in antigen-non-specific CD8 Tm cells is established, as demonstrated by the impaired anti-LM protection of antigen-non-specific CD8 Tm cells from progesterone treated mice in a hypermethylation-dependent manner. In contrast to the significant differences in LM bacterial burdens, we observed moderate but not statistically significant improvement in the number of viable fetuses and abortion rate in pregnant mice receiving antigen-non-specific CD8 Tm cells compared to those receiving no cell transfer or receiving CD8 Tm cells from progesterone treated mice. This might be explained by that IFN-γ production by antigen-non-specific CD8 Tm cells following LM infection disrupts maternal-fetal tolerance mechanisms, partially compensating the fetus-protective effects of reduced maternal and possibly placental bacterial burden[47,48]. Further mechanistic studies are critically required to extend our knowledge on whether and how maternal immune responses against LM infection during pregnancy may independently impact pregnancy outcomes.
It has been suggested that progesterone induces regulatory T cell expansion during pregnancy, which facilitates maternal-fetal tolerance but impairs anti-infectious immunity[47]. Earlier studies also suggested that placental trophoblasts as a protected niche to harbor bacteria that then re-seed maternal organs, causing persistent LM infection until expulsion of the infected placental tissues[48]. Our data thus provide a new mechanistic explanation to the T cell immune suppression and increased susceptibility to LM during pregnancy, which may work simultaneously and/or sequentially with the previously proposed mechanisms[47,48]. As CD8 Tm cell-derived IFN-γ is also required in innate-like protection against local virus infections[49], it remains possible that such an epigenetic modification of CD8 Tm cells by progesterone also contributes to the increased susceptibility to viral infections during pregnancy.
Transcription of IFN-γ gene is regulated by a variety of sequential epigenetic mechanisms including DNA methylation, transcription factors and chromatin modulation[29–32]. It has been shown in human T cells that hypermethylation of IFN-γ gene promoter CpG sites are related to immune suppression[50]. In mice, IFN-γ gene promoter CpG sites are nearly completely demethylated in effector CD8 T cells that are readily producing IFN-γ[31]. Although CD8 T m cells and naïve CD8 T cells have comparable overall IFN-γ gene methylation levels, quick demethylation of IFN-γ gene was observed in CD8 Tm cells but not naïve CD8 T cells early upon cognate antigen stimulation[31]. These findings strongly support an idea that DNA methylation is a gate-keeping molecular mechanism in controlling IFN-γ gene transcription in CD8 T cells including CD8 Tm cells. Here we show that hypermethylation of IFN-γ gene is associated with increased serum progesterone levels in pregnant women. Furthermore, hypermethylation of IFN-γ gene promoter region reduces IFN-γ production by CD8 Tm cells in pregnant women. Such a phenotype of hypermethylation-dependent IFN-γ reduction is also observed in CD8 Tm cells in pregnant mice. The hypermethylation of representative regulatory elements within the IFN-γ gene locus in CD8 Tm cells in pregnant mice further support our conclusions that progesterone induces hypermethylation-dependent IFN-γ reduction in CD8 Tm cells[51]. We also show that exogenous progesterone supplementation in non-pregnant female mice induces similar IFN-γ gene hypermethylation and IFN-γ reduction in CD8 Tm cells as observed in pregnant mice. These data demonstrate the indispensable roles of progesterone in CD8 Tm cell functions during pregnancy, although this does not exclude the potential impacts of other pregnancy-associated hormones on CD8 Tm cells. In a previous study, a rapid and nearly complete demethylation of IFN-γ gene was observed in CD8 Tm cells upon cognate antigen stimulation[31]. In our current model however, only partial demethylation was observed in antigen-non-specific CD8 Tm cells within 72h post LM infection. Such a partial demethylation is consistent to a relatively low frequency of IFN-γ-producing antigen-non-specific CD8 Tm cells upon cytokine stimulation ex vivo or LM infection in vivo in our current model. These findings indicate that DNA methylation-based regulation of IFN-γ gene transcription may have different upper stream signaling pathways when CD8 Tm cells are activated by TCR-dependent versus TCR-independent stimulants[52].
In conclusion, our data demonstrate that increased progesterone during pregnancy induces IFN-γ gene hypermethylation in CD8 Tm cells, resulting in reduced IFN-γ production and impaired anti-LM immune protective functions of antigen-non-specific CD8 Tm cells. These findings thus provide new mechanistic insights into the increased susceptibility to intracellular pathogens during pregnancy, as well as an unappreciated immune regulatory role of progesterone in CD8 Tm cells.
All experimental animal manipulations were conducted in accordance with the National Institute of Health Guide for the Care and Use of Laboratory Animals along with approval from the Scientific Investigation Board of Chinese PLA General Hospital. The project license number is NSFC81441006.
Written informed consent was obtained from all subjects for the use of personal medical data and peripheral blood cells in the current study. This study was conducted according to the Declaration of Helsinki and all procedures involving human subjects were approved by the Ethics Committee of the General Hospital of Chinese People's Armed Police Forces.
Peripheral blood was collected at before, weeks 14 and 28 of pregnancy, and around 1 year after delivery from subjects who had prenatal exams in the Obstetric Out-Patient Clinic of the General Hospital of Chinese People's Armed Police Forces. At each time point, routine clinical examination on serum concentration of progesterone was performed in the Department of Clinical Laboratory Examinations of the General Hospital of Chinese People's Armed Police Forces. Peripheral blood mononuclear cells (PBMCs) were obtained by Ficoll-Paque gradient centrifuge[53].
CD8 Tm cells were negatively selected from non-adherent PBMCs by using human memory CD8 T cell enrichment kit (StemCell; Vancouver, BC, Canada). Purity of enriched human CD8 T memory cells was typically 90% to 95%, as determined by flow cytometry analysis on a CD3+CD8+CD45RO+ phenotype.
T2 (HLA-A0201+ TAP-deficient lymphoblastoid cell line) cells were kindly provided by Professor Xuetao Cao from Chinese Academy of Medical Sciences, and were cultured in RPMI 1640 culture media(Hyclone Laboratories; South Logan, UT, USA) supplemented with 10% fetal bovine serum (FBS; Hyclone Laboratories; South Logan, UT, USA), penicillin and streptomycin, at 37°C in a CO2 incubator. Human IFN-γ ELISpot kit was purchased from DAKEWE Biotech (Shenzhen, China). For T-cell receptor (TCR)-independent IFN-γ production, enriched CD8 Tm cells were cultured at 5×103 cells/well in Serum-Free Media for ELISpot (DAKEWE; Shenzhen, China) supplemented with phytohaemagglutinin (PHA, 2.5μg/ml). For TCR-dependent IFN-γ production, enriched PBMC CD8 Tm cells from HLA-A0201+ donors were cultured at 2×104 cells/well with T2 cells (CD8Tm:T2 = 10:1) pulsed with cytomegalovirus (CMV) pp65 peptide 495-503(NLVPMVATV, CMVpp65) at 10mg/ml. For demethylation treatment, enriched CD8 Tm cells were pretreated with decitabine (0.5μM; Xian-Janssen Pharmaceuticals Ltd, Xi’an, China) for 24 hours and the same concentration of decitabine was supplemented to the ex vivo stimulation system[53]. Cells were cultured in triplicate ELISpot wells for 16 hours according to manufacturer’s instructions and plates were read by DAKEWE Biotech (Shenzhen, China) on a AID EliSpot Read Classic (AID GmbH, Strassberg, Germany).
Wild-type (WT) C57BL/6J (female and male, 6–8 weeks of age) were purchased from the Joint Ventures Sipper BK Experimental Animal Co. Ltd. (Shanghai, China). Breeders of Interferon-γ deficient (IFNG-/-) mice (B6.129S7-Ifngtm1Ts/J) were purchased from the Jackson Laboratory (Bar Harbor, ME, USA). Female IFNG-/- mice at 6 to 12 weeks of age were used in the experiments. Mice were housed in specific pathogen-free conditions in central animal facility of Chinese PLA General Hospital.
Recombinant replication-deficient human type 5 adenovirus expressing ovalbumin (rAdHuOVA) was constructed and kindly provided by Dr. Xiaohua Tan from Beijing Military General Hospital. Female WT C57BL/6J mice or IFNG-/- mice were immunized intramuscularly with rAdHuOVA at 5×107 PFUs/animal. In the experiments where indicated, mice were immunized i.p. with 108.5 egg infective dose units of H1N1 influenza A virus (A/PR8/34 strain; originally from ATCC and was a kind gift from Dr. Xiaohua Tan from Beijing Military General Hospital) to generate memory CD8 T cells[54–56]. At 40 days post immunization, OVA or influenza virus A (IVA) antigen-specific CD8 Tm cells were generated. And at 40 days post immunization, female mice with OVA antigen-specific CD8 Tm cells were mated with male C57BL/6J mice to induce pregnancy or leave non-pregnant. Female mice at 14 to 16 days of pregnancy were infected with LM. To determine the impact of progesterone on CD8 Tm cells, immunized non-pregnant female mice (on day 40 post immunization) were injected subcutaneously with progesterone (0.75mg/animal/day on days 1–7, and 1.5mg/animal/day on days 8–14; Sigma-Aldrich, St. Louis, MO, USA) suspended in 0.1ml olive oil or vehicle (0.1ml olive oil) for 14 consecutive days[8,57]. At various time points following progesterone administration, serum concentration of progesterone (P4) analysis was performed in the Department of Clinical Laboratory Examinations of the General Hospital of Chinese People's Armed Police Forces. In some progesterone treated mice, decitabine (1mg/kg/day) diluted in PBS was injected intraperitoneally on days 10–14 of progesterone administration[58].
A virulent strain of Listeria monocytogenes (LM; strain 10403S) was grown in brain-heart infusion broth (BHI; BD Biosciences, San Jose, CA, USA). At mid-log growth phase, colony forming units (CFUs) were counted following overnight incubation on BHI agar. For bacterial infection, mice were intravenously infected with LM diluted in PBS at 2.5×103 CFUs/animal intravenously. At 72 h after LM challenge, spleen and liver were harvested and dissociated on metal screens in 10 ml of PBS containing 0.05% Triton-X100. Serial dilutions were performed in the same buffer and plated on BHI agar plates. Colonies on plates were counted after overnight culture and CFUs per organ were calculated. To determine pregnancy outcomes following LM infection, pregnant mice at day 13–15 of pregnancy, with or without OVA antigen-specific CD8 Tm cell transfer, were infected with LM at 2.5×103 CFUs/animal intravenously. Uteri were examined 4 days post infection for post-implantation scars that indicate aborted fetuses, as well as for viable fetuses. Abortion rate = number of aborted fetuses/(number of aborted + viable fetuses)×100% [59].
Unless otherwise specified, all reagents for flow cytometry were purchased from BD Biosciences (San Jose, CA, USA). For intracellular staining of human CD8Tm cells, enriched PBMC CD8 Tm cells were stimulated ex vivo in triplicate wells with PMA(100ng/ml) and Ionomycin(1μg/ml) for 5 hours in the presence of GolgiPlug, followed by staining with human CD45-AF488(Biolegend, San Diego, CA USA), anti-human CD3-PerCP-Cy5.5, anti-human CD8-PE-Cy7, and anti-human CD45RO-APC. Stained cells were then fixed, permeabilized and stained with anti-human IFN-γ-PE. For ex vivo mouse T cell stimulation, single cell suspension from peripheral blood, the spleen and MLN of rAdHuOVA immunized mice were cultured in the presence of recombinant murine IL-12 (5 ng/ml) and IL-18 (10 ng/ml; both cytokines were from Peprotech, Rocky Hill, NJ, USA) for 24 hours. Five hours before harvest, GolgiPlug was supplemented to culture media. For intracellular flow cytometry staining of mouse cells, single cell suspension of peripheral blood, the spleen and MLN from LM infected mice were cultured ex vivo for 5 hours in the presence of GolgiPlug and stained with anti-mouse CD3-V450, anti-mouse CD8-APC-Cy7, anti-mouse CD44-APC, anti-mouse NK1.1-AF700 and H2Kb/OVA.SIINFEKL tetramer-PE (H2KbOVA257-264; NIH Tetramer Core Facility, Atlanta, GA, USA). Stained cells were then fixed, permeabilized and stained with anti-mouse IFN-γ-FITC. In selected experiments, cells were stained with anti-mouse CD3-V450, anti-mouse CD8-APC-Cy7, H2Kb/OVA.SIINFEKL tetramer-PE, and anti-mouse NKG2D-APC. FACS stained cells were acquired on a LSR II cytometer (BD Biosciences, San Jose, CA, USA). FACS data were analyzed by using FlowJo software version 10 (TreeStar, Ashland, OR, USA).
For adoptive transfer of CD8 Tm cells, cells from the spleen and MLN of rAdHuOVA or influenza virus A/PR8/34 immunized mice were enriched by using mouse CD8 negative selection kit (StemCell; Vancouver, BC, Canada). Enriched CD8+ cells were stained with anti-CD3, anti-CD8, anti-CD44 antibodies and H2KbOVA257-264 tetramer for OVA antigen specific CD8 Tm cells or H2DbNP366-374 –PE (NIH Tetramer Core Facility, Atlanta, GA, USA) for IVA antigen-specific CD8 Tm cells [54], followed by flow sorting on a Moflo XDP cell sorter (Beckman-Coulter; Brea, CA, USA) based on a CD3+CD8+CD44+tetramer+ phenotype. Purity of sorted CD8 Tm cells was >95%. Viability of purified CD8 Tm cells were >97% in all groups as determined by trypan blue exclusion. OVA or IVA antigen-specific CD8 Tm cells were adoptively transferred into recipient mice intravenously at 2×106 cells/animal. Recipient mice were infected with LM (2.5×103 CFUs/animal) 4 hour after CD8 Tm cell adoptive transfer. At 24 hours post infection, the presence of adoptively transferred OVA antigen-specific CD8 Tm cells and their IFN-γ-producing capacity were determined by flow cytometry analysis. To block IFN-γ in vivo in IFNG-/- mice adoptively transferred with CD8 Tm cells, mice were injected intraperitoneally with 200 μg/animal of the anti-IFN-γ antibody (BioxCell, West Lebanon, NH, USA) on day -1 of infection. Dose of anti-IFN-γ antibody was repeated at 100 μg/animal/day on days 0, 1, and 2 of infection.
Total RNA was extracted from enriched human PBMC CD8 Tm cells(hCD3+hCD8+hCD45RO+) using miRNeasy Mini Kit (Qiagen, Germantown, MD, USA) according to manufacturer’s instruction. RT was performed using Reverse Transcription System (Promega; Madison, WI, USA) on 1 μg of total RNA[60]. Expression of IFN-γ and GAPDH was quantified by SYBRgreen real-time quantitative PCR analysis on an Mx3000p light cycler (Stratagene; La Jolla, CA, USA), and data were analyzed using Mx3000p software (Stratagene; La Jolla, CA, USA). Primers for human IFN-γ (forward and reverse): 5’-GCAGGTCATTCAGATGTAGCGG-3’ and 5’-TGTCTTCCTTGATGGTCTCCACAC-3’. Primers for human GAPDH (forward and reverse): 5’-GAGTCAACGGATTTGGTCGT-3’ and 5’-TTGATTTTGGAGGGATCTCG-3’. IFN-γ mRNA expression was expressed as 2-ΔCT relative to GAPDH.
Genomic DNA was prepared from purified human PBMC CD8 Tm cells(hCD3+hCD8+hCD45RO+), or murine splenic CD8 Tm cells (CD3+CD8+CD44+H2Kb-OVA257-264 tetramer+) at 40 days post immunization, by using the Wizard Genomic DNA Purification Kit (Promega; Madison, WI, USA). Bisulfite-treatment of genomic DNA was performed as previously described[60], followed by PCR amplification using the Epitectbisulfit kit (Qiagen; Germantown, MD, USA). For methylation analysis on human IFN-γ gene promoter CpG sites, the following primer pair was used: forward, 5’-TGTGAATGAAGAGTTAATATTTTATTA-3’; reverse, 5’-TTGGTAGTAATAGTTAAGAGAATTTA-3’[50]. For methylation analysis on mouse IFN-γ gene promoter CpG sites, bisulfite-treated DNA was amplified in semi-nested PCR using primers: 5’-GGTGTGAAGTAAAAGTGTTTTTAGAGAATTTTAT-3’ and 5’-CAATAACAACCAAAAACAACCATAAAAAAAAACT-3’, then 5’-GGTGTGAAGTAAAAGTGTTTTTAGAGAATTTTAT-3’ and 5’-CCATAAAAAAAAACTACAAAACCAAAATACAATA-3’[33]. For methylation analysis on regulatory elements of mouse IFN-γ gene locus, the following primers were used: Locus -54: Primer pair 1, 5’- GTGGTTAAGATAGGTTTGTTAGTGGTTTGTT-3’, 5’- ATTACACATCTACATAATCTAAAAACTTCCTA-3’; Primer pair 2, 5’- GGTTTGTGGATATTAGTGATGTTGAG-3’, 5’- AAACACTTCCTTCAACTTCCCCTACTATA-3’. Locus -6: 5’- TTTAATTTATGGGATAAATGAGTTA-3’, 5’- AAATACTATCACCCCAATAACACATC-3’. Locus +18: 5’- TAATGTGAGTTGGAATATTAAGAATTT-3’, 5’- TCTAAATAAACAAATCACCAAATCTCA-3’. Locus +20: 5’- GATAAGTAGTTTAAAGGTTATATGT-3’, 5’- CTAAATCCCTTACTAACCTACATCC-3’. Locus +55: Primer pair 1, 5’- GAAGGTTTTATGTTTAGGTTAGAAATATTTT-3’, 5’- TACCTATCTCTTACCCAAAATATTATCTATA-3’; Primer pair 2, 5’- GATGTTTGGAGAGAGATAAAATATAGGTTAGTT-3’, 5’- TTTCCTACAAATAATTCTCTAATTA-3’[51]. The PCR products were gel purified and cloned into the pGEM-T vector (Promega; Madison, WI, USA). The inserted PCR fragments of individual clones were sequenced using an ABI PRISMDNA sequencer (Applied Biosystems; Foster City, CA, USA). For all samples, 10 reads or 6 reads per CpG site were used to determine the average percentage of methylated CpG.
Two-tailed unpaired Student’s t-test was used for statistical comparison between two groups in mouse experiments. Two-tailed paired Student’s t-test was used for statistical comparison between control and decitabine treatment groups in human cell experiments. One-way ANOVA and Tukey’s multiple comparisons test was used to compare between multiple groups. Pearson correlation analysis was used to determine the potential correlation between two parameters. All statistical analysis was performed by using the GraphPad Prism software (version 6.01; GraphPad Software, La Jolla, CA, USA). Values of P < 0.05 were considered statistically significant.
|
10.1371/journal.ppat.1006733 | Inhibition of group-I metabotropic glutamate receptors protects against prion toxicity | Prion infections cause inexorable, progressive neurological dysfunction and neurodegeneration. Expression of the cellular prion protein PrPC is required for toxicity, suggesting the existence of deleterious PrPC-dependent signaling cascades. Because group-I metabotropic glutamate receptors (mGluR1 and mGluR5) can form complexes with the cellular prion protein (PrPC), we investigated the impact of mGluR1 and mGluR5 inhibition on prion toxicity ex vivo and in vivo. We found that pharmacological inhibition of mGluR1 and mGluR5 antagonized dose-dependently the neurotoxicity triggered by prion infection and by prion-mimetic anti-PrPC antibodies in organotypic brain slices. Prion-mimetic antibodies increased mGluR5 clustering around dendritic spines, mimicking the toxicity of Aβ oligomers. Oral treatment with the mGluR5 inhibitor, MPEP, delayed the onset of motor deficits and moderately prolonged survival of prion-infected mice. Although group-I mGluR inhibition was not curative, these results suggest that it may alleviate the neurological dysfunctions induced by prion diseases.
| Prion diseases are a result of ordered accumulation of the misfolded conformer of cellular prion protein (PrPC), a GPI anchored protein expressed on the cell surface. Similar pathogenetic principles operate in several other neurodegenerative diseases. Currently no disease-modifying therapies exist and the situation is compounded by a dearth of validated therapeutic targets. In our present study, we have discovered that genetic ablation, or pharmacological inhibition, of group-I (i.e. activating) metabotropic glutamate receptors is beneficial against prion neurotoxicity in vitro and in vivo. Mice treated with these inhibitors exhibited impressive suppression of neurological signs and a delayed onset of the symptoms. These results further suggest that activation of these metabotropic glutamate receptors is a downstream event of prion replication and targeting these receptors could be a therapeutic option to alleviate the neurological symptoms, thereby ameliorating the quality of life in patients having prion infection.
| The decisive event in the pathogenesis of prion diseases is the conversion of the normal cellular prion protein (PrPC) into an aggregated conformational variant called PrPSc [1]. Expression of PrPC at the cell surface is not only required for the self-propagation of prions, but also for mediating the toxicity induced by PrPSc [2], a process that results in endoplasmic reticulum (ER) stress and ultimately in impaired protein translation [3]. But how can PrPC, an extracellular GPI-linked protein, initiate intracellular central nervous system (CNS) toxicity? Most likely this process requires mediation by transmembrane constituents. Indeed PrPC has been shown to interact with transmembrane signal-transducing proteins [4] and disturbing these interactions might lead to the neurotoxicity seen in prion diseases [5].
Among the proteins interacting with PrPC are glutamate receptors [6]. N-methyl-D-aspartate receptors (NMDAR) are crucial regulators of glutamatergic transmission, and loss of both synapses and neurons has been attributed to inappropriate NMDAR activation [7, 8]. Metabotropic glutamate receptors (mGluRs) may also play a role in prion diseases. Changes in mGluR1, leading to reduced expression levels of phospholipases, were observed in the cerebral cortex of Creutzfeldt-Jakob disease (CJD) patients [9]. Also, impairment of the mGluR1/1-phosphatidylinositol 4,5-bisphosphate phosphodiesterase 1 (PLC1)/protein kinase C (PKC) signaling pathway has been observed in a murine model of BSE. Abnormal mGluR1 signaling correlated with PrPSc deposition, histological changes, and clinical scores [10].
A role for group-I mGluRs is emerging in a multitude of CNS disorders including Fragile X syndrome, ischemia, multiple sclerosis, amyotrophic lateral sclerosis, Huntington’s, and Parkinson’s disease [11–18]. In Alzheimer’s disease (AD), PrPC and mGluR5 may directly contribute to disease manifestation and toxicity of amyloid-β (Aβ) aggregates. Aβ oligomers can bind to PrPC at the cell surface [19] and form complexes that contain mGluR5 [20]. In a mouse model of Aβ deposition, cognitive decline and synaptic alterations were rescued by mGluR5 inhibition [21]. Furthermore, PrPC-mGluR5 coupling is involved in Aβ-mediated inhibition of LTP and Aβ-facilitated LTD in vivo [22], and genetic ablation of mGluR5 reverses disease-related memory deficits in a murine model of AD (APPswe/PS1ΔE9) [23]. In another study, exposure of cortical APPswe/PS1ΔE9 neuronal cultures to Aβ oligomers upregulated mGluR1 and PrPC α-cleavage, whereas activation of group-I mGluRs increased PrPC shedding from the membrane [24]. In primary hippocampal neurons, membrane-bound Aβ oligomers induce toxicity by promoting clustering of mGluR5 in synapses, resulting in elevated intracellular calcium and synaptic failure [25]. All these studies suggest an involvement of group-I mGluRs in the pathogenesis of AD. On the other hand, others have reported that neither PrPC ablation nor overexpression had any effect on neurotoxicity in AD models [26–29]. As a possible explanation for these discrepancies, it has been suggested that only a limited oligomeric fraction of Aβ [30] interacts with mGluR5 [31].
Here we focused on the role of group-I mGluR-PrPC interaction in prion disease. We found that toxic prion-mimetic compounds increased mGluR5 clustering and accumulation at dendritic heads, close to the synaptic source of glutamate. Moreover, pharmacological inhibition of mGluR1 and mGluR5, as well as genetic ablation of the Grm5 gene encoding mGluR5, protected organotypic slice cultures against the toxicity of prions and of prion-mimetic compounds. Finally, pharmacological inhibition of mGluR5 improved the neurological status and, to some extent, the survival of prion-infected mice.
Cerebellar and hippocampal organotypic cultured slices (COCS and HOCS, respectively) [32, 33] prepared from PrPC overexpressing tga20 mice [34] can be infected with the Rocky Mountain Laboratory (RML) strain of prions and undergo neurodegeneration after ca. 5 weeks [32]. The time course and extent of neurodegeneration can be measured by morphometric assessment of the area of the cerebellar granule cell layer (CGL) immunoreactive to antibodies against the neuronal NeuN antigen.
We inoculated COCS and HOCS with brain homogenate from CD1 mice that had been infected with RML prions (passage #6, henceforth called RML6). For control, slices were inoculated with non-infectious brain homogenate (NBH) derived from healthy CD1 mice. Starting at 21 days post infection, slices were treated with a range of concentrations of either N-cyclohexyl-6-N-methylthiazolo[3,2-a]benzimidazole-2-carboxamide (YM202074) [35], 2-methyl-6-(phenylethynyl)-pyridine (MPEP) [36] or Mavoglurant (AFQ056) [37] which specifically inhibit mGluR1 and mGluR5, respectively.
MPEP, AFQ056 and YM202074 prevented CGL loss in COCS at concentrations as low as 10 nM (Fig 1A and 1B) and 36 nM (Fig 1C, 1D, 1G and 1H), respectively. The protective effect of YM202074 and MPEP was further confirmed in wild-type slices (S1A and S1B Fig). Extremely high MPEP concentrations (3–10 μM) were not intrinsically toxic (S1C Fig) as previously reported [36], but failed to protect against prion toxicity in tga20 mice (S1C and S1D Fig). Also in HOCS, prepared from 4–6 days old tga20 mice, MPEP significantly suppressed neuronal loss after prion infection at concentrations as low as 36 nM (Fig 1E and 1F).
The beneficial effects of mGluR5 inhibition ex vivo encouraged us to assess whether MPEP can potentially rescue prion pathogenesis in vivo. C57BL/6J male mice were inoculated intracerebrally with 3 or 5 log LD50 units of RML6 prions as described [38] and chronically treated with MPEP. Control mice were inoculated with NBH. In order to record the neurological deficits associated with prion disease, we utilized the rotarod behavioral test which measures a combination of motor performance, coordination and balance [39]. Rotarod performance was similar in RML6- and NBH-inoculated mice until 18 weeks following prion inoculation. Starting from 19 weeks post inoculation, mice receiving control food showed a progressive decline in rotarod performance. The performance of MPEP-treated mice declined, but less rapidly. This improvement was lasting and detectable until the very late stages of the disease (22–23 weeks post inoculation; Fig 2A and 2B), suggesting that the progression of the disease was delayed by MPEP.
At very late time points, the general health status of all mice deteriorated to an extent that made it impossible to accurately measure their rotarod performance and eventually required euthanasia. Nevertheless, MPEP-treated mice showed a modest, though significant, prolongation of survival (Fig 2C and 2D). The median survival for untreated vs MPEP-treated RML6-inoculated C57BL/6J mice was, respectively, 183 vs 190 days post inoculation (dpi) after injection with 3 log LD50 units of prions and 188 vs 195 dpi after inoculation with 5 log LD50 units (P = 0.0008 and 0.0231 respectively; log-rank test). Control mice injected with NBH and treated with MPEP exhibited stable rotarod performance during the entire test period, up to 23 weeks post-injection (S2A Fig). No significant changes in average food and water consumption were observed between control and treatment groups during the experiment (S2B Fig). To determine the exposure of the brain to MPEP, mice treated with control and MPEP food were sacrificed at two time points, corresponding to the active and the inactive phase of the mice across the circadian circle. The average brain-to-blood ratio for the MPEP concentration was around 1, indicating good brain penetration of MPEP (S2C Fig, S1 Table).
Antibody-derived molecules targeting the globular domain (GD) of PrPC (termed GDLs) are acutely neurotoxic [40, 41] and activate similar cascades as bona fide prion infection [42]. Single chain POM1 miniantibodies (scPOM1), fusion proteins containing only the variable regions of the heavy (VH) and light chains (VL) of the antibody connected with a short linker peptide, were previously shown to be sufficient to induce toxicity in COCS [41]. To investigate if pharmacological inhibition of mGluR1 and mGluR5 rescues GDL toxicity, we exposed tga20 COCS to the GDL agent scPOM1, followed by YM202074, MPEP and AFQ056 treatments. Treatment with scPOM1 led to almost complete CGL loss within 8 days of treatment. No CGL loss occurred in control treatment where scPOM1 was blocked by pre-incubation with a molar excess of recombinant PrP (recPrP). Treatment with MPEP significantly reduced CGL loss in scPOM1-treated slices. As with prion infections, MPEP treatment (at concentrations as low as 10 nM) was sufficient to rescue the loss of CGL, whereas high concentrations (≥1μM) did not show protective activity (Fig 3A and 3B). Even lower MPEP concentrations (3nM) were sufficient to rescue scPOM1-induced toxicity in COCS (S3E and S3F Fig). AFQ056 and YM202074 treatment (at concentrations as low as 36nM) also significantly reduced the toxicity of scPOM1 (Fig 3C, 3D, 3G and 3H) in COCS.
The protective effect of mGluR1 and mGluR5 inhibitors (YM202074 and MPEP respectively) was further confirmed in wild-type slices. No additional effect was observed upon double MPEP/YM202074 inhibition (S3A and S3B Fig). Similarly to COCS, HOCS treated with scPOM1 exhibited conspicuous toxicity after 8 days of treatment. Neuronal loss was monitored by morphometric analysis of NeuN immunofluorescence, and was readily visible in GDL-treated samples, whereas the survival of hippocampal neurons exposed to scPOM1 (Fig 3E and 3F) was greatly increased by treatment with MPEP. In contrast, no protection was observed upon treatment with the selective group III agonist L-2-amino-4-phosphonobutyrate (L-AP4) [43] and the potent group II/III antagonist (RS)-α-Cyclopropyl-4-phosphonophenylglycine (CPPG) [44] of metabotropic glutamate receptors (S3C and S3D Fig). Hence toxicity of both infectious prions and prion-mimetic GDLs was prevented by pharmacological inhibition of mGluR1 or mGluR5.
Cerebellar organotypic slice cultures from Grm5-/-, Grm5+/- and Grm5+/+ littermates were treated with the anti-GD single-chain miniantibody scPOM1 [45], which acts as a prion-mimetic compound. Exposure to scPOM1 led to the loss of cerebellar granular layer (CGL) neurons in Grm5+/+ slices, but neither in Grm5-/- nor in Grm5+/- slices (Fig 4A and 4B). We then inoculated cerebellar and hippocampal organotypic slice cultures from Grm5-/-, Grm5+/- and Grm5+/+ littermates with RML6 prions or control NBH homogenate. In COCS, both Grm5-/-and Grm5+/- slices are protected against RML6 toxicity (Fig 4C and 4D). In HOCS, genetic ablation of mGluR5 was protective against prion-induced toxicity (Fig 4E and 4F).
To assess the role of mGluR5 in prion infections in vivo, we infected Grm5-/-, Grm5+/- and Grm5+/+ littermates with RML6 prions (5 log LD50). In line with a recently published study [46], no significant difference in survival was observed between Grm5-/-, Grm5+/- and Grm5+/+ mice (S4A Fig).
The latter finding was unexpected and prompted us to investigate the possibility of compensatory mechanisms. Both group-I metabotropic glutamate receptors, mGluR1 and mGluR5, can associate with PrPC and induce similar intracellular pathways [47] suggesting functional redundancy between these two receptors. In order to detect a possible epistasis between mGluR1 and mGluR5, we assessed mGluR1 and mGluR5 protein levels in cerebellum, cortex and hippocampus of Grm5-/-, Grm5+/- and Grm5+/+ mice (S4C and S4D Fig).
At 10 days of age, mGluR5 expression was similar in cerebellum, hippocampus and cortex as described [48], whereas mGluR1 was highest in the cerebellum (S4C Fig). Interestingly, we observed an increased expression of mGluR1 in all the three tested regions of Grm5-/- brains. We further assessed mGluR1 and mGluR5 levels at later time points (45–180 days). Expression of mGluR5 decreased in all brain regions with increasing age, whereas expression of mGluR1 remained stable. However, we detected increased mGluR1 expression in Grm5-/- brains. In the cortex, we observed increased expression of mGluR1 in samples from 45-day old Grm5-/- mice compared to Grm5+/+ littermates (S4D Fig, middle right panel). In the hippocampus, we observed increased expression of mGluR1 in samples from 90-day old Grm5-/- mice (S4D Fig, bottom right panel) and in samples from both Grm5+/- and Grm5-/- 180-day old mice (S4D Fig, lower right panel, lanes 7, 8 & 9 and quantification). In the cerebellum, we observed increased expression of mGluR1 in samples from 90-day old Grm5-/- mice compared to wild-type control littermates (S4D Fig, upper right panel).
We then tested whether treatment with MPEP also enhances the expression of mGluR1. mGluR1 expression levels were assessed in whole-brain lysates from 1-year old control wild-type mice, NBH-inoculated wild-type mice, and NBH-inoculated wild-type mice that received MPEP food. However, no differences were observed in the mGluR1 expression levels between the samples (S4B Fig), suggesting that compensatory Grm1 upregulation is developmentally controlled.
PrPC interacts with mGluR1 and mGluR5 [21, 47]. We confirmed these results by immunoprecipitating brain homogenates from wild-type (C57BL/6J) or Prnp knockout mice (Prnpo/o) using antibody POM1 against PrPC, followed by Western blotting with antibodies to mGluR1 and mGluR5. The group-I mGluRs, which migrate as SDS-resistant oligomers at 250kDa [49], were found to co-precipitate with PrPC (Fig 5A). When we blocked the antigen-recognition domain of POM1 with recombinant PrP, mGluR1 and mGluR5 no longer co-precipitated with PrPC (Fig 5A). Western blots of brain lysates (total extracts; TEs) did not reveal any changes in the concentration of mGluR1 and mGluR5 protein between wild-type tga20 and Prnpo/o homogenates (Figs 5A and S5A). In contrast, mGluR6 and mGluR2/3 did not co-precipitate, confirming the specificity of the interaction (S5B Fig).
The residues 91–153 of PrPC participate to the interaction with mGluR5 [20]. To confirm these findings and to identify the domain of PrPC mediating its interaction with mGluR5, we studied a panel of transgenic mice expressing variants of PrPC bearing deletions in the flexible tail (FT) regions, designated ΔC, ΔCC, ΔF, ΔOR, and ΔHC [50–54] (S5E Fig). In each line of mice, we immunoprecipitated PrPC from brain using POM1 antibody (specific information and binding sites on PrPC are provided in S5F Fig and Table 1) and measured the co-precipitation of mGluR5. Most FT-mutated PrPC variants showed an impaired capacity to co-precipitate mGluR5, with deletions of residues 51–90 and 32–134 showing the most striking reduction (S5C Fig). Conversely, when we performed immunoprecipitations of mGluR5 followed by Western blotting for PrPC, we found that deletions spanning residues 111–134 affected the interaction most profoundly (Fig 5B).
We also analyzed the capacity of PrPC mutants to immunoprecipitate mGluR1. While all examined FT mutations decreased the interaction of PrPC with mGluR1, deletions affecting residues 51–90 showed the most significant reduction (S5D Fig). Immunoprecipitation of mGluR1 revealed that PrPC deletions spanning residues 51–90 and 111–134 had the strongest effect on its interaction with mGluR1 (Fig 5C). Finally, we observed that deletion of mGluR5 had no effect on co-precipitation of PrPC with mGluR1 (Fig 5C), indicating that mGluR1 and mGluR5 interact with PrPC independently of each other.
These results suggest that the interaction domain between PrPC and mGluR5 resides at the N-terminal region of PrPC and is larger than previously inferred, with residues 32–114 participating to the in vivo interaction. The interaction domain between PrPC and mGluR1 also resides at the N-terminal region of PrPC and spans residues 51–90 and 111–134.
PrPSc deposition is accompanied by neurodegeneration, vacuole formation and activation of microglia and astrocytes [55]. MPEP treatment did not affect the accumulation of PrPSc in prion-infected mice and slices (S6A–S6C Fig), yet it reduced vacuole formation. Although the numbers of vacuoles in control and MPEP treated groups were similar, vacuoles were smaller in cerebella of MPEP-treated mice (Fig 6A and 6B). Astrogliosis, assessed by immunohistochemistry for glial fibrillary acidic protein (GFAP), was prominent in terminally sick prion-infected mice but not in NBH-inoculated mice. MPEP treatment reduced the astrogliosis in the hippocampus of prion-infected mice (Fig 6C and 6D), but not in the cerebellar granule cell layer (S6D Fig), as expected from the decreased expression of mGluR5 in the cerebellum of older mice. These findings corroborate the interpretation that MPEP reduces prion toxicity even if it does not affect prion load.
Clusters of mGluR5 accumulate around excitatory synapses, but are also found at extra-synaptic sites (S7A Fig). Increased size of synaptic mGluR5s clusters is associated with toxic calcium influx [21, 25, 56]. Therefore, we asked whether the prion-mimetic POM1 antibody altered the clustering of mGluR5s. POM2 and POM3 antibodies were also used in parallel (for details about POM antibodies and their epitopes, see Table 1). Specific information and binding sites on PrPC for all antibodies are provided in (S5F Fig, Table 1) and materials and methods.
Exposure of live neurons to POM1, significantly increased the size of mGluR5s clusters compared to POM2 or POM3 exposure (Fig 7A and 7B), however no change was observed with the NMDA and AMPA receptor clusters (S7B–S7E Fig), suggesting formation of abnormal, potentially deleterious mGluR5 signaling platforms [57]. Next, we examined the fluorescence of dendritic spines of neurons expressing an mGluR5-pHluorin fusion protein. Spines in mGluR5-pHluorin transfected neurons indeed co-localize with post-synaptic marker Homer, which is also a scaffolding protein for mGluR5 (S7F Fig). We observed increased accumulation of mGluR5s in dendritic spines following exposure to POM1, but not to POM2 or POM3 (Fig 7C and 7D).
Both mGluR5 and PrPC are enriched in postsynaptic densities [21]. In order to assess if the changes in mGluR5s level in spines correlated with PrPC level in spines, we performed photo-activated localization microscopy (PALM) on neurons expressing a PrPC tagged with dendra2 fusion protein [58] (Fig 7E). PALM images were obtained from single-molecule detection with a pointing accuracy of 20 nm [58]. The PrPC-Dendra fluorescence patterns showed both clustered and diffused staining (Fig 7E, control); we observed an increased enrichment within dendritic spines following POM1 but not POM1+2 exposure (Fig 7E and 7F). Furthermore, exposure to Fab1-POM2, which was previously found to protect against POM1 toxicity [41], induced a small but significant reduction in PrPC enrichment within dendritic spines. Therefore, Fab1-POM1 and Fab1-POM2 may exert opposite effects on the topology and size of mGluR5 clusters, with POM1 inducing abnormal accumulation and translocation to dendritic spines.
Prion toxicity is ultimately mediated by unfolded-protein responses [3, 59], yet it is unclear how these are triggered by PrPSc which is primarily extracellular. The group-I metabotropic glutamate receptors mGluR5 and mGluR1, G protein-coupled receptors that interact with PrPC [19, 21, 25], may represent one such link. We found that mGluR5 and mGluR1 inhibitors prevented neurodegeneration in prion-infected organotypic slice cultures and protected against prion-mimetic globular-domain ligands [41]. Inhibition of group-I mGluRs may reduce glutamatergic signaling and calcium overload in prion-infected cells [60], similarly to models of Alzheimer’s disease [21, 25].
PrPC associates with group-I mGluRs [47] and modulates the signaling activity of mGluR5 [20]. If prion toxicity depends on the direct interaction of PrPC to group-I mGluRs, it may modify the subcellular distribution of mGluR5. Indeed, prion-mimetic antibodies selectively increased clustering of mGluR5 (but not of AMPA and NMDA receptors) in dendritic spine heads, potentially sensitizing them to synaptic glutamate. Prion-mimetic antibodies also increased the level of PrPC in spines, reinforcing the notion that mGluR5 and PrPC are part of the same complex whose accumulation at excitatory synapses instigates neurotoxicity in prion diseases. The impact of POM1 on mGluR5 enrichment within dendritic spines is modest, possibly because only a small fraction of mGluR5 is associated with PrPC. Increased cell surface clustering may also slow down endocytosis, thereby increasing the amount of functional mGluR5s [21, 23, 61]. Thus, mGluR5 clustering at synapses may amplify responses to glutamate, thereby exaggerating Ca2+ influx and leading to spine loss, a primary event in prion diseases [62]. The POM2 antibody [45] against the Flexible Tail (FT) of PrPC is neuroprotective in vivo and in vitro. Since both POM2 and mGluR5 bind to the N-terminus of PrPC, binding of mGluR5 to PrPC may facilitate its activation whereas POM2 may compete for PrPC binding (Fig 8).
Although mGluR5 inhibition delayed neurological deterioration, survival was only modestly (though significantly) improved. These findings support the concept that mGluR5 inhibition alleviates the symptoms of the disease whereas prion replication progresses unabated. Eventually, the prion load may exert neurotoxicity through mGluR5-independent mechanisms including mGluR1 activation. Not all neurons express mGluR5 [63, 64]; neurons essential for survival may be mGluR5-negative and possibly mGluR1-positive.
Upregulation of mGluR5 can go along with glial activation [56, 65, 66]. We observed reduced GFAP immunoreactivity in hippocampi of MPEP-treated animals (Fig 6C). Conversely, MPEP was unable to suppress glial activation in adult cerebella (S6D Fig) where mGluR5 expression is low, suggesting that dampened neuroinflammation was beneficial.
Genetic ablation of Grm5 was protective against the toxicity of prion-mimetic antibodies and prion infections in organotypic slices. This effect was haploinsufficient, as hemizygous Grm5+/- slices were also protected. Surprisingly, a previous report [46] and this study show that Grm5 ablation does not ameliorate the clinical manifestation of scrapie in vivo. This discrepancy is most likely due to the conspicuous mGluR1 upregulation in Grm5-/- and Grm5+/- mice.
Co-immunoprecipitations from transgenic mice expressing PrPC with amino-proximal deletions [50–54] showed that both mGluR1 and mGluR5 independently interact with the N-proximal flexible tail of PrPC. However, the boundaries of the interacting domain differ, with PrPC residues 32–134 (with residues 51–90 (ΔOR) and 111–134 (ΔHC) acting as important interaction sub-regions) mediating the interaction with mGluR5. The interaction domain appears to extend over the previously reported borders [31]. The interaction domain between PrPC and mGluR1 also resides at the N-terminal region of PrPC and spans residues 51–90 (ΔOR region) and 111–134 (ΔHC region).
Although both Grm5 genetic deletion and mGluR5 pharmacological inhibition (MPEP) did not prevent prion disease, MPEP significantly improved locomotor abilities until the later stage of disease, decreased the size of spongiform vacuoles, and reduced the extent of hippocampal astrogliosis. These observations are aligned with reports of abnormal expression of group-I mGluRs and mGluR1 signaling in Creutzfeldt-Jakob disease and bovine spongiform encephalopathy [10, 67]. Additional mGluRs may also play a role, and a genome-wide association study identified an mGluR8 variant as a marker for sCJD risk outside the PRNP locus [68].
The above data suggest that group-I mGluRs inhibition may attenuate dysfunctions associated with prion diseases, for which there are no disease-modifying therapies. It is unsurprising that mGluR5 antagonists have only a moderate effect on survival, since this therapeutic modality is likely to affect downstream consequences of prion toxicity rather than quenching prion propagation. Because of their orthogonal mode of action, these antagonists may represent ideal compounds for combination therapy with compounds inhibiting prion replication. Because they are well-tolerated and have high bioavailability and blood-brain-barrier penetration [15, 69, 70], mGluR5 antagonists may be useful for enhancing the quality of life of prion patients—a legitimate and important aim even if the overall life expectancy may not be dramatically improved.
The purpose of this study was to evaluate the therapeutic potential of group I metabotropic glutamate receptor (mGluR1, mGluR5) inhibition in ex vivo and in vivo models of prion disease. We selected highly specific and well-studied pharmacological inhibitors of mGluR1 and mGluR5, YM202074 and MPEP and AGQ056 respectively, with known specificity and efficiency. To ensure availability of the inhibitors to the brain of prion-infected mice thorough pharmacokinetic and pharmacodynamic analyses were performed. We further extended our study to transgenic mice, knock out for the glutamate receptors being studied. For slice experiments, treatments were randomly assigned to individual wells. For mouse experiments, treatments were randomly assigned to age- and sex-matched mice; experimenters were blinded to experimental group while performing the animal experiments. For experiments with transgenic mice, similar number of heterozygotes and wild-type littermates were included as controls. Mice were sacrificed at the terminal stage of the disease. For analysis, random numbers were assigned to each subject or experimental group.
All animal procedures were approved by the local Ethical Committee (Animal Experimentation Committee of the Canton of Zurich, permit 200/2007; 41/2012; 90/2013) in accordance with the Swiss federal, Ethical Principles and Guidelines for Experimenting on animals (3rd edition, 2005). All efforts were made to minimize the suffering and the number of animals used.
C57BL/6J wild-type mice were purchased from Jackson laboratories. Male mice were selected because they do not have estrous cycles that can complicate pharmacology. Prnpo/o and Prnpo/o;tga20+/+ (tga20), were on a mixed 129Sv/BL6 background [71, 72]. Transgenic mice expressing mutated PrPC were utilized for immunoprecipitation experiments. The production and relevance to disease phenotype of the Tg mice expressing N-terminal deletion mutants of PrPC (termed ΔC, ΔCC, ΔF, ΔOR, and ΔHC) have been previously reported [50–54]. Grm5+/- embryos [73, 74] were acquired from Dr. Gasparini and were revitalized at the transgenics facility of the University Hospital of Zurich. Grm5 null mice were derived from breeding of these mice.
2-Methyl-6-(phenylethynyl)-pyridine (MPEP) [36] chronic treatment was initiated at the time of prion inoculation. A dose of 30 mg of MPEP/kg of body weight was selected [75]. The drug was incorporated into chow to achieve voluntary consumption and constant drug administration. Control, untreated groups received the same type of food lacking the drug. For this study, mice between 2 and 4 months of age at the time of prion inoculation/beginning of MPEP treatment were utilized.
To determine PK values in mice fed with food pellets containing MPEP (250mg/kg; Provimi Kliba SA, Rinaustrasse 380, CH-4303 Kaiseraugst), 10 C57BL/6J mice were fed MPEP-food pellets for 15 days and sacrificed to measure the blood/brain ratio of MPEP. Based on an average intake of 3 gram food pellets per day and a body weight of approximately 25 g, a dose of 30mg/kg/day was established. The MPEP concentration was determined by liquid chromatography separation followed by mass spectrometry (LC-MS). Control mice (a total of 8 C57BL/6J mice) received normal food. Mice were sacrificed at two different time points, corresponding to the active and the inactive phase of the mice across the circadian circle and exposures of MPEP in blood and brain were measured.
Organotypic cerebellar cultured slices, 350 μm thick, were prepared from 9–12 day-old pups according to a previously published protocol [32]. Organotypic hippocampal cultured slices, 350 μm thick, were prepared from 4–6 day-old pups according to a previously published protocol [33]. Cultures were kept in a standard cell incubator (37°C, 5% CO2, 95% humidity) and the culture medium was changed three times per week.
Inoculations were performed with either infectious brain lysate (RML6) or non-infectious brain homogenate (NBH). Slices were inoculated (as free-floating sections for 1 h at 4°C) with 100μg brain homogenate per 10 slices. After washing in GBSSK, they were cultured on a 6-well Millicell-CM Biopore PTFE membrane insert (Millipore) according to previously published protocol [60]. Drug-treated tga20 slices were maintained until 45 dpi, fixed and analyzed by NeuN morphometry (analySIS vc5.0 software). Neurotoxicity was defined as significant NeuN+ neuronal layer loss over NBH treatment. Slices prepared from GRM5-/-, GRM5+/- and GRM5+/+ littermates were maintained until 60 dpi, fixed and analyzed by NeuN morphometry (analySIS vc5.0 software). Neurotoxicity was defined as significant NeuN+ neuronal layer loss over NBH treatment.
For globular domain ligand (GDL) treatment, toxicity in slices was induced by exposure to ligands, toxic anti- PrPC antibodies targeting the globular domain, such as single chain scPOM1 mini-antibody, after a 14-day recovery period; allowing the initial gliosis induced by tissue preparation to subside, according to previously published protocol [41]. tga20 COCS were exposed to scPOM1 (200 nM, 8 dpe), or to control treatment (200 nM scPOM1/210nM recPrP, 8 dpe), immunostained for the neuronal marker NeuN and counterstained with DAPI. Slices were imaged and analysed as previously described. Antibody treatment was randomly assigned to individual wells.
Treatment with the specific inhibitors 2-Methyl-6-(phenylethynyl)-pyridine (MPEP) [36], AFQ056 (Mavoglurant) [37] or N-cyclohexyl-6-N-methylthiazolo[3,2-a]benzimidazole-2-carboxamide (YM202074) [35] was initiated at the time of GDL addition (14dpe) for the GDL toxicity model (treated slices were maintained until 28 dpe for POM1 treatment and until 22dpe for scPOM1 treatment) [41] and at 21 days post-inoculation (dpi) for prion-infected slices, when PrPSc accumulation was already discernible [32]. Drug treatments were re-added at every media change [36]. Post-treatment slices were fixed in 4% paraformaldehyde (PFA), immunostained for the neuronal marker NeuN and counterstained with DAPI. Slices were imaged at 4x magnification on a fluorescence microscope (BX-61, Olympus) analyzed by NeuN morphometry (analySIS vc5.0 software). Neuroprotection was defined as significant neuronal layer rescue over toxic-antibody treated, non-drug treated slices.
Inoculum of the RML6 strain of mouse-adapted scrapie prion was prepared from pooled 10% w/v brain homogenates of RML6 terminally sick CD1 mice. C57BL/6J mice were inoculated with serial dilutions (10−3 and 10−5) of the RML6 inoculum. C57BL/6J mice were injected intracerebrally (i.c.) with 30μl of brain homogenate prepared in a solution of PBS/5% BSA, containing 3log LD50 units or 5log LD50 units of the RML6 strain. Control mice received 30μl of NBH derived from healthy CD1 mice. Scrapie was diagnosed according to clinical criteria (ataxia, kyphosis, priapism, and hind leg paresis). Mice were sacrificed on the day of onset of terminal clinical signs of scrapie. The operator was blinded to drug treatment.
The rotarod test was used to assess motor coordination and endurance at defined timepoints after prion inoculations. A rotarod machine (Ugo Basile) with five cylinders (3cm diameter) separated by dividers (25cm diameter) in five lanes, each 57mm wide, was utilized. Before the training sessions, the mice were habituated to stay on the rotating rod (4 rpm lowest speed) for 3 sessions lasting 1–2 minutes each and separated by 10 minute intervals. The test phase started 30 minutes after the last habituation session and consisted of 3 trials separated by 15 minute inter-trial intervals. For each test session the mouse was placed on a rotating rod, which accelerated from 5 to 40 rpm. Each test session lasted a maximum of 5min. Latency to fall was assessed when the mouse was no longer capable of riding on the accelerating rod and slipped from the drum. Test sessions were always performed at the same time of the day, mice were tested in a randomized manner and the operator was blind to drug treatment.
Adult Prnpo/o, tga20+/+ (tga20), and C57BL/6J mice were euthanized and their brains were dissected. Brain samples were snap frozen in liquid nitrogen. Samples were subsequently homogenized in ice cold Lysis Buffer (1% Igepal (NP-40) in 1x PBS, pH 7.4) supplemented with protease (EDTA-free) and phosphatase inhibitor cocktail mix (Roche). Protein concentration was determined using the bicinchoninic acid assay (Pierce). Following immunoprecipitation of PrPC with a specific anti-PrP monoclonal antibody (POM1 or POM2) and addition of Dynabeads M-280 Sheep anti-mouse (#311201D, Thermo Fischer Scientific), samples were prepared in loading buffer (NuPAGE, Invitrogen) and incubated at 37°C for 30 min. For the immunoprecipitation data shown in S4B and S4C Fig, the samples were incubated at 95°C for 5 min; this resulted in disruption of dimers of mGluR5. However this did not have any effect on the immunoprecipitated fractions. The samples were migrated on 4–12% NuPage gels and transfered onto the PVDF membrane. For reverse immunoprecipitation experiments, the subsequent experimental set-up was used. Following immunoprecipitation of mGluR1 or mGluR5 with a specific anti-mGluR1/5 polyclonal antibody (Cell Signalling Technology #12551 or #55920 respectively) and addition of Dynabeads Protein G (#10003D, Thermo Fisher Scientific), samples were prepared in loading buffer (NuPAGE, Invitrogen) and incubated at 37°C for 10–30 min [76]. The samples were migrated on 4–12% NuPage gels and transferred onto the PVDF membrane.
All compounds were purchased from Sigma-Aldrich unless otherwise stated. Monoclonal anti PrP antibody POM1 (1:5000) was generated as described previously [45]. Anti-mGluRs antibodies against representative receptors of each group, targeting the N-terminal domain were utilized: anti-mGluR5 #ab53090 (Abcam) or AB5675 (Millipore), anti-mGluR1 [EPR13540] (ab183712) (Abcam), anti-mGluR2+3 #ab6438 (Abcam) and anti-mGluR6 #AGC-026 (Alomone labs). Secondary antibodies were horseradish peroxidase (HRP)- conjugated rabbit anti–mouse IgG1 (1:10,000, Zymed) and goat anti–rabbit IgG1 (1:10,000, Zymed). Blots were developed using SuperSignal West Pico chemiluminescent substrate (Pierce) and visualized using the VersaDoc system (model 3000, Bio-Rad). Rocky Mountain Laboratory strain (RML; passage #6) prions (RML6) were amplified in CD1 mice by intracerebral inoculation into the lateral forebrain of 30 μl of 1% (wt/vol) brain homogenate. The mGluR5 antagonists MPEP and AFQ056 were kindly provided by Novartis. The mGluR1 antagonist YM202074 was purchased from Tocris Bioscience (Ellisville, USA).
Immunohistochemistry of fixed organotypic slices and subsequent NeuN morphometric analysis was performed according to previously published protocols [41, 60].
Stainings were performed on sections from brain tissues fixed in formalin and treated with concentrated formic acid to inactivate prions. Partially protease-resistant prion protein deposits, astrogliosis and microglia deposition were visualized by staining brain sections with the SAF84 antibody (1:200, SPI bio), GFAP (1:1000, Millipore) and IBA1 (1:2500, WAKO) respectively on a NexES immunohistochemistry robot (Ventana instruments) using an IVIEW DAB Detection Kit (Ventana), after preceding incubation with protease 1 (Ventana). Images of DAB stained sections were acquired using the NanoZoomer scanner (Hamamatsu Photonics) and NanoZoomer digital pathology software (NDPview; Hamamatsu Photonics). Quantifications of IBA1, GFAP staining and vacuoles in mouse sections were performed on acquired images; regions of interest were drawn on a Digital Image Hub (Leica Biosystems) and analyzed as previously described [77].
Hippocampal neurons were prepared from embryonic day 18 (E18) C57/BL6 mice (Janvier Labs, France). Freshly dissociated (trypsin) cells were plated (80,000 cells per 18 mm coverslip per ml) in neuronal attachment media consisting of 10% horse serum, 1 mM sodium pyruvate, and 2 mM glutamine in MEM for 3h. The attachment medium was replaced and cells were maintained in serum-free neurobasal medium supplemented with B27 (1X) and glutamine (2 mM). 300 μl of fresh medium was added once a week.
mGluR5-pHluorin construct [78]was generated and kindly provided by Lili Wang and Christian Specht. Dendra2 was inserted between residues Q222 and A223 of mouse prion protein. GluN2A-GFP was kindly provided by Andrea Yao and Pierre Paoletti. Transfection was performed on DIV 17–18 neurons using Lipofectamine as described recently [58]. Transfection medium (TM) was composed of 1 mM sodium pyruvate and 2 mM glutamine in nerobasal medium (Invitrogen). 0.5 μg of plasmid and 2 μl of lipofectamine- 2000 reagent were used for each coverslip. All in vitro experiments were performed on mature neurons (DIV 21–24)
Immunocytochemistry of mGluR5 (rabbit polyclonal, Millipore, AB5675, 1:200 dilution) or GluR2-AMPA receptor (rabbit polyclonal, Synaptic System, 182103, 1:400 dilution) was performed following methanol fixation / permeabilization (10 min at -20°C; methanol pre-stored at -20°). Image thresholding using wavelet decomposition to identify fluorescent clusters (mGluR5 and GluR2-AMPA immunoreactivity or GluN2-GFP fluorescence) has been described in previous studies [25, 58]. Size of clusters denotes the total fluorescence intensity of the given cluster. Images were acquired using Leica Inverted Spinning Disk microscope (DM5000B, Coolsnap HQ2 camera, Cobolt lasers) using 100X objective (field of view = 1392 x 1040 pixels) and a pixel size of 60.5nm. For estimation of mGluR5 fluorescence within dendritic spines, ratio of fluorescence within a circular region of fixed size (6 pixel) on spine head to the shaft below was measured using ImageJ program.
PALM was performed on live neurons expressing PrPc-Dendra2 and the microscope setup and lasers used have been recently described in detail [58]. Unconverted Dendra2 has excitation and emission maxima at 490 and 507 nm (green range) while converted Dendra2 protein has excitation and emission maxima at 553 and 573 nm (red range). First, all signal in red channel was photo-bleached to allow detection of single molecule events arising due to the switching of Dendra2 from green to red channel. Single molecule events of Dendra2 were imaged using laser 561 nm (0.5kW, used at 300-400mW) while activating with 405 nm laser (100 mW power, used at 2–5 mW). PrPc-Dendra2 was imaged for 5000–6000 frames. Single molecule detections using in-house software has been used and described in previous publications [58]. Density of detections (number/area) of single-molecule on spine head was divided by density of detections over a dendritic shaft to obtain spine enrichment of PrPC-Dendra2.
Dendrites were not filled with any additional post-synaptic marker. Mature neurons (DIV 21–24) were transfected with mGluR5-SuperEcliptic pHluorin. The pHluorin-tag allows the visualization of only cell-surface mGluR5s and the neuronal membrane, which is then visually recognizable. We have recently used this plasmid to compute the diffusion dynamics of mGluR5s within dendritic spines [78] In this study, we quantified the spines enrichment of all recognizable spines; considering that visually recognizable spines in mGluR5-pHluorin transfected neurons indeed colocalize with post-synaptic marker, Homer (which is also the scaffold of mGluR5).
Detailed image analysis information is provided in the figure legends. For NeuN morphometric analysis (Figs 1, 3, 4, S1 and S3), NeuN values are normalized to the median NeuN value of the NBH or Ctrl samples respectively. Two-way ANOVA, followed by Bonferroni correction or Log-rank (Mantel-Cox) test was performed in Fig 2, to measure statistical differences between groups. One-way ANOVA followed by Dunnet’s post-hoc test was performed to measure statistical differences between groups. Two-way ANOVA, followed by Bonferroni correction was performed for Fig 4G. For Western Blot quantification in S4 Fig, mGluR1/actin ratios were normalized to the mean Grm5+/+ sample mGluR1/actin ratio in each timepoint (45days, 90days, 180days). One-way ANOVA followed by Tukey’s post-hoc test was performed to measure the statistical differences between the groups. For IP quantification in Figs 5 and S5, densitometric quantitation of PrP signal or mGluR1/5 respectively from the immunoprecipitation was normalized over the ration of PrP/Actin or mGluR1/Actin or mGluR5/Actin signal in TEs respectively. One-way ANOVA followed by Tukey’s post-hoc test was performed to measure the statistical differences between the groups. For immunohistochemistry analysis in Figs 6 and S6, number of GFAP+ cells or vacuoles was quantified in different brain regions. GFAP expression, quantified as the percentage of the “brown” surface occupied by the GFAP staining over the total measured area. Vacuolation, quantified as the percentage of “white” surface occupied over the total measured area. Two-way ANOVA, followed by Bonferroni correction was performed to measure statistical differences between groups. Non-parametric Mann-Whitney test was performed in Fig 7 to measure the statistical differences between the distributions. GraphPad Prism (GraphPad Software) was chosen for the statistical analysis.
|
10.1371/journal.pgen.1003279 | Autophagy Induction Is a Tor- and Tp53-Independent Cell Survival Response in a Zebrafish Model of Disrupted Ribosome Biogenesis | Ribosome biogenesis underpins cell growth and division. Disruptions in ribosome biogenesis and translation initiation are deleterious to development and underlie a spectrum of diseases known collectively as ribosomopathies. Here, we describe a novel zebrafish mutant, titania (ttis450), which harbours a recessive lethal mutation in pwp2h, a gene encoding a protein component of the small subunit processome. The biochemical impacts of this lesion are decreased production of mature 18S rRNA molecules, activation of Tp53, and impaired ribosome biogenesis. In ttis450, the growth of the endodermal organs, eyes, brain, and craniofacial structures is severely arrested and autophagy is up-regulated, allowing intestinal epithelial cells to evade cell death. Inhibiting autophagy in ttis450 larvae markedly reduces their lifespan. Somewhat surprisingly, autophagy induction in ttis450 larvae is independent of the state of the Tor pathway and proceeds unabated in Tp53-mutant larvae. These data demonstrate that autophagy is a survival mechanism invoked in response to ribosomal stress. This response may be of relevance to therapeutic strategies aimed at killing cancer cells by targeting ribosome biogenesis. In certain contexts, these treatments may promote autophagy and contribute to cancer cells evading cell death.
| Autophagy is an act of self-preservation whereby a cell responds to stressful conditions such as nutrient depletion and intense muscular activity by digesting its own cytoplasmic organelles and proteins to fuel its longer-term survival. An understanding of the wide spectrum of physiological stimuli that can trigger this beneficial cellular mechanism is only just starting to emerge. However, this process also has a negative side, since autophagy is exploited in certain pathological conditions, including cancer, to extend the lifespan of cells that would otherwise die. Our analysis of a new zebrafish mutant, titania (ttis450), with defective digestive organs and abnormal craniofacial structure, sheds further light on the physiological and pathological ramifications of autophagy. In (ttis450), an inherited mutation in a gene required for ribosome production provides a powerful stimulus to autophagy in affected tissues, allowing them to evade cell death. The phenotypic consequences of impaired ribosome biogenesis in our zebrafish model are reminiscent of some of the clinical features associated with a group of human syndromes known as ribosomopathies.
| The generation of new ribosomes is the most energy-consuming process in the cell [1]. It requires the coordinated transcription and maturation of 4 different ribosomal RNA (rRNA) molecules and 70 small nucleolar RNAs (snoRNAs) together with the synthesis of approximately 80 ribosomal proteins (RPs) and an additional 170 associated proteins [2]. The regulation of this complex, multi-step process is the major factor determining the potential of a cell to grow and divide [3]. In times of nutrient availability and/or hormonal and growth factor signalling, the onset of ribosome biogenesis is tightly coupled to the translational requirements of a rapidly proliferating cell. In contrast, ribosome biogenesis is down-regulated to conserve energy and restrict unwarranted cell growth and division when the cellular environment is nutrient poor or challenged by harmful stimuli such as hypoxia, reactive oxygen species or genotoxic stress. Inherited impairment mutations in genes that encode components of the ribosome biogenesis machinery or ribosome structure underlie a number of human syndromes, collectively known as ribosomopathies, with a broad range of clinical phenotypes [4]. There is a growing appreciation that sporadically acquired mutations in genes that contribute to ribosome function also increase susceptibility to human cancer, particularly leukemia and lymphoma, although the precise mechanisms involved are only just beginning to emerge [5].
The process of human ribosome biogenesis initiates in the nucleolus with the transcription by RNA polymerase (Pol) I of a 45S pre-rRNA precursor (35S in yeast), which contains the mature 28S, 18S and 5.8S rRNAs interspersed by spacer sequences. A series of processing and chemical modification events mediated by discrete multiprotein/RNA complexes known as the 90S, 66S and 43S pre-ribosomal particles generate the mature 18S, 28S and 5.8S species, respectively and assembles them into the 40S and 60S ribosomal subunits prior to their export from the nucleus to the cytoplasm where they associate to form the functional 80S ribosomes [6]. In yeast, the 90S particle, also known as the small-subunit processome, has been shown to be strictly required for the production of 40S ribosomal subunits containing 18S rRNA [7].
One of the mechanisms through which ribosome biogenesis is coupled to cell growth and proliferation is the Target of rapamycin (Tor) pathway, which is activated by cell surface growth factor and insulin receptors and other growth promoting sensors that detect when nutrients such as amino acids are plentiful. Activation of the Tor pathway stimulates the phosphorylation of S6 kinase (S6K) and 4E-Binding Protein 1 (4EBP1), which regulate ribosome biogenesis and mRNA translation [8], [9]. Activation of Tor also inhibits macroautophagy (hereafter referred to as autophagy), an evolutionarily conserved process that provides a survival mechanism during periods of cell starvation by promoting intracellular recycling of organelles, such as mitochondria and ribosomes [10], [11].
Autophagy describes a complex multi-step process whereby cells sequester a portion of their cytoplasm inside double-membrane vesicles called autophagosomes, which then fuse with lysosomes to form autolysosomes [12]. Inside these vesicles, the captured material, together with the inner membrane, is digested and the released nutrients are recycled. In metazoa, autophagy mediates the catabolic turnover of malfunctioning, damaged or superfluous proteins and organelles to maintain cellular homeostasis during development and in adult life [13]. It is activated in response to multiple forms of cellular stress, including nutrient deprivation, endoplasmic reticulum (ER) stress, accumulation of reactive oxygen species, DNA damage, invasion by intracellular pathogens and intense exercise [14], [15]. Some of these triggers induce autophagy through activation of Tumour protein 53 (Tp53), which increases the expression of the β1 and β2 subunits of AMP-activated protein kinase (AMPK), an evolutionarily conserved sensor of cellular energy levels [16]. AMPK responds to reductions in the ratio of ATP:AMP nucleotides by phosphorylating multiple targets with functions related to energy metabolism, including the Tuberous sclerosis complex (Tsc) protein, Tsc2 and Raptor. These phosphorylation events indirectly inhibit the Torc1 complex, which in its active state inhibits autophagy by negatively regulating the protein kinase, Ulk1 (mammalian orthologue of yeast Atg1). Ulk1, together with Atg13, Fip200 and Atg101, are the key components of a complex that initiates mammalian autophagosome formation [17], [18]. Recent work proposes that AMPK may also induce autophagy independently of Torc1 inhibition by directly phosphorylating Ulk1 [19]–[21]. However, a clear understanding of the AMPK-Ulk1-Torc1 network is yet to emerge [22].
In this study, we employed a zebrafish intestinal mutant, titanias450 (ttis450), as an in vivo model to examine the connection between rRNA processing and autophagy. ttis450 was identified on the basis of its hypoplastic intestinal morphology at 96 hours post-fertilization (hpf) in a focused ENU mutagenesis screen designed to identify mutants with defects in the size and morphology of the endoderm-derived organs [23]. Using positional cloning we identified periodic tryptophan protein 2 homologue (pwp2h) as the mutated gene in ttis450. In yeast, Pwp2 has been shown to be an essential scaffold component of the 90S pre-ribosomal particle, facilitating the binding of proteins such as the U3 snoRNP to the 5′ end of the 35S rRNA precursor [24]. Depletion of Pwp2 in yeast cells results in reduced production of mature 18S rRNA and 40S ribosomal subunits [24], [25]. In agreement with these results, we show that zebrafish Pwp2h plays a conserved role in rRNA processing and ribosome biogenesis. Moreover, we use this in vivo model system to demonstrate a connection between rRNA processing and autophagy which has, to our knowledge, been hitherto unappreciated.
ttis450 is one of several intestinal mutants identified in an ENU mutagenesis screen (the Liverplus screen) conducted on a transgenic line of zebrafish (Tg(XlEef1a1:GFP)s854) harbouring a GFP transgene (“gutGFP”) expressed specifically in the digestive organs [23], [26], [27]. Abnormalities in the gross morphology of ttis450 larvae are first detectable at 72 hpf and became more severe with time. At 120 hpf, the wildtype (WT) intestinal epithelium exhibits a columnar morphology and starts to elaborate folds; in contrast, the intestinal epithelium in ttis450 remains thin and unfolded (Figure 1A and 1B). ttis450 larvae also exhibit smaller eyes (microphthalmia), a smaller, misshapen head, an uninflated swim bladder and impaired yolk absorption (Figure 1A). At 120 hpf, the ttis450 pancreas and liver are both substantially smaller than in WT (Figure 1C).
By 120 hpf, the rostral intestine (intestinal bulb region) in ttis450 larvae is markedly smaller than in WT and the intestinal epithelial cells (IECs) are cuboidal rather than columnar in shape (Figure 1C, 1D). The intestinal lumen appears clear of cellular debris. Cells in the mid and posterior intestine are also smaller and less polarized than in WT (Figure 1D). The mean apicobasal height of the cells in the intestinal bulb region of ttis450 larvae is approximately 40% less than that in WT (Figure 1E). However, cellular differentiation is not inhibited as similar numbers of mucin-producing goblet cells are found in the mid-intestinal region of ttis450 larvae as in WT (Figure 1D).
The reduction in cell size is accompanied by changes in the proportion of cells in different phases of the cell cycle. At 72 hpf, the intestinal epithelium is the most rapidly proliferating tissue in the zebrafish embryo [28], [29]. Using BrdU incorporation analysis, we detected fewer ttis450 IECs in S phase than WT IECs (Figure S1A, S1B). Fluorescent activated cell sorting (FACS) of cells disaggregated from WT and ttis450 larvae carrying the gutGFP transgene allowed us to analyze the proliferation of cells derived specifically from the liver, pancreas and intestine. We observed a significant accumulation of ttis450 cells in the G1 phase of the cell cycle at 96 hpf (88% in ttis450 compared to 70% in WT) and a corresponding reduction of ttis450 cells in S phase (8% in ttis450 compared to 28% in WT). No significant difference in the number of cells in G2 was observed (Figure 1F).
The ttis450 phenotype is completely penetrant, and the animals die at 8–9 days post-fertilization (dpf). Heterozygous ttis450 carriers are phenotypically indistinguishable from WT siblings.
We identified the mutated gene responsible for the abnormal digestive organ development in ttis450 by mapping the ttis450 locus to a 260-kilobase interval on chromosome 1 encompassing 5 genes (Figure 2A). One of these genes, pwp2h, comprises 21 exons spanning 2928 base pairs (Figure 2B) and encodes a protein of 937 amino acids containing 13 WD-40 repeat domains. WD-40 repeats generally serve as platforms for the assembly of proteins in multi-protein complexes and are conserved from yeast to mammals. We identified an A to T base change in the conserved splice acceptor site in intron 9 of pwp2h in ttis450 mutants (Figure 2C) resulting in utilization of a cryptic splice site 11 bp upstream of exon 10, thereby generating a frame-shift and nonsense mutation in codon 421 (Figure S2A) and truncating the Pwp2h protein in the seventh WD domain (Figure S3).
The tti phenotype is recapitulated by microinjection of 1–4 cell zebrafish embryos with an antisense morpholino oligonucleotide targeted to pwp2h mRNA (Figure S2B, S2C). That mutant pwp2h is responsible for the ttis450 phenotype was confirmed by non-complementation with an independent allele of pwp2h, ttis927 (Figure S2D–S2G). ttis927 was identified in an ENU mutagenesis screen (the 2-CLIP screen) [30] conducted on the (ins:dsRed)m1081;Tg(fabp10:dsRed;ela3l:GFP)gz12 transgenic background [31] to facilitate assessment of pancreas and liver development. ttis927 harbours a missense mutation in pwp2h: a T to A transversion in exon 5 (Figure S2H) resulting in the replacement of a valine with glutamic acid (Figure S2I) in the second WD-40 domain (Figure S3). The phenotypes of ttis450 and ttis927 larvae are essentially indistinguishable.
In order to assess the expression pattern of pwp2h during zebrafish embryogenesis, we performed wholemount in situ hybridization (WISH). In WT embryos pwp2h mRNA is ubiquitously expressed between 4–12 hpf and then becomes restricted to the brain and eyes at 24 hpf (Figure 2D–2G). By 48 hpf pwp2h mRNA is expressed in the pharyngeal cartilages and primitive gut, including the liver and pancreas anlagen (Figure 2H). By 72 hpf expression in the eye is largely extinguished and restricted to the pharyngeal cartilages, liver, intestine and pancreas (Figure 2I). By 96 hpf, pwp2h expression in the intestine is diminishing but is sustained in the pharyngeal cartilages, liver and pancreas (Figure 2J). By 120–144 hpf, the pancreas is the only tissue in which pwp2h mRNA is detected (Figure 2K, 2L). Expression of pwp2h is absent in ttis450 embryos from 24 hpf onwards (Figure 2M, 2N) indicating that upon exhaustion of maternally deposited supplies of WT pwp2h mRNA, the zygotically expressed mutant mRNA probably undergoes nonsense-mediated decay (NMD). These expression data are consistent with the eye, brain, pharyngeal cartilages and digestive organs being the most severely affected organs in ttis450 larvae.
In all species, rRNA is transcribed as a large pre-rRNA transcript which undergoes a series of enzymatic cleavage steps within the nucleolus by large ribonucleoprotein complexes to produce mature 18S, 28S and 5.8S rRNAs (Figure 3B). To investigate rRNA processing in ttis450 larvae, we conducted Northern blot analysis (Figure 3A) using probes designed to hybridize to the external (5′ETS) and internal-transcribed (ITS1 and ITS2) spacer regions of zebrafish 45S pre-rRNA (Figure 3B). These probes detect the full-length rRNA precursor and all intermediate species but not the fully mature forms of rRNA. This analysis revealed a 2.5 fold accumulation of the full-length precursor ‘a’ in ttis450 and an accumulation of the intermediates ‘b’ and ‘c’ (4.6 fold and 1.3 fold, respectively). These observations are consistent with a block in the processing of the full-length rRNA precursor. We also noted a 2.6 fold decrease in ttis450 larvae in the level of ‘d’, the immediate precursor of 18S rRNA (Figure 3A). Furthermore, E-bioanalyser analysis revealed a marked reduction in the production of mature 18S rRNA in ttis450 larvae (Figure 3C); however, the production of mature 28S rRNA was unaffected (Figure 3C). These changes altered the ratio of 28S/18S rRNA in ttis450 larvae, which is 2.8 at 120 hpf, compared to 1.8 in WT (Figure 3D).
To investigate the impact of Pwp2h deficiency on ribosome formation, we prepared extracts of WT and tti zebrafish larvae at 96 hpf and fractionated the ribosomal subunits on sucrose density gradients (Figure 3E). The areas under the peaks corresponding to the 40S subunits and 80S monosomes in ttis450 lysates are markedly smaller compared to those in WT (reduced approximately 4 fold and 2-fold, respectively). Meanwhile, the area under the peak corresponding to the 60S subunits is increased by approximately 4.5 fold (Figure 3F). Collectively, these data are consistent with Pwp2h deficiency primarily impacting on 40S subunit formation.
To determine the impact of impaired ribosome biogenesis at the ultrastructural level, we used transmission electron microscopy (TEM) (Figure 4A–4H). While WT intestinal epithelium is folded and the cells exhibit apicobasal polarity and a highly elaborated apical brush border (Figure 4A, 4C, 4E, 4G), IECs in ttis450 are smaller and the microvilli are shorter and relatively sparse (Figure 4B, 4D, 4F, 4H). The ttis450 nuclei contain prominent condensed nucleoli, suggesting ribosomal stress [32]. Also conspicuous at 96 hpf in the IECs of ttis450 larvae, but essentially absent in WT, are cytoplasmic vesicles containing debris (Figure 4B, 4B′). At 120 hpf, these structures are bigger in size and electron dense (Figure 4D, 4D′). At 144 hpf, vesicles more akin to those observed at 96 hpf are present (Figure 4H, 4H′, 4H″). Similar transient structures have been previously identified in cells undergoing autophagy. We therefore pursued the hypothesis that the cytoplasmic vesicles in ttis450 larvae correspond to autophagosomes and autolysosomes: vesicles that sequester and digest organelles.
Autophagy is a dynamic process comprising autophagosome synthesis, delivery of autophagic substrates to lysosomes and substrate degradation in autolysosomes [10], [12]. In order to investigate whether the electron dense vesicles observed at 120 hpf (Figure 4D) correspond to autolysosomes, we exposed WT and ttis450 larvae at 106 hpf for 14 h to chloroquine, an autophagy inhibitor that blocks the fusion of autophagosomes with lysosomes and thereby prevents digestion of the vesicle contents [33]. After chloroquine treatment few, if any, electron dense cytoplasmic vesicles (autolysosomes) are found in the intestinal epithelium of ttis450 larvae (Figure 4F). Instead, the IECs in ttis450 larvae contain vesicles more reminiscent of autophagosomes (Figure 4F, 4F′, 4F″). We counted >3 autophagosomes/cell (3.25±0.144, n = 60) in the IECs of ttis450 larvae, compared to <1 (0.6±0.058, n = 60) in WT IECs. Thus chloroquine inhibition of autophagic flux results in a significantly higher number of autophagosome-like structures in ttis450 larvae compared to WT.
To investigate this further, we examined LC3 localisation in WT and ttis450 larvae using wholemount immunocytochemistry (Figure 5A–5G). LC3, the mammalian orthologue of yeast Atg8, is a robust marker of autophagosomes. Upon induction of autophagy, the cytoplasmic form of LC3 (LC3I) is converted by cleavage and lipidation to a transient, autophagosomal membrane-bound form of LC3 (LC3II). Disrupting the fusion of autophagosomes with lysosomes with chloroquine prolongs the half-life of LC3II and facilitates the accumulation of LC3II-containing autophagosomes, which appear as punctate structures using LC3 immunocytochemistry. We observed more puncta in the IECs of chloroquine-treated WT larvae (Figure 5C) compared to untreated WT larvae (Figure 5A). Consistent with impaired ribosome biogenesis stimulating autophagy, we counted approximately 5 times more puncta in the IECs of chloroquine-treated ttis450 larvae (Figure 5D) compared to the IECs of chloroquine-treated WT siblings (Figure 5C; compare 2nd and 4th bars in Figure 5G). We next exposed WT and ttis450 larvae to rapamycin, which through its specific inhibition of Torc1 [34], [35] provides a powerful stimulus to autophagy in yeast, zebrafish and mice. We found that the number of puncta in WT larvae treated with rapamycin and chloroquine together (Figure 5E, 5G) was similar to the number of puncta in ttis450 larvae treated with chloroquine alone (Figure 5D, 5G). Finally, treating ttis450 larvae with rapamycin and chloroquine together (Figure 5F) resulted in more abundant puncta than in both chloroquine-treated ttis450 larvae and rapamycin and chloroquine-treated WT larvae (Figure 5G). Upon Western blot analysis of whole larval lysates (Figure 5H, 5I), we found that LC3II levels in chloroquine-treated ttis450 larvae were significantly higher than in chloroquine-treated WT larvae but not significantly different from those in WT larvae treated with rapamycin and chloroquine together (Figure 5I). Together these experiments demonstrate that the vesicles identified in the IECs of ttis450 larvae are autophagosomes, and, to the best of our knowledge, provide the first evidence for a link between impaired ribosome biogenesis and autophagy.
To determine the extent of autophagy in ttis450 larvae, we injected RNA encoding a mCherry-LC3 fusion protein into the yolk of 1–4 cell stage zebrafish embryos and evaluated the formation of puncta after prior treatment with chloroquine for 14 h at three time-points (Figure S4). At 72 hpf, abundant puncta are present in the eye (Figure S4B) and brain (Figure S4B′) of ttis450 larvae compared to WT larvae (Figure S4A, S4A′). At this time-point, there are very few puncta in the digestive organs (Figure S4C, S4D). A similar picture was observed at 96 hpf (data not shown). At 120 hpf, the number of puncta in the brain (Figure S4F′) in ttis450 larvae is now comparable to that observed in WT (Figure S4E′), while higher numbers of puncta are still found in the eye (Figure S4F). At 120 hpf there are more abundant puncta in the intestine and pancreas of ttis450 larvae (Figure S4H) compared to these organs in WT (Figure S4E and S4G, respectively). This pattern of autophagy induction mirrors the tempero-spatial expression of pwp2h during zebrafish development, and is consistent with these tissues being the most affected by impaired ribosome biogenesis in ttis450 larvae.
To determine whether autophagy is a specific response to impaired ribosome biogenesis, we conducted LC3 analysis of two additional zebrafish intestinal mutants, setebos (sets453) and caliban (clbns846), which exhibit phenotypes that are essentially indistinguishable from that of ttis450 when viewed under the light microscope or upon histological analysis. Whereas sets453 harbours a mutation in a gene which impairs 28S rRNA production and ribosome biogenesis (APB et al., in preparation), the mutation in clbns846 lies in a gene encoding an essential mRNA splicing factor (SJM et al., in preparation). We observed that sets453 larvae, like ttis450 larvae, contain higher LC3II levels compared to WT siblings in the presence of chloroquine (Figure S5A, S5B) and their IECs contain abundant autophagosome-like structures when analysed by TEM (data not shown). In contrast, the LC3II levels in clbns846 larvae are indistinguishable from those in WT siblings (Figure S5A, S5B) and the intestinal epithelium of clbns846 mutants do not contain autophagosomes or autolysosomes when inspected at the ultrastructural level (Figure S5C–S5H). These data suggest that the induction of autophagy in IECs is a specific response to impaired ribosome biogenesis, rather than a non-specific response to impaired cell growth.
We followed the morphological changes in the intestinal epithelium and liver of ttis450 larvae until 7 dpf, just before the larvae die at 8–9 dpf. At 7 dpf, the IECs are substantially smaller in ttis450 larvae than in their WT counterparts and neither ttis450 nor WT larvae contain detached cells in the intestinal lumen (Figure S6A–S6D). The ttis450 IECs no longer contain conspicuous autophagosomes, though electron dense vesicles are present in abundance in adjacent liver cells (Figure S6E–S6F). To investigate the impact of inhibiting autophagy in ttis450 larvae, we blocked autophagosome formation by injecting 1 ng of an antisense morpholino oligonucleotide (MO), which targets the translation start-site of atg5 mRNA [36], into 1–4 cell stage embryos derived from pair-wise matings of heterozygous ttis450 adults. At 72 hpf, uninjected, vehicle-injected and atg5 MO-injected ttis450 larvae were identified and subjected to LC3 analysis. We found significantly lower LC3II levels in the atg5 MO-injected ttis450 larvae compared to uninjected and vehicle-injected controls (Figure 6A). Moreover, from 72–120 hpf, we noticed that atg5 MO-injected ttis450 larvae start to develop oedema around the head, eye, heart and intestine (Figure S7D). As a consequence, 50% of atg5 MO-injected ttis450 larvae die by 5 dpf and all atg5 MO-injected ttis450 larvae are dead by 7 dpf (Figure 6B). This contrasts markedly with untreated or vehicle-injected ttis450 larvae, which survive until 8–9 dpf (Figure 6B). The longevity of WT larvae injected with the atg5 MO is not affected. Ultrastructural analysis at 120 hpf revealed detached, shrunken cells in the intestinal lumen of atg5 MO-treated tis450 larvae (Figure 6D–6F) that were never seen in the intestinal lumen of ttis450 larvae injected with vehicle or WT siblings injected with atg5 MO (Figure 6C). Together these data demonstrate that autophagy extends the lifespan of ttis450 larvae and prolongs the survival of IECs.
To explore the relationship between the Tor pathway and autophagy in ttis450 larvae, we analysed the levels of phosphorylated RPS6 (p-RPS6), a downstream target of Torc1 activity. Using Western blot analysis, we found that p-RPS6 levels decrease markedly in WT larvae between 72–120 hpf as previously reported [37] (Figure 7A, 7B). Somewhat surprisingly, p-RPS6 levels persist in ttis450 larvae until 120 hpf, when they are 4-fold higher than in WT siblings (Figure 7A, 7B). We also noticed that the overall level of RPS6 protein is less in ttis450 larvae compared to WT, perhaps reflecting the fact that RPS6 is a structural component of the 40S subunits, which are fewer in ttis450 larvae. Using immunocytochemistry we examined p-RPS6 expression in histological sections of WT and ttis450 larvae. At 96 hpf, we observed robust p-RPS6 expression in the intestinal epithelium and liver of WT and ttis450 larvae (Figure 7C). The high p-RPS6 levels in the ttis450 intestinal epithelium raise the possibility that elevated p-RPS6 stimulates autophagy directly in ttis450 larvae, as this occurrence has been recognised previously, including in the Drosophila fat body during starvation [38], [39]. To test this, we blocked p-RPS6 accumulation using rapamycin. We found that prior exposure to rapamycin for 14 h eliminated the p-RPS6 signal in both WT and ttis450 larvae at 96 hpf (Figure 7D), thereby unequivocally linking the persistent and elevated p-RPS6 signal in ttis450 larvae to Torc1 activity. Moreover, rapamycin treatment of ttis450 larvae in the presence and absence of chloroquine results in elevated levels of LC3II (Figure 7E) and LC3II-containing autophagosome formation (Figure 5F, 5G). These augmented levels of autophagy, achieved through rapamycin blockade of RPS6 phosphorylation, exclude the possibility that elevated p-RPS6 is responsible for the induction of autophagy in ttis450 larvae. Indeed, these data suggest that autophagy induction in ttis450 larvae is independent of the level of activation of the Tor pathway and the levels of p-RPS6.
We corroborated this finding with a genetic approach by crossing ttis450 onto the tsc2vu242/vu242 background [40]. Tsc2 is a negative regulator of Torc1 and tsc2vu242/vu242 zebrafish larvae exhibit a variety of defects including an enlarged liver at 7 dpf [40], consistent with Tor playing a positive role in digestive organ growth. The development of the ttis450 phenotype, including the induction of autophagy, is not perturbed on the tsc2vu242/vu242 background (Figure S8A–S8F). Interestingly, ttis450 larvae at 96 hpf contain higher levels of pRPS6 than tsc2vu242/vu242 larvae (Figure S8E, S8F) and the levels of p-RPS6 are higher still in compound ttis450;tsc2vu242/vu242 mutants (Figure S8E, S8F). In conclusion, these data show that impaired ribosome biogenesis induces autophagy in ttis450 larvae through a mechanism that does not require inhibition of the Tor pathway and is independent of p-RPS6 levels.
Defects in 18S and 28S rRNA processing have been shown to activate Tp53 [41], which in turn can stimulate autophagy [42]. While WT larvae contained negligible levels of Tp53 protein at 96 hpf, ttis450 larvae display readily detectable levels of Tp53 protein at this time-point (Figure 8A) and increased transcription of Tp53 target genes, including ΔN113p53, p21, cyclinG1 and mdm2 (Figure 8B–8E). To determine whether Tp53 plays a role in the induction of autophagy in ttis450, we generated ttis450 larvae expressing a mutant form of Tp53 (Tp53M214K) with negligible DNA-binding activity [43]. While this mutation severely diminished the elevated ΔN113p53, p21, cyclinG1 and mdm2 expression levels in ttis450 larvae at 96 hpf as expected (Figure 8B–8E), the level of LC3II in compound ttis450;tp53M214K/M214K mutants in the presence of chloroquine was significantly higher than in tp53M214K/M214K mutants (Figure 8F–8H). In addition, ultrastructural analysis revealed similar numbers of autolysosomes in ttis450 mutants at 120 hpf, independent of whether they were on the tp53M214K/M214K background or not (Figure 8H). Therefore the induction of autophagy in response to Pwp2h depletion proceeds unabated in ttis450 larvae that are devoid of functional Tp53 protein.
This study shows, in the context of an intact vertebrate organism, that Pwp2h is critical for the production of mature 18S rRNA, an integral component of the 40S ribosomal subunit. In zebrafish, as in yeast, Pwp2h depletion results in reduced levels of the immediate precursor to mature 18S rRNA and a concomitant decrease in the production of mature 18S rRNA and assembly of 40S ribosomal subunits. Thus the role of Pwp2h in the 90S pre-ribosomal particle or small subunit processome is conserved from yeast to vertebrates.
In our pwp2h-deficient model, titania (ttis450), the growth of the endodermal organs, eyes, brain and craniofacial structures is severely arrested and autophagy is markedly up-regulated. To the best of our knowledge, this is the first time that a link between impaired ribosome biogenesis and autophagy has been demonstrated. We further show that elevated rates of autophagy support the survival of intestinal epithelial cells and increase the lifespan of ttis450 larvae, thereby demonstrating that autophagy is a survival mechanism invoked in response to ribosomal stress. In our zebrafish model, autophagy induction does not depend on inhibition of the Tor pathway or activation of Tp53.
The death of ttis450 larvae at 8–9 dpf demonstrates that pwp2h encodes a protein that is indispensable for life. However, the development of ttis450 larvae until 72 hpf is supported by the deposition of maternal, wild-type pwp2h mRNA (and/or protein) into oocytes by their heterozygous mother. At 72 hpf, the tissues in which pwp2h is most highly expressed are the intestinal epithelium, pharyngeal arches, liver, dorsal midbrain, cerebellum, dorsal hindbrain, retinal epithelium and pancreas. These tissues are also the most rapidly proliferating tissues in WT larvae at 72 hpf [28] and the most severely affected tissues in ttis450 larvae. Thus the tissue-specific phenotype of ttis450 larvae may be explained by maternally-derived WT pwp2h mRNA being exhausted first in developing organs containing highly proliferative cells.
In WT zebrafish larvae there is a transient spike in Torc1 activity (as measured by p-RPS6) at around 72 hpf that is coincident with the activation of anabolic pathways required for cell growth and proliferation during the endoderm to intestine transition [37]. Torc1 is thought to play a role in developing organisms as an organ size checkpoint, potentiating growth signals that promote the rapid expansion of organs until they reach a genetically programmed cell size [44]. Therefore the persistent and robust activity of Torc1 we observe in the intestinal epithelium and liver of ttis450 larvae at 96 hpf may be a consequence of these organs being markedly smaller than their WT counterparts at this stage.
The gross phenotype of ttis450 is highly reminiscent of another zebrafish mutant, nil per os (npo), in which the morphogenesis of the intestinal epithelium is also arrested. In npo the failure of the primitive gut endoderm to transform into a monolayer of polarized and differentiated epithelium is caused by a mutation in rbm19, a gene encoding a protein with six RNA recognition motifs that is also thought to play a role in ribosome biogenesis [45]. The same authors showed that essentially the same hypoplastic intestinal phenotype was recapitulated by exposure of WT zebrafish larvae to the Torc1 inhibitor, rapamycin [46], which presumably stimulated autophagy. It would be interesting to determine whether the growth arrest of the digestive organs in the npo mutant is also accompanied by autophagy.
The degree of activation of the Tor pathway is thought to be one of the major factors governing autophagy. However, Tor inhibition is not the mechanism responsible for autophagy in ttis450 larvae and recent work suggests that autophagy regulation is a very complex process involving the integration of signals from many diverse signalling pathways [47]. Indeed, proteomic analysis of binding partners of components of the autophagy machinery suggests that several hundred molecules participate in the regulation of the human autophagy network [48]. While much recent attention has been focused on the direct phosphorylation of Ulk1/Atg1 by AMPK, acting either cooperatively or independently of Tor to exert autophagy control [19]–[21], there are many reports of other kinases capable of controlling autophagy by a variety of Tor-independent mechanisms [49]–[51]. The dissociation of the key BH3 domain-containing autophagy protein, Beclin 1 (mammalian orthologue of yeast Atg6) from its inhibitors Bcl2 and Bcl-XL as a result of phosphorylation of one or other components is also a critical determinant in the induction of autophagy [52]. In the case of ttis450 larvae, it is plausible that autophagy induction may involve a targeted pathway, selective for ribosomes [11], which by analogy with mitophagy [53], is invoked to digest damaged cargo such as non-functional organelles.
Somewhat surprisingly, we also ruled out involvement of Tp53 in the induction of autophagy in ttis450 larvae, even though Tp53 protein is active in ttis450 larvae at 96 hpf. However, we believe the increased expression of Tp53 target genes such as p21 and cyclinG1 may be responsible, at least in part, for the reduction in the number of cells in the S phase of the cell cycle we observed at this time-point. To explain this, we surmise that as ribosome biogenesis is progressively impaired, the ttis450 larvae mount a two-stage response to Pwp2h depletion. Initially, the cells undergo a Tp53-mediated cell cycle arrest. However, as the synthesis of new proteins, including Tp53 and its targets, is progressively impaired, the cells invoke autophagy to prolong their survival.
The notion of the existence of a second type of programmed cell death, distinct from apoptosis, which emanates from catastrophic levels of autophagy, is a hotly debated topic [54]. Using TEM, we did not see any evidence of cell death in the IECs of ttis450 larvae, even at 7–8 dpf just before the larvae die, affirming that the levels of autophagy induced in the IECs of ttis450 larvae prolong cell survival rather than trigger cell death. We proved this by disrupting the formation of the early autophagosome by inhibiting the translation of atg5 mRNA. This resulted in the death of IECs in ttis450 larvae and a markedly reduced lifespan.
As mentioned previously, ttis450 larvae exhibit impaired development of the craniofacial cartilages, exocrine pancreas and brain, tissues that are often clinically abnormal in patients with certain human ribosomopathies, including Diamond Blackfan anaemia and Schwachman Diamond syndrome [4]. Recently, two new zebrafish models of dyskeratosis congenita (DC) based on mutations in components of the H/ACA RNP complex were described [55], [56]. Like ttis450, these mutants display impaired production of 18S rRNA and induction of Tp53 target genes, consistent with previous studies demonstrating that defects in ribosome biogenesis induce Tp53 activation and cell cycle arrest [41]. Moreover, hematopoietic stem cells in these mutants were depleted via a Tp53-dependent mechanism, providing a plausible explanation for why DC patients are susceptible to bone marrow failure [55], [56]. In one of these mutants, the gut and craniofacial structures were also reported to be underdeveloped and, as observed in ttis450, these defects persisted on a Tp53 mutant background [55]. We speculate that the p53-independent features of this model of DC may be caused by elevated rates of autophagy. If so, and these findings are confirmed in human DC, it will be important to determine whether elevated autophagic activity contributes to prolonged cell survival prior to considering clinical interventions to limit this process.
There is currently a great deal of interest in the development of novel therapeutics that target the cancerous translation apparatus through the combined inhibition of ribosome biogenesis, translation initiation and translation elongation [5]. To avoid inadvertently prolonging cancer cell survival, these approaches could benefit from a detailed understanding of the mechanisms and cellular contexts that induce autophagy in response to ribosomal stress. While such insights may be forthcoming from studies performed on cell lines, it is likely that complementary experiments carried out in the context of an entire vertebrate organism, such as the zebrafish model introduced here, may also be fruitful.
All experimental procedures on zebrafish embryos and larvae were approved by the Ludwig Institute for Cancer Research/Department of Surgery - Royal Melbourne Hospital Animal Ethics Committee.
Zebrafish embryos were obtained from pair-wise matings of heterozygous ttis450, seteboss450 and calibans846 zebrafish on the Tg(XlEef1a1:GFP)s854 (gutGFP) background and from ttis450 heterozygotes carrying two mutant alleles of Tp53 (ttis450;Tp53M214K/M214K) [43] and raised at 28.5°C. ttis927 was propagated on the Tg(ins:dsRed)m1081;Tg(fabp10:dsRed;ela3l:GFP)gz12 (2-CLIP) background [31]. The Tp53M214K/M214K line (gift of Thomas Look and David Lane) and tsc2vu24 line were obtained through TILLING [40], [43]. The tsc2 and pwp2h loci in zebrafish are both on chromosome 1 so in order to generate sufficient ttis450;tsc2vu24 compound mutants for analysis, we identified and in-crossed recombinants harbouring the two mutations in a cis configuration. To prevent melanization and maintain transparency, embryos were treated with 0.003% 1-phenyl-2-thiourea (PTU; Sigma Aldrich) in embryo medium. Imaging of live larvae was carried out using a LeicaM2 FLIII microscope after anaesthetizing with 200 mg/L benzocaine (Sigma-Aldrich, St. Louis, MO) in embryo medium. All images were imported into CorelDRAWX4 (Corel Corporation, Ottawa, Ontario, Canada). Image manipulation was limited to levels, hue and saturation adjustments.
Histology was performed as described [27]. Mucins and other carbohydrates secreted by intestinal goblet cells were stained using alcian blue-periodic acid-Schiff reagent [27]. For WISH, larvae were processed as described [57], [58] To generate pwp2h riboprobes a cDNA template was amplified by RT-PCR. For primer sequences see Text S1. These were then transcribed using the digoxigenin DNA Labelling Kit (Roche Diagnostics) according to the manufacturer's instructions. Hybridized riboprobes were detected using an anti-DIG antibody conjugated to alkaline phosphatase according to the manufacturer's instructions (Roche Diagnostics). Larvae were imaged on a Nikon SMZ 1500 microscope.
100–200 WT and ttis450 larvae were rinsed in PBST (PBS containing 0.5% Tween 20) three times prior to incubating in 1 mL Hank's Balanced Salt Solution containing 0.25% trypsin, 0.1% EDTA, 40 µg/mL Proteinase K and 10 µg/mL collagenase for 30 min at 37°C. Larvae were then homogenised in 7 mL PBS containing 5% FBS. The cell suspension was strained through a 40 µM nylon cell strainer (BD Falcon) and spun at 2000 rpm for 10 min at 4°C. The pellet was washed twice with cold PBS/5% FBS and resuspended in 500 µl PBS. Ice-cold methanol (900 µl) was added to the pellet and cells were left on ice for 1 h prior to centrifugation as above. The pellet was resuspended in 0.5 mL PBS containing 40 µg/mL propidium iodide and 0.5 mg/mL RNaseA for 30–60 min at room temperature (RT). GFP positive cells were sorted on a FACSCalibur™ Optics instrument (Benton Dickinson) and analysis was performed using the ModFit LT program.
To identify cells in the S-phase of the cell cycle, the incorporation of bromodeoxyuridine (BrdU) by live larvae was analysed as described [27]. To measure cell height, images of sagittal histological sections were captured on a Nikon Eclipse 80i microscope and then analysed using MetaMorph Microscopy Automation & Image Analysis Software.
For genetic mapping, ttis450 heterozygotes on the gutGFP background were crossed onto the polymorphic WIK strain. Mutant larvae were identified by craniofacial and intestinal defects visible at 96 hpf under brightfield and fluorescence illumination. Subsequent mapping was performed as described [28].
Protein sequence alignment of Pwp2h from zebrafish, yeast, mouse and human was performed using the clustalW2 program with default parameters. WD domains were identified using the Simple Modular Architecture Research Tool (SMART) software.
A novel EcoN1 restriction enzyme site created by the ttis450 mutation produced a restriction fragment length polymorphism (RFLP) that was exploited for genotyping. Primers were used to amplify a 653-base pair (bp) fragment spanning exons 9 to 11 containing the ttis450 mutation. For primer sequences see Text S1.
Total cellular RNA was prepared from WT and ttis450 larvae (120 hpf) by homogenizing 20–50 larvae in Solution D (4.2 M guanidinium thiocyanate, 25 mM NaCitrate, 30% Sarkosyl BDH NL30) as described [59]. Northern blot analysis was conducted on 2 µg samples using α-32P-labelled probes designed to hybridize to zebrafish 5′ETS, ITS1 and ITS2 sequences, which were PCR-amplified from genomic DNA using previously described primers [60]. Radioactive signals were detected using a Phosphorimager and Storm 820 scanner (Amersham Biosciences) and analysed using ImageQuant TL software.
Solutions of total RNA extracted from WT and ttis450 larvae were analysed on an Agilent 2100 E-Bioanalyser according to the manufacturer's instructions.
50–100 WT and ttis450 larvae at 96 hpf were resuspended in cold lysis buffer (50 mM Tris-HCl pH 7.4, 150 mM KCl, 2.5 mM MgCL2, 1% Triton X-100, 0.5% sodium deoxycholate, 3 mM DTT) containing 120 U/mL RNase inhibitor (Invitrogen) and Complete Protease Inhibitor Cocktail (Roche) and sheared through a 23G needle. Lysates were incubated on ice for 30 min and centrifuged (12,000 rpm, 20 min at 4°C) to pellet nuclei and cellular debris. Cytoplasmic extract (2 mg) was loaded onto a continuous low salt (80 mM NaCl) 3.1–30.1% (w/v) sucrose gradient (14 mL) [61] generated using an ISCO gradient maker. Samples were separated by centrifugation using a SW41 rotor at 40000 rpm for 4 h at 4°C, and fractionated (1 mL) using a Foxy Jr fraction collector. Absorbance at 260 nM was determined with an ISCO UA-6 absorbance detector. In each case, quantitation of 40S, 60S, and 80S was performed by measuring the area under the relevant peak using Metamorph Image Analysis Software.
For TEM, larvae were fixed in 2.5% glutaraldehyde, 2% paraformaldehyde (Electron Microscopy Sciences, Hatfield, PA) in PBS for 2 h at R.T, rinsed in 0.08 M Sorensen's Phosphate buffer pH 7.4 and then stored in 0.08 M Sorensen's buffer with 5% sucrose. Post-fixation was with 2% osmium tetroxide in PBS followed by dehydration through a graded series of alcohols, 2 acetone rinses and embedding in Spurrs resin [62]. Sections approximately 80 nm thick were cut with a diamond knife (Diatome, Switzerland) on a Ultracut-S ultramicrotome (Leica, Mannheim, Germany) and contrasted with uranyl acetate and lead citrate. Images were captured with a Megaview II cooled CCD camera (Soft Imaging Solutions, Olympus, Australia) in a JEOL 1011 TEM. Transverse sections were obtained through the anterior intestinal region known as the intestinal bulb.
For transverse sections, embryos were fixed in 2% paraformaldehyde overnight at 4°C, embedded vertically in 4% low melting temperature agarose (Cambrex BioScience, East Rutherford, NJ) in disposable cryomolds (Sakura Finetek, Torrance, CA), and sectioned at 200 µm intervals using a Leica (Solms, Germany) VT1000S vibrating microtome. Floating sections were transferred to the wells of a 24-well plate containing PBD (PBS containing 0.1% Tween-20 and 0.5% Triton-X) and then replaced with antibody blocking solution (PBD containing 1% (w/v) BSA and 1% (v/v) FCS) for 2 h at RT. The blocking solution was removed and the sections incubated with LC3B primary antibody diluted to 1∶500 in PBD containing 0.2% (w/v) BSA at 4°C overnight. The sections were rinsed three times in PBST (PBS containing 0.1% Tween-20) for 20 min at RT, followed by antibody blocking solution for 2 h at RT. The sections were then incubated overnight at 4°C in PBD containing 0.2% (w/v) BSA, Alexa Fluor 488 (1∶500), rhodamine-phalloidin (1∶150; Biotium, Hayward, CA) and 5 µg/mL Hoechst33342 (Sigma Aldrich). Sections were rinsed three times in PBST for 20 min at RT prior to imaging on an Olympus FV1000 scanning confocal microscope. Enumeration of LC3 puncta was performed using Metamorph. Details of antibodies and stains are available in Text S1.
Larvae were lysed (2 µL per embryo) in cold RIPA cell lysis buffer (50 mM Tris-HCl pH 7.4, 150 mM NaCl, 2 mM EDTA, 1% NP-40, 0.1% SDS) containing Complete Protease Inhibitor Cocktail (Roche) and sheared through a 23G needle. Lysates were incubated on ice for 30 min and then centrifuged for 20 min at 13,000 rpm at 4°C to pellet nuclei and cellular debris. Samples containing 40–80 µg of protein were heated to 95°C for 5 min with 5 X Protein Loading Dye (0.03 M Tris-HCl, pH 6.8, 13.8% glycerol, 1% SDS, 0.05% bromophenol blue, 2.7% β-mercaptoethanol) and loaded onto a 12% polyacrylamide gel. The proteins were transferred to PVDF membranes using an iBlot Gel Transfer Device (Invitrogen) according to the manufacturer's instructions. For RPS6, p-RPS6, LC3 and Actin, subsequent blocking, antibody incubation and membrane exposure were performed using the Odyssey system (LI-COR Biosciences). For Tp53, blocking and antibody incubation were performed in PBST/5% skim milk powder and membranes developed using the SuperSignal West Femto Chemilluminescent Substrate (Thermo Scientific). Signals were quantitated by densitometry and expressed as relative levels by reference to the level in untreated WT larvae, which was set at 1. Details of antibodies are provided in Text S1.
DNA encoding the fluorophore mCherry fused to the N terminus of LC3 was PCR amplified and transcribed into mRNA using the mMessage mMachine SP6 kit (Ambion Life Technologies, Mulgrave, Australia). For primer sequences see Text S1. mRNA (400 pg) was injected into the yolk of 1–4 cell stage embryos and exposed to 2.5 µM chloroquine (Fluka Sigma-Aldrich, Sydney, Australia) in embryo medium for 14 h at various time-points during development prior to mounting in 1.5% low melting point agarose for imaging with an Olympus FV1000 scanning confocal microscope.
Live WT, ttis450, sets453 clbns846 larvae were exposed to 2.5 µM chloroquine and/or 10 µM rapamycin in embryo medium at 28°C. Larvae were collected 14 h later for protein extraction and Western blot analysis of LC3II levels as described above.
Antisense morpholino oligonucleotides (MOs) targeted to the initiation of translation codons of pwp2h or atg5 mRNA were injected into the yolk of 1–4 cell stage WT or ttis450 embryos. 2 nL of MO at a concentration of 120×10−15 mol (total = 1 ng) and 180×10−14 mol (total = 15 ng) were used to knockdown atg5 and pwp2h mRNA translation, respectively. For MO sequences see Text S1.
Using immunocytochemical analysis, LC3II-containing autophagosomes were identified as puncta in thick transverse sections of ttis450 larvae. Puncta in 20 cells in 3 independent sections were counted using Metamorph. For TEM sections, the numbers of autophagosome-like structures in 20 cells in 3 independent sections were counted manually.
cDNA was reverse transcribed from total RNA (1–2 µg) extracted from WT and ttis450 larvae at 96 hpf using the Superscript III First Strand Synthesis System (Invitrogen) according to manufacturer's instructions. qRT-PCR was performed using the SensiMix SYBR Kit (Bioline) according to manufacturer's instructions. For primer sequences see Text S1.
Student's t-test was used to compare the means of two populations in Graphpad Prism 5.0. Error bars represent the mean +/− standard deviation (n≥3). A P value<0.05 was used to define statistical significance.
|
10.1371/journal.pcbi.1002263 | Computational Modeling and Analysis of Insulin Induced Eukaryotic Translation Initiation | Insulin, the primary hormone regulating the level of glucose in the bloodstream, modulates a variety of cellular and enzymatic processes in normal and diseased cells. Insulin signals are processed by a complex network of biochemical interactions which ultimately induce gene expression programs or other processes such as translation initiation. Surprisingly, despite the wealth of literature on insulin signaling, the relative importance of the components linking insulin with translation initiation remains unclear. We addressed this question by developing and interrogating a family of mathematical models of insulin induced translation initiation. The insulin network was modeled using mass-action kinetics within an ordinary differential equation (ODE) framework. A family of model parameters was estimated, starting from an initial best fit parameter set, using 24 experimental data sets taken from literature. The residual between model simulations and each of the experimental constraints were simultaneously minimized using multiobjective optimization. Interrogation of the model population, using sensitivity and robustness analysis, identified an insulin-dependent switch that controlled translation initiation. Our analysis suggested that without insulin, a balance between the pro-initiation activity of the GTP-binding protein Rheb and anti-initiation activity of PTEN controlled basal initiation. On the other hand, in the presence of insulin a combination of PI3K and Rheb activity controlled inducible initiation, where PI3K was only critical in the presence of insulin. Other well known regulatory mechanisms governing insulin action, for example IRS-1 negative feedback, modulated the relative importance of PI3K and Rheb but did not fundamentally change the signal flow.
| Insulin is a hormone produced by the body that regulates uptake of glucose from the bloodstream. The cellular response to insulin is governed by a complex network of intracellular interactions that ultimately influence cell growth and metabolism. Because of its central role in physiology, insulin signaling has been extensively studied. Yet despite this wealth of research, the relative importance of components in insulin signaling remains unclear. Mechanistic computer simulations have been shown to provide insight into the function of complex systems, such as insulin signaling. In this work we constructed and interrogated a mathematical computer simulation of insulin signaling to better understand the important components of the insulin signaling network. We determined the most important network components and identified network perturbations that can induce dramatic shifts in cellular phenotype. Our results offer an in-depth analysis of the insulin signaling pathway and provide a unique paradigm towards understanding how malfunctions in insulin signaling can result in numerous disease states.
| Insulin, the primary hormone regulating the level of glucose in the bloodstream, modulates a variety of cellular and enzymatic processes in normal and diseased cells [1]–[7]. The regulation of cellular function by insulin and insulin-like growth factors I/II (IGF-I/II) is a highly complex process [8]–[14]. Insulin and IGF-I/II interact with insulin receptors (IR), and type I/II IGF receptors (IGF-IR/IIR) in addition to other transmembrane receptors [10]. These interactions ultimately induce gene expression programs or other processes such as translation initiation. Translation rates of many cell cycle and survival proteins are modulated by growth factor, hormone or other mitogenic signals [15]. Insulin induces the activation of class I Phosphoinositide 3-kinases (PI3Ks), which in turn activate the serine/threonine protein kinase Akt and the mammalian target of rapamycin (mTOR). The PI3K/Akt/mTOR signaling axis is important to a variety of cellular programs, including apoptosis [16], cell size control [17] and translation initiation. Among other functions, activation of the PI3K/Akt/mTOR axis results in the phosphorylation of eukaryotic translation initiation factor 4E-binding protein (4E-BPx) family members [18]. Phosphorylation of 4E-BPx causes the release of the eukaryotic translation initiation factor 4E (eIF4E), which is critical to directing ribosomes to the 7-methyl-guanosine cap of eukaryotic mRNAs. Previously, the availability of eIF4E has been shown to be rate limiting for translation initiation in many eukaryotic cell-lines [15], [19]. Given its central role in cell biology, evolutionarily optimized infrastructure like translation might be expected to be robust or highly redundant. Surprisingly, deregulated translation, especially involving growth-factor or insulin induced initiation mechanisms, has been implicated in a spectrum of cancers [20].
Despite the wealth of literature on insulin signaling, the relative importance of the components linking insulin with translation initiation remains unclear. Many investigators have explored this question using both experimental and computational tools. For example, Caron et al. recently published a comprehensive map of the mTOR signaling network, including a detailed portrait of insulin induced mTOR activation and its downstream role in translation initiation [21]. Taniguchi et al. proposed three criteria to identify the critical nodes of insulin signaling: network divergence, degree of regulation and potential crosstalk [10]. Using these criteria, they identified insulin-receptor (IR), PI3K and Akt as the critical nodes of insulin action. Several insightful mathematical models of insulin-signaling have also been published [22]–[25]. While these models vary in their focus and biological scope, none has exclusively focused on how insulin stimulates translation initiation. This particular question was addressed by Nayak et al., who analyzed a family of detailed mathematical models of growth factor and insulin induced translation initiation [26]. Like the Taniguchi et al. hypothesis, their study suggested that Akt/mTOR were structurally fragile, and likely the key elements integrating growth factor signaling with translation. However, the Nayak et al. model neglected several key features of insulin processing, e.g., negative feedback of IR resulting from mTOR activity.
The objective of this study was to rank-order the importance of components of insulin-induced translation initiation using computational tools. Toward this objective, we analyzed an ensemble of mechanistic mathematical models of insulin induced translation initiation that was a significant extension of our previous work [26]. First, we expanded the original model connectivity to include a detailed description of the regulation and activity of insulin, insulin-like growth factor and platelet-derived growth factor (PDGF) receptor family members (including negative feedback). Second, we refined the description of the phosphorylation state of Akt and its downstream role in the activation of the mTORC1 and mTORC2 complexes. Lastly, we used new model estimation and interrogation techniques to generate and analyze an uncorrelated population of initiation models that were simultaneously consistent with 24 qualitative and quantitative data sets. Interrogation of this model population, using sensitivity and robustness analysis, identified an insulin-dependent switch that controlled translation initiation. Without insulin, a balance between the pro-initiation activity of the GTP-binding protein Rheb and anti-initiation activity of PTEN controlled basal initiation. Rheb knockdown simulations confirmed decreased initiation in the majority of the model population, while translation initiation increased for all models in the population following a PTEN deletion. On the other hand, a combination of PI3K and Rheb activity controlled insulin inducible initiation. PI3K deletion in the presence of insulin removed the ability of the network to process insulin signals, but did not remove initiation altogether. PI3K deletion reduced initiation to approximately 60% of its maximum level. Interestingly, the relative contribution of PI3K versus Rheb to the overall initiation level could be tuned by controlling IRS-1 feedback. In the absence of feedback, PI3K was more important than Rheb to signal propagation, while the opposite was true in the presence of feedback. Taken together, our modeling study supported the Taniguchi et al. hypothesis that PI3K was a critical node in the insulin-induced initiation network. However, we also found that the role of PI3K was nuanced; PI3K in combination with Rheb controlled initiation in the presence of insulin, while the combination of PTEN and Rheb controlled basal initiation.
The translation initiation model consisted of 250 protein, lipid or mRNA species interconnected by 573 interactions (Fig. 1). The model described the integration of insulin and growth-factor signaling with 80S assembly. While other eukaryotic translation initiation mechanisms exist, we focused only on cap-mediated translation as the dominant translation mechanism [27]. The model interactome was taken from literature (SBML file available in the supplemental materials Protocol S1); the connectivity of insulin- and growth-factor induced translation initiation has been extensively studied [14], [28]. The model interactome was not specific to a single cell line. Rather, it was a canonical representation of the pathways involved in insulin and growth-factor induced initiation. Using a canonical network allowed us to explore general features of insulin or growth-factor induced translation initiation without cell line specific artifacts. Binding of insulin or IGF-I/II with IR or IGF-I/IIR promotes the autophosphorylation of the cytosolic domains of these receptors at tyrosine residues. Receptor autophosphorylation promotes the formation of adaptor complexes, which are anchored in place by insulin receptor substrate (IRSx) family members; IRSx are required for the assembly of adaptor complexes involving the SHC-transforming protein 1 (Shc), Son of Sevenless (SoS), growth factor receptor-bound protein 2 (Grb2) and Ras proteins [29]–[31]. In the model we considered only the IRS-1 protein and neglected other IRSx family members. Adaptor complex formation ultimately culminates in the activation of the catalytic subunit of PI3K. Among their many roles, PI3Ks catalyze the phosphorylation of the phospholipid PIP2 to PIP3 [6]. PIP3 is critical to the localization of 3-phosphoinositide-dependent kinase 1 (PDK1) to the membrane, where it phosphorylates the master kinase Akt at Thr308 [32]. Akt is further phosphorylated at Ser473 by the rictor-mammalian target of rapamycin (mTORC2) protein [33]. Once phosphorylated, Akt promotes translation initiation by directly or indirectly activating the mTORC1 protein [1]. Akt directly activates mTORC1 through a novel binding partner known as PRAS40 [34], [35]. However, mTORC1 can also be activated by the GTP bound form of the Ras homologue enriched in brain (Rheb) protein. Without insulin, Rheb is regulated by the tuberous sclerosis complex TSC1/2, which has GTPase activating protein (GAP) activity. Akt directly phosphorylates TSC1/2 which inhibits its GAP activity and allows Rheb-mediated activation of mTORC1 [36], [37]. Activated mTORC1 plays two key roles in translation initiation; first, it activates ribosomal protein S6 kinase beta-1 (S6K1) and second it phosphorylates eukaryotic translation initiation factor 4E-binding protein (4E-BPx) family members [38]. In this study, we included only 4E-BP1 and modeled a single deactivating phosphorylation site. Phosphorylated 4E-BP1 releases eIF4E which, along with other initiation factors, is critical to directing ribosomes to the 7-methyl-guanosine cap structure of eukaryotic mRNAs [28].
Several mechanisms attenuate insulin and growth-factor induced translation initiation. First, insulin signal propagation can be controlled by disrupting adaptor complex formation. For example, we included tyrosine phosphatases and competitive inhibitors such as protein-tyrosine phosphatase 1B (PTP1B), src homology phosphotyrosyl phosphatase 2 (SHP2), growth factor receptor-bound protein 10 (Grb10) and suppressor of cytokine signaling 1/3 (SOCS1/3) which interfere with adaptor complex formation and activity [10], [39]–[41]. Second, several mechanisms control PIP3 formation, PDK1 recruitment and Akt phosphorylation [10]. In the model, we included the phosphatase and tensin homolog (PTEN) protein, which dephosphorylates PIP3 [42], as well as the SH2 (Src homology 2)-containing inositol phosphatase-1 (SHIP1) protein which hydrolyses the 5-phosphates from PIP3 [43]. Lastly, S6K1 inhibits IRS-1 activity by phosphorylation at Ser318 [44]. S6K1/IRS-1 feedback has been shown to be important in insulin resistance and cancer [14], [45]–[47].
Translation initiation was modeled using mass-action kinetics within an ordinary differential equation (ODE) framework. ODEs and mass-action kinetics are common methods of modeling biological pathways [48]–[50]. However, ODEs have several important limitations that could be addressed with other model formulations e.g., Partial Differential Equation (PDE) based models. PDEs naturally describe spatially distributed intracellular processes or can be used to model population dynamics using population balance methods [51]. However, the computational burden associated with solving and analyzing systems of PDEs, especially at the scale of the current study, would be substantial. Alternatively, we have addressed both of these ODE shortcomings (without resorting to a PDE formulation) by including well-mixed compartments to account for spatially localized species and processes and have considered an ensemble of models in our analysis to coarse-grain population phenomena. Irregardless of whether we have an ODE or PDE model formulation, both classes of model typically require the identification of a large number of unknown model parameters. The initiation model had 823 unknown parameters (573 kinetic parameters and 250 initial conditions), which were not uniquely identifiable (data not shown). We estimated an experimentally constrained population of parameters using multiobjective optimization. Model parameters were estimated, starting from an initial best fit parameter set, using 24 in vitro and in vivo data sets taken from literature (Table 1). These training data were taken from multiple independent studies (in different cell lines) exploring insulin and IGF-I/II signaling or in-vitro translation initiation. These data were largely western blot measurements of the total or phospho-specific abundance of proteins following the addition of a stimulus or inhibitor. While the use of multiple cell-lines was not ideal, it did allow us to capture a consensus picture of insulin or IGF-I/II initiated signaling (which was useful in understanding the general operational principles of the network). However, one should be careful when applying consensus models to specific cell lines or tissues, as these generally may behave qualitatively differently.
The residual between model simulations and each of the experimental constraints was simultaneously minimized using the multiobjective POETs algorithm [52]. We used a leave-three-out cross validation strategy to independently estimate prediction and training error during parameter identification (Table 1). Additionally, a random control (100 random parameter sets) was run to check the training/prediction fitness above random (Table 1). The training error for 23 of the 24 objectives was statistically significantly better than the random control at a 95% confidence level. Additionally, for 20 of the 24 objectives, the model prediction error was also significantly better than the random control (p0.05). Of the four remaining objectives (O4,O5,O12 and O13), three involved phosphorylated Akt (O4 and O12) or IRS-1 (O13), each of which had redundant measurements in the objective set that were significant. While the remaining objective, which involved IRS-1 levels (O5), was not significantly better than the random control, the absolute error was small.
The ensemble of translation models recapitulated diverse training data across multiple cell lines. POETs generated 18,886 probable models with Pareto rank 4. Model parameters had coefficients of variation (CV) ranging from 0.65 to 1.10. Further, 89% (512 of 573) of the model parameters were constrained with a CV 1. The performance of 5,818 rank-zero models is shown in Fig. 2. The majority of objective functions were uncorrelated e.g., O4O13 or O12O13 or directly proportional e.g., O3O11 or O9O15. Uncorrelated or proportional objectives suggested the model population simultaneously described each training constraint. However, several other objectives were inversely proportional e.g., O12O14. For these pairs, the model was unable to simultaneously fit both training data sets. Surprisingly, these objectives were the same protein pAkt(Thr308) O9O12 and pS6K1(Thr389) O3O14, taken from either different cell lines or different labs. This suggested conflicts in the data e.g., cell line variation or differences in specific laboratory protocols, rather than structural inaccuracies in the model, were responsible for the inverse relationship. The key indicators of eukaryotic translation initiation are the phosphorylation of S6K1 and 4E-BP1 [38]. Both Tzatos et al. and Villalonga et al. performed insightful studies exploring the dynamics of S6K1 and 4E-BP1 phosphorylation in L6 Myotubes and RhoE 3T3 cells [53], [54]. The ensemble recapitulated these observations with error distributions that were statistically significantly better than random parameters (, ; , ) (Fig. 3A and 3B, Table 1). The model population also recapitulated IGF1 induced Akt and S6K1 phosphorylation (, ; , ) (Fig. 3E and 3F, Table 1). Lorsh et al. studied ribosomal assembly dynamics in rabbit reticulocytes, suggesting the formation of the eIF2∶GTP∶Met-tRNA tertiary complex was rate limiting in 80S formation [55]. Our model captured 80S assembly dynamics, including the crucial lag phase in the first two minutes of stimulation (, ) (Fig. 3C, Table 1). Inhibitor data was also used for model training. Without insulin, PI3K was not activated and pAkt (Ser473) levels remained low (Fig. 3D, lane 1). Following insulin stimulation, PI3K activation resulted in increased pAkt(Ser473) levels (Fig. 3D, lane 2). Wortmannin, a PI3K inhibitor, significantly decreased pAkt(Ser473) (Fig. 3D, lane 3). While our model population qualitatively captured this decrease, the levels of pAkt(Ser473) were higher than those observed experimentally. The model was not trained using mTORC1/2 measurements, however species immediately upstream and downstream of mTORC1/2, namely pAkt(Ser473) or S6K1 were used in model training. Without insulin, pAkt(Ser473) and S6K1(Thr421/Ser424) levels were low (Fig. 3E/F, lanes 1). Addition of insulin increased pAkt(Ser473) and S6K1(Thr421/Ser424). Upon rapamycin addition, mTORC1 was inhibited and the levels of phosphorylated S6K1 decreased (Fig. 3E, lane 3). However, because of its position upstream of mTORC1, pAkt(Set473) levels were unchanged (Fig. 3E, lane 3).
The model was validated by comparing simulations with in vivo and in vitro data sets not used for training or cross-validation (Table 2). For four of the five prediction data sets, the model demonstrated errors statistically significantly better than a random control (p0.05). However, the remaining prediction case (P3), while not significantly different than random, has a small error relative to the other objectives. Data from Lorsh et al. was used to validate the dynamics of intermediate ribosomal complexes [55]. The level of 43S mRNA was quantified using both GTP and a non-degradable GTP-like homologue GMP-PNP (Fig. 4A). Data involving GMP-PNP was used for training while data involving GTP was used only for validation (, ). Garami et al. explored insulin-induced Rheb activation and the role of TSC1/2 in the presence and absence of wortmannin and rapamycin [56]. We first compared measured versus simulated Rheb-GTP levels, with and without insulin, in the absence of inhibitors. While we captured the qualitative trends, we over-predicted the percentage of GTP bound Rheb (, ) (Fig. 4B). The model also failed to predict sustained Rheb-GTP levels in the presence of rapamycin. This suggested that sustained pAkt(Ser473) levels (observed in Fig. 3E) were not correlated with increased Rheb-GTP activity. Garami et al. also measured the levels of GTP bound Rheb in both wild-type and TSC2 knockout cells. Because of TSC2's regulatory role, a TSC2 knockout significantly increased Rheb-GTP levels (, ) (Fig. 4C). Lastly, the model predicted the levels of 4E-BP1 bound eIF4E in response to heat shock (, ) (Fig. 4D) [57]. Because the model was not trained on stress-induced translation inhibition, this result further demonstrated the predictive power of the model population.
Sensitivity analysis generated falsifiable predictions about the fragility or robustness of structural features of the initiation architecture. First order sensitivity coefficients were computed for 40 parameter sets selected from the ensemble (materials and methods), time-averaged and rank-ordered for the 250 species in the model, in the presence and absence of insulin and IRS-1 feedback. The sensitive components of insulin signaling shifted from Rheb in the absence of insulin to a combination of Rheb and PI3K in the presence of insulin. Sensitivity coefficients () were calculated with and without insulin over the complete 100 min response (Fig. 5A). Globally, processes involved with 80S formation were consistently ranked among the most sensitive, irrespective of insulin. However, the sensitivity of other signal processing components changed with insulin status. For example, without insulin, Rheb/Rheb-GDP were highly fragile (rank0.25), while PI3K, PIP2, PIP3 and PTEN were highly robust (rank0.0). Surprisingly, the relative sensitivity of these network components changed in the presence of insulin. While the fragility of Rheb/Rheb-GDP shifted modestly upward with insulin, the sensitivity of PI3K and its downstream complexes increased dramatically (rank0.45) following insulin stimulation. This suggested that the combination of PI3K and Rheb activity was critical to insulin action over the full 100 min time window. However, it was unclear whether PI3K was always important, or if there was a temporal window in which PI3K became important following insulin stimulation. To explore this question, we time-averaged the sensitivity coefficients over early- and late-phase time periods following insulin stimulation (Fig. 5B). The 0–5 minute time period captured the initial network dynamics, while the 30–100 minute time period captured the network at a quasi-steady state. Generally, network components were more sensitive under dynamic operation (species beneath the 45 line), compared with steady state. However, there were exceptions to this trend. For example, PI3K, PTEN and TSC1/2 were equally sensitive in both time frames, suggesting these species played important roles in both dynamic and steady state signaling. On the other hand, the Rheb rank decreased from to as the network moved toward steady state. Taken together, the sensitivity results suggested that Rheb activity controlled the background level of translation initiation while the PI3K axis in combination with Rheb regulated insulin-induced initiation. Moreover, the transition between PTEN and PI3K control occurred directly after the addition of insulin, giving rise to switch like behavior.
IRS-1 phosphorylation, a well known negative feedback mechanism [14], [45]–[47], attenuated PI3K sensitivity. We explored the role of IRS-1 feedback by comparing sensitivity coefficients under insulin stimulation in the presence and absence of IRS-1 feedback (Fig. 5C). The most significant change without feedback was the sensitivity of the IR∶IRS-1 and adaptor complexes (Fig. 5C, black fill); IR∶IRS-1, which anchors the adaptor complex to the activated receptor and is immediately upstream of PI3K activation, changed from NSS rank 0.04 to 0.32. The sensitivity of the PI3K/Akt signaling axis also increased in the absence of feedback (Fig. 5C, grey fill). Surprisingly, the sensitivity of Rheb and many ribosomal components decreased in the absence of feedback. Similar results were observed when sensitivity coefficients were time averaged over the 0 to 5 min time window (Fig. 5D). These sensitivity calculations suggest that IRS-1 feedback plays a significant role in insulin signaling by modulating the relative importance of PI3K versus Rheb. Thus, IRS-1 feedback though not directly identified as a fragile regulatory motif, has significant effects on network function.
Lastly, the architectural features of the initiation network identified by sensitivity analysis, as either fragile or robust, were likely parameter independent. While first-order sensitivity coefficients are local, we sampled a family of uncorrelated parameter sets (mean correlation of approximately 0.6) to generate a set of consensus conclusions. By sampling over many uncorrelated sets, we calculated how our conclusions changed with different unrelated parameter sets. The distribution of ranking (standard-error shown in Fig. 5) suggested that despite parametric uncertainty, sensitivity analysis over an uncorrelated model population produced a consensus estimate of the strongly fragile or robust elements of the insulin signaling network. Previously, we (and others) have shown that monte-carlo parameter set sampling produced similar results in several studies across many signaling networks [49], [58]–[60].
Knockdown simulations were conducted for 92 proteins to estimate the functional connectedness of the initiation network. The effects of the perturbations were quantified by calculating the relative change () in translational activity (80S formation) for each simulated knockout in the presence (Fig. 6A) and absence (Fig. 6B) of insulin. Knockdown simulations were conducted using 400 models selected from the ensemble based on error and correlation (materials and methods). Proteins were classified based on their impact on translational activity: little or no effect (, white fill), moderate decrease (, dark grey), critical (, light grey) and increase (, black). Generally, knockdowns in the presence of insulin were more likely to decrease initiation (Fig. 6A). Knockdown analysis identified 24 proteins (or 26% of the network) that were critical to translation initiation irrespective of insulin status; these critical components included mTORC1, S6K1, several initiation factors and other ribosomal components. Sensitivity analysis suggested basal translation was governed by Rheb, while insulin-induced initiation was governed by PI3K. Robustness analysis showed that perturbations in PI3K signaling, in the presence of insulin, restored initiation control to Rheb. Initiation was reduced by 40% by disrupting species immediately upstream or downstream of PI3K; a moderate reduction in the presence of insulin demonstrated that initiation was governed by both PI3K and Rheb. Lastly, deletion of TSC1/2 (negative regulator of Rheb) or 4E-BP1 (sequesters the cap-binding protein eIF4E), increased initiation in the presence of insulin. Interestingly, for several proteins the direction or magnitude of change in initiation activity depended upon the presence or absence of insulin. For example, PTEN deletion significantly increased initiation (1) in the absence of insulin, but had no effect when insulin was present. On the other hand, PI3K deletion had a moderate reduction on 80S formation in the presence of insulin, but only a small effect in the absence of insulin (Fig. 6B). These results suggested that PI3K and PTEN were conditionally fragile proteins; in the presence of insulin, PI3K is a critical signal processing node, while PTEN acts to restrain inadvertent basal initiation.
Paradoxically, Rheb and mTORC2 subunit (sin1, rictor) knockdowns increased initiation. Our expectation from sensitivity analysis was that a Rheb knockdown would reduce initiation, irrespective of insulin status. However, this was not universally true; some members of the model population showed increased initiation (Fig. 6C). Following the deletion of PTEN, approximately 80% (or 323 of the 400 models sampled) had increased initiation in the absence of insulin. Of these models, 16% (or 51 of 323) had at least a two fold increase in translational activity. This result was expected; deletion of a protein species resulted in a qualitatively similar change in initiation across the ensemble of models. However, for Rheb knockdowns, members of the ensemble demonstrated qualitatively different behavior. For 84% (or 334 of 400) of the models sampled, Rheb knockdowns significantly down-regulated initiation. Thus, the vast majority of models behaved as expected. Interestingly, 20 models (or 5% of the models sampled) had increased translation initiation in the presence of a Rheb knockdown, with 15 models demonstrating greater than a two-fold change (Fig. 6C). Thus, the model population estimated by POETs contained models with qualitatively different behavior. Histograms of sin1 and rictor knockdowns showed a similar trend (results not shown). We explored the flux vectors of these outlying parameter sets to better understand the mechanistic effect of Rheb and rictor/sin1 knockouts. All of the outlying models were in regions of parameter space where the association between Rheb and GTP was very high. Strong Rheb/GTP binding resulted in abnormally high signal flux to mTORC1 despite the inhibitory effects of TSC1/2 (Fig. 6D, top-left). Consequently, less GTP was available for the energy-dependent steps of translation initiation (i.e. formation of eIF2-GTP-met-tRNA tertiary complex). Additionally, strong association between Rheb and GTP resulted in high levels of activated mTORC1 and S6K1. However, despite the high levels of mTORC1, GTP-dependent pre-initiation reactions were rate limiting (Fig. 6D, labeled*). Thus, Rheb knockdown released the network from its GTP limitation and shifted the predominant signaling mode to mTORC2. This shift in signaling, while lowering the activated mTORC1/S6K1 level, ultimately resulted in higher levels of initiation (Fig. 6 bottom-left). On the other hand, the rictor/sin1 knockdown behaved differently. The rate-limiting step for the rictor/sin1 knockdowns was mTORC1 activation: more Rheb-GTP was present than there was mTORC1 to be activated (Fig. 6D top-right). Thus, knockdown of rictor/sin1 prevented the assembly of mTORC2 and freed the mTOR subunit to be used for mTORC1 assembly. This shift toward mTORC1 assembly and activation relieved the Rheb-GTP/mTORC1 bottleneck, resulting in increased initiation.
In this study, we developed and analyzed a population of insulin and growth factor induced translation initiation models. These models described the integration of insulin and growth-factor signals with 80S assembly. A family of model parameters was estimated from 24 transient and steady state data sets using multiobjective optimization. In addition to the training data, the model family also predicted novel data sets not used during model training. The population of initiation models was analyzed using sensitivity and robustness analysis to identify the key components of insulin-induced translation initiation. Without insulin, a balance between the pro-initiation activity of the GTP-binding protein Rheb and anti-initiation activity of PTEN controlled basal initiation. Rheb knockdown simulations confirmed decreased initiation in the majority of the model population. Surprisingly, we also identified a model subpopulation in which deletion of Rheb or mTORC2 components increased initiation. In these cases, removal of Rheb or mTORC2 components relieved a rate-limiting bottleneck e.g., constrained levels of GTP, leading to increased initiation. On the other hand, in the absence of insulin, translation initiation increased for all models in the population following a PTEN deletion. In the presence of insulin, Rheb and PTEN were no longer the dominant arbiters of initiation; a combination of PI3K and Rheb activity controlled inducible initiation, where PI3K was only critical in the presence of insulin. PI3K deletion in the presence of insulin removed the ability of the network to process insulin signals, but did not remove initiation altogether. PI3K deletion reduced initiation to approximately 60% of its maximum level. Interestingly, the relative contribution of PI3K versus Rheb to the overall initiation level could be tuned by IRS-1 feedback. In the absence of feedback, PI3K was more important than Rheb to signal propagation, while the opposite was true in the presence of feedback.
PI3K and PTEN in combination with Rheb are components of a switch that regulates inducible and basal translation initiation. In the absence of insulin, a balance between the pro-initiation activity of Rheb and the anti-initiation activity of PTEN regulated basal initiation. On the other hand, in the presence of insulin, control shifted to a combination of Rheb and PI3K, where PI3K activity regulated the inducible fraction of initiation. Thus, deletion of PTEN, constitutive activation of PI3K or constitutively active Rheb could all induce aberrant translation initiation without an insulin or growth factor signal. Yuan and Cantley noted that every major species in the PI3K pathway is mutated or over-expressed in a wide variety of solid tumors [6]. For example, activating mutations in PIK3CA, the gene encoding the catalytic subunit of PI3K, induces oncogene signaling in colon, brain and gastric cancers [61]. On the other hand, PTEN mutations have long been implicated in a spectrum of cancer types [62]. Both PIK3CA and PTEN mutations induce a pro-initiation operational mode in the absence of growth factor. Likewise, constitutive Rheb activity induces a variety of pleiotropic traits involving translation. For example, Saucedo et al. showed that Rheb over-expression in Drosophila melanogaster increased cell size, wing area and G1/S cell cycle progression [63]. Rheb and TSC1/2 mutations are also frequently observed in cancer [64], [65]. Taken together, our study supports the supposition of Taniguchi et al. that PI3K is a critical arbiter of insulin-induced translation initiation [10]. However, we have also shown that initiation control and particularly the role of PI3K was more nuanced; while insulin or growth-factor inducible initiation was controlled by PI3K, basal initiation was controlled by Rheb. Moreover, in the absence of insulin, PTEN was the critical upstream initiation regulator, not PI3K. This suggested that the relative level of the phosphorylated phospholipids PIP2 and PIP3 was actually the key mediator of initiation. Lastly, Taniguchi et al. suggested that Akt was also a key node involved in insulin action. Our previous model directly supports this, however, the current model does not. Rather, our analysis suggested that Rheb was the downstream controller of initiation. These two points of view are not contradictory however, as Rheb activation is driven by phosphorylated Akt.
The initiation model connectivity was assembled from an extensive literature review, however, several potentially important signaling mechanisms were not included. First, we should revisit the role of PRAS40. Currently, PRAS40 acts as a cofactor that aids in pAkt(Ser473)-mediated activation of mTORC1. Sancak et al suggested that PRAS40 sequesters mTORC1, and only after phosphorylation by Akt does it releases from mTORC1 [34]. Other groups have also shown that mTORC1 can phosphorylate and inhibit PRAS40, thus providing a positive feedback mechanism for Akt-mediated mTORC1 activation [66], [67]. A more complete description of PRAS40 will enhance our ability to interrogate Akt dependent mTORC1 activation. Second, we need to refine the description of IRS-1 feedback. Currently, we assume a single deactivating phosphorylation event at Ser308. However, several studies have shown that IRS-1 can be phosphorylated at multiple serine sites, which are both activating and deactivating [44], [68]. Additionally, PTEN is known to dephosphorylate activated PDGF receptors and attenuate their activity, a feature not included currently [69]. A more complete description of IRS-1 phosphorylation could help define how, and under what conditions, IRS-1 regulation attenuates PI3K activation. Third, we modeled the regulation of 4E-BPx as a single phosphorylation event where phosphorylated 4E-BPx was unable to bind to eIF4E. In reality, 4E-BPx family members, such as 4E-BP1, have several phosphorylation sites [70] and the release of eIF4E is driven only after multiple conserved phosphorylation events [71]. Additionally, eIF4E can itself be phosphorylated at Ser209; while there is agreement that the phosphorylation of eIF4E does have a regulatory significance, the data is contradictory as to whether it is positive or negative [72]. Fourth, signaling downstream of mTORC1 has also been shown to mediate translation modes beyond those included in our model. eIF3 has been identified as a scaffolding protein that recruits mTORC1 to untranslated mRNA and facilitates S6K1 and 4E-BP1 phosphorylation [73]. S6K1 can also activate eIF4B, a protein that helps eIF4A to unwind the secondary structure of untranslated mRNA [74]. Further, a recently discovered scaffold protein, SKAR, has been shown to assist S6K1 recruitment to mRNA [75]. Lastly, because of mTORC1's unique cellular role, it would be interesting to explore how other aspects of metabolism interact with insulin signaling to mediate decisions between translation, lipid synthesis or proliferation. In these studies, one could imagine constructing in-vivo mouse models to explore the physiological role of mTORC1 signaling in important diseases such as diabetes or cancer.
The translation initiation model was formulated as a set of coupled non-linear ordinary differential equations (ODEs):(1)The symbol denotes the stoichiometric matrix (). The quantity denotes the concentration vector of proteins (). The term denotes the vector of reaction rates (). The element of the matrix , denoted by , described how protein was involved in rate . If , then protein was consumed in . Conversely, if , protein was produced by . Lastly, if , then protein was not involved in rate . We assumed mass-action kinetics for each interaction in the network. The rate expression for interaction was given by:(2)The set denotes reactants for reaction while denotes the stoichiometric coefficient (element of the matrix ) governing species in reaction . The quantity denotes the rate constant governing reaction . All reversible interactions were split into two irreversible steps. Model equations were generated using UNIVERSAL from an SBML input file (available in the supplemental materials Protocol S1). UNIVERSAL is an open source Objective-C/Java code generator, which is freely available as a Google Code project (http://code.google.com/p/universal-code-generator/). The model equations were solved using the LSODE routine in OCTAVE (v 3.0.5; www.octave.org) on an Apple workstation (Apple, Cupertino, CA; OS X v10.6.4).
When calculating the response of the model to the addition of insulin or other growth factors, we first ran to steady state and then issued the perturbation. The steady state was estimated numerically by repeatedly solving the model equations and estimating the difference between subsequent time points:(3)The quantities and denote the simulated concentration vector at time and , respectively. The vector-norm was used as the distance metric, where s and = 0.001 for all simulations.
We used multiobjective optimization in combination with cross-validation to estimate an ensemble of initiation models. Multiobjective optimization in combination with cross-validation allowed us to address qualitative conflicts in the training data, and to protect against model over-training. While computationally more complex than single-objective formulations, multiobjective optimization is an important tool to address qualitative conflicts in training data that arise from experimental error or cell-line artifacts [76]. Multiobjective optimization balances these conflicts allowing us to identify a consensus model population. In this study we used the Pareto Optimal Ensemble Technique (POETs) to perform the optimization. POETs integrates standard search strategies e.g., Simulated Annealing (SA) or Pattern Search (PS) with a Pareto-rank fitness assignment [52]. Denote a candidate parameter set at iteration as . The squared error for for training set was defined as:(4)The symbol denotes scaled experimental observations (from training set ) while denotes the scaled simulation output (from training set ). The quantity denotes the sampled time-index and denotes the number of time points for experiment . In this study, the experimental data used for model training was typically the band intensity from immunoblots, where intensity was estimated using the ImageJ software package [77]. The scaled measurement for species at time in condition is given by:(5)Under this scaling, the lowest intensity band equaled zero while the highest intensity band equaled one. A similar scaling was defined for the simulation output. By doing this scaling, we trained the model on the relative change in blot intensity, over conditions or time (depending upon the experiment). Thus, when using multiple data sets (possibly from different sources) that were qualitatively similar but quantitatively different e.g., slightly different blot intensities over time or condition, we captured the underlying trends in the scaled data.
We computed the Pareto rank of by comparing the simulation error at iteration against the simulation archive . We used the Fonseca and Fleming ranking scheme [78] to estimate the number of parameter sets that dominate . Parameter sets with increasing rank are progressively further away from the optimal trade-off surface. The parameter set was accepted or rejected by POETs with probability :(6)where is the annealing temperature and denotes the Pareto rank for . The annealing temperature was discretized into 10 quanta between and and adjusted according to the schedule where was defined as . The initial temperature was given by , where was used in this study and the final temperature was . The epoch-counter was incremented after the addition of 100 members to the ensemble. Thus, as the ensemble grew, the likelihood of accepting parameter sets with a large Pareto rank decreased. To generate parameter diversity, we randomly perturbed each parameter by . We performed a local pattern search every steps to minimize the residual for a single randomly selected objective. The local pattern-search algorithm has been described previously [79].
A leave-three-out cross-validation strategy was used to simultaneously calculate the training and prediction error during the parameter estimation procedure [80]. The 24 training data sets were partitioned into eight subsets, each containing 21 data sets for training and three data sets for validation. The leave-three-out scheme generated 18,886 probable models. From the approximately 6000 rank zero models, we iteratively selected 50 random models from each cross-validation trial with the lowest correlation and shortest Euclidian distance to the origin (minimum error). This selection technique produced sub-ensembles with low set-to-set correlation (0.50) and minimum training error.
Sensitivity coefficients were calculated for 40 models selected from the ensemble (rank-zero, low-correlation, minimum error selection). First-order sensitivity coefficients at time :(7)were computed by solving the kinetic-sensitivity equations [81]:(8)subject to the initial condition . The quantity denotes the parameter index, denotes the number of parameters in the model, denotes the Jacobian matrix, and denotes the th column of the matrix of first-derivatives of the mass balances with respect to the parameters. Sensitivity coefficients were calculated by repeatedly solving the extended kinetic-sensitivity system for forty parameters sets selected from the final 400 member ensemble. These sets were chosen to be comparable to the final 400 member ensemble on the basis of parametric coefficient of variation (CV); the sets selected for sensitivity analysis had a mean CV of 0.850.5 and a mean correlation of approximately 0.6. Thus, there were diverse and uncorrelated. The Jacobian and the vector were calculated at each time step using their analytical expressions generated by UNIVERSAL.
The resulting sensitivity coefficients were scaled and time-averaged (Trapezoid rule):(9)where denotes the final simulation time. The time-averaged sensitivity coefficients were then organized into an array for each ensemble member:(10)where denotes the index of the ensemble member, denotes the number of parameters, denotes the number of ensemble samples and denotes the number of model species. To estimate the relative fragility or robustness of species and reactions in the network, we decomposed the matrix using Singular Value Decomposition (SVD):(11)Coefficients of the left (right) singular vectors corresponding to largest singular values of were rank-ordered to estimate important species (reaction) combinations. Only coefficients with magnitude greater than a threshold ( = 0.001) were considered. The fraction of the vectors in which a reaction or species index occurred was used to determine its importance (sensitivity ranking). The sensitivity ranking was compared between different conditions to understand how control in the network shifted as a function of perturbation or time (Fig. 5).
Robustness coefficients were calculated as shown previously [60]. Robustness coefficients (denoted by ) are the ratio of the integrated concentration of a network marker in the presence (numerator) and absence (denominator) of a structural or operational perturbation. The quantities and denote the initial and final simulation time, respectively, while and denote the indices for the marker and the perturbation respectively. If , then the perturbation increased the marker concentration. Conversely, if the perturbation decreased the marker concentration. Lastly, if the perturbation did not influence the marker concentration. Robustness coefficients were calculated over 400 models selected from the ensemble (rank-zero, low-correlation, minimum error selection). Convergence analysis suggested that the qualitative conclusions drawn from the robustness analysis would not change if more than N = 400 parameter sets were sampled (Fig. S1).
|
10.1371/journal.pntd.0005694 | Increasing airline travel may facilitate co-circulation of multiple dengue virus serotypes in Asia | The incidence of dengue has grown dramatically in recent decades worldwide, especially in Southeast Asia and the Americas with substantial transmission in 2014–2015. Yet the mechanisms underlying the spatio-temporal circulation of dengue virus (DENV) serotypes at large geographical scales remain elusive. Here we investigate the co-circulation in Asia of DENV serotypes 1–3 from 1956 to 2015, using a statistical framework that jointly estimates migration history and quantifies potential predictors of viral spatial diffusion, including socio-economic, air transportation and maritime mobility data. We find that the spread of DENV-1, -2 and -3 lineages in Asia is significantly associated with air traffic. Our analyses suggest the network centrality of air traffic hubs such as Thailand and India contribute to seeding dengue epidemics, whilst China, Cambodia, Indonesia, and Singapore may establish viral diffusion links with multiple countries in Asia. Phylogeographic reconstructions help to explain how growing air transportation networks could influence the dynamics of DENV circulation.
| In the past half century the incidence of dengue fever worldwide has increased 30-fold, with an estimated ~390 million infections per year. The years 2014 and 2015 were characterized by large dengue outbreaks worldwide, which are a threat to public health, especially in Asian countries. Here we use a phylogeographic approach to reconstruct historical virus movement and evaluate multiple potential predictors of the spatial spread of dengue virus (DENV) among Asian countries. We show that air traffic has played a more important role than maritime transport in shaping large-scale geographic dispersal of DENV in Asia. Our study suggests that the patterns of DENV spread result from the interplay between intensity and structure of human mobility through the air transportation network.
| Dengue virus (DENV) is a growing threat to public health, with nearly 390 million infections every year worldwide, of which ~96 million are symptomatic [1,2]. An estimated 2.5 billion people are at risk of dengue infection [3]. Dengue is regarded as the world’s most important mosquito-borne viral disease and is endemic in more than 100 countries [4], with most disease burden limited to tropical and subtropical regions [5]. However, in 2014 an outbreak of dengue occurred in Japan for the first time in over 70 years. This occurred despite the country’s temperate climate, and viral phylogenetic analysis suggests that the Japanese outbreak resulted from international travel from Southeast Asia [6]. Furthermore, the number of reported dengue cases and outbreaks continues to increase. In 2014, a dengue outbreak affected several Asian countries, including China, Thailand, Vietnam and Japan [7,8].
The majority of DENV infections are asymptomatic or cause a mild febrile disease known as dengue fever, while the more severe forms of dengue infection—dengue hemorrhagic fever (DHF) and dengue shock syndrome (DSS)—may be life threating with >20% mortality [9]. Epidemiological studies have indicated that DHF often occurs when a dengue-immune person acquires a second infection with a different DENV serotype [10,11], and it has been hypothesized that DHF/DSS may result from a process of antibody-dependent enhancement [12,13]. The geographical areas in which transmission of multiple dengue serotypes occurs has grown in recent years and the pattern of co-circulation is conspicuously different from that which prevailed decades ago [14,15]. Furthermore, a growing number of DHF/DSS cases in the last 50 years, especially in Asia, demonstrates the need for a better understanding of how DENV genetic diversity and transmission jointly shape dengue epidemics.
The growing scale of human mobility, particularly through air transportation, underscores an increase of pathogen introductions into geographic areas suitable for transmission, potentially contributing to the emergence and the re-emergence of epidemics [16,17]. Recent examples include the SARS epidemic, novel influenza A virus strains, the Middle East respiratory syndrome (MERS) coronavirus, and Zika virus [18–21]. At local scales, endemic circulation of DENV strains can be driven by viral lineage introduction events that eventually lead to lineage replacement [22] and viral introductions are expected to increase with human mobility. Together, these observations suggest that increasing human mobility through air-traffic networks may impact the spread of pathogens at large geographical scales. Here we combine Asian socio-economic data, air and maritime transportation network data and viral genetic data to identify key drivers of the spread of DENV serotypes 1, 2 and 3 in Asia, and to reconstruct the patterns of virus circulation among Asian countries over the past half century.
Maximum phylogenetic information would be obtained by studying whole DENV genomes. However, available DENV whole genomes from Asia have insufficient coverage through time and space for reliable analysis, and most available sequences comprise partial or complete coding E gene sequences. In order to generate a data set with both acceptable phylogenetic diversity and spatiotemporal sampling, we used DENV E gene sequences in subsequent analyses. DENV (DENV-1 to DENV-3) envelope (E) gene sequences with known collection dates and locations of sampling in Asia were collected from GenBank. DENV-4 was not included in this study because too few samples were available (only 64 sequences from 11 countries; Fig 1). The remaining strains comprised a total of 2,202 sequences sampled between 1956 and 2015, from 20 distinct countries or geographic regions (Fig 1). Sequences were grouped by serotype and aligned separately using MAFFT [23]. Recombination was inspected using the methods implemented in RDP3 and SimPlot [24]. After removing duplicate and recombinant strains, the final data set contained 1,272 DENV-1 sequences, 628 DENV-2 sequences and 302 DENV-3 sequences. In the complete sequence dataset, some countries, such as Vietnam, Cambodia, Thailand, and Singapore, were over-represented. In order to control for possible bias from uneven sampling, we randomly subsampled the complete sequence datasets by location and sampling time. At most 10 sequences were sampled per country and per year in order to create a more equitable spatio-temporal sampling distribution. After sub-sampling, the total number of sequences analyzed here was 327 for DENV-1, 357 for DENV-2, and 202 for DENV-3, sampled over a total of 59 years (S1 Fig). Details of the sequences in each data set, including information on the year of isolation, sampling location, and accession numbers, are provided in Supplementary Information (S1 Table).
For each serotype subsampled data set, we first estimated the correlation between root-to-tip genetic divergence and sequence sampling dates, using TempEst [25]. This preliminary analysis indicated a good temporal signal for all serotypes (S2 Fig). To reconstruct past population dynamics, we used a coalescent-based Gaussian Markov random field (GMRF) method with the time-aware smoothing parameter [26], as implemented in BEAST v1.8.2 [27]. A GTR+I+Γ nucleotide substitution model and an uncorrelated lognormal relaxed molecular clock model were used, with a prior distribution for the evolutionary rate parameter set to a Γ distribution with shape = 0.001 and scale = 1000. The BEAGLE library was used to accelerate computation [28]. For each serotype, three independent analyses of 150 million generations were performed, sampling parameters and trees every 15,000 generations. Analyses were combined after the removal of a burn-in of 10–20% of the samples and were checked visually in Tracer v.1.5.
To reconstruct the spatial dynamics of DENV-1–3, we used a Bayesian Markov chain Monte Carlo (MCMC) phylogeographic discrete approach [29,30] that estimates the ancestral locations along each branch of a viral phylogeny, as implemented in BEAST v1.8.2 [27]. A GTR+ I+Γ nucleotide substitution model was used in this analysis. Three independent MCMC chains were run for 150 million states, sampling every 15,000 states after the removal of 10% burn-in. Maximum clade credibility (MCC) trees were summarized using TreeAnnotator and visualized using SPREAD [31].
To provide a minimal set of location state changes that provide an adequate description of viral spread, we used Bayesian stochastic search variable selection (BSSVS) [30,32]. BSSVS uses Bayes factors (BF) and binary indicator variables (I) to identify statistically supported viral lineage movement routes. Values of BF > 6 and I > 0.5 were considered as denoting a significant migration pathway, where BF > 1,000 indicates decisive statistical support, 100 ≤ BF < 1,000 indicates very strong support, 30 ≤ BF < 100 indicates strong support, 10 ≤ BF < 30 indicates substantial support and 6 ≤ BF < 10 indicates support [32]. We next estimated the number of expected transitions among location-pairs (“Markov jum” counts) along the phylogeny branches [33], which provided a quantitative measure of successful viral introductions among countries [34].
We investigated the air transportation network in Asia using mobility data from ICAO (International Civil Aviation Organization; http://www.icao.int/Pages/default.aspx). This database contains the number of passengers traveling among 373 airports in Asia during the years 1982, and 1992–2012, and included all scheduled flights both for large and small aircraft. Country-level movement of passengers was obtained by aggregating airport-level movement for each country.
To obtain insight into temporal changes across the air transportation network in Asia, directed and weighted air flow networks were constructed, with countries as the network nodes. Network edges were weighted by the number of air passengers connecting pairs of countries. Hubs within the Asian air transportation network were also identified using two network topology properties: degree centrality and betweenness centrality. Degree centrality measures the number of edges connected to a node. Here, the degree of a given country refers to the number of airlines linking to it in the airline network. Betweenness centrality depends on the proportion of shortest paths between all pairs of vertices that pass through a given node. Thus the betweenness centrality of a given country is a measure of the extent to which a country lies on routes between other countries in the airline network and is calculated as:
Betweenness of node k=∑s≠v≠tσst(k)σst
(1)
where σst is the total number of shortest paths from node s to node t and σst (k) is the number of those paths that pass through k.
To represent the shipping connectivity between pairs of countries, the liner shipping bilateral connectivity index (LSBCI) in Asia from 2008 to 2014 was obtained from United Nations Conference on Trade and Development (UNCTAD; http://unctadstat.unctad.org/EN/Index.html), which captures the amount of goods transported between two countries. Container port throughput (CPT) for each Asian country was also obtained from UNCTAD.
To investigate the potential predictors driving DENV-1, DENV-2 and DENV-3 spatial spread, we used a generalized linear model (GLM) extension to Bayesian phylogeographic inference [20,35]. The phylogeographic GLM method parameterizes the continuous-time Markov chain (CTMC) matrix of among-location lineage migration parameters as a log-linear function of several potential predictors. Each function includes a coefficient β (in log space) and a binary indicator variable δ. BSSVS [30] was used to estimate the posterior probability that each predictor is included or excluded from the model. We considered several potential predictors of DENV diffusion. These included log-transformed and standardized measures of demographic and economic data, geographical distances among countries, absolute latitude, DENV sample sizes. Finally, predictors of human mobility from shipping connectivity and air transportation movement were included (averaged values were estimated using data in 1982, 2000, and 2012).
Specific details of potential predictors are as follows: (i) average distance between two countries was estimated as the average of pairwise distances between all pairs of airports in the two countries, (ii) absolute latitudes for each country were calculated as the latitudes of their geometric center, (iii) passenger flow represents the number of passengers on fights between each pair of countries, (iv) LSBCI (linear shipping bilateral connectivity index) represents goods transported by the sea between pairs of countries, (v) population sizes and densities for each country in 2000 were obtained from the UN World Population Prospects database (http://www.un.org/en/development/desa/population/), (vi) gross domestic product (GDP) for each country in 2000 was collected from the UN Statistics Division (https://unstats.un.org) and (vii) average temperature and rainfall for each Asian country was extracted from the data set provided by WorldClim [36]. In addition, we included sample sizes (number of dengue sequences per country) for origin and destination locations as potential predictor variables, in order to test the impact of heterogeneous sampling.
Fig 1A depicts the geographic locations of the sequences used in this study, from a total of 20 distinct countries or geographic regions in Asia; most sequences were sampled in Southeast Asia. To reduce potential bias in the reconstruction of spatial spread that may arise from over-sampling particular locations, we subsampled the original data (see Methods). It is important to note that, in recent years, particularly since the 1980s, genetic evidence of multiple dengue virus sequences from a given location has increased (Fig 1B) in Asian countries (S3 Fig).
Evolutionary analysis of 327 DENV-1 E gene sequences showed that Asian sequences fell into five distinct lineages, genotypes I—V. The 357 DENV-2 E gene sequences were classified into five genotypes: Asia I, Asia II, America/Asia, Cosmopolitan, and the America genotype. Further, the 202 DENV-3 E gene sequences comprised four genotypes, genotypes I and III—V. The estimated evolutionary rates of the E gene of each serotype, under the selected evolutionary model, were 7.50×10−4 (95% highest posterior density, HPD: 6.73×10−4–8.28×10−4) substitutions per site per year (s/s/y) for DENV-1, 8.14×10−4 (95% HPD: 7.38×10−4–8.94×10−4) s/s/y for DENV-2, and 7.96×10−4 (95% HPD: 7.01×10−4–8.89×10−4) s/s/y for DENV-3. The corresponding estimates of the time of the most recent common ancestors (TMRCAs) of DENV-1–3 in Asia were 1911 (95% HPD: 1891–1928), 1901 (95%HPD 1868–1931, after removal of sylvatic lineages) and 1929 (95% HPD: 1916–1941) respectively.
Air transportation has historically exhibited significant growth [37]. We found that Asian air transportation grew substantially from 1980 onwards (Fig 2A). We observe that several countries act as hubs in airline network, e.g. India, Singapore, and Thailand, while other countries exhibit growing network centrality over time, such as China and Malaysia (Fig 2B). It is interesting to note that the sequences obtained from dengue patients in 2014 in Japan shared a very high identity with a sequence from China, which shares one of the busiest Asian flight routes with Japan (Fig 2B). These observations prompt us to question the role of air passenger transportation in DENV spread in Asia.
To infer the contribution of candidate factors driving DENV diffusion in Asia we used a generalized linear model that simultaneously estimates ancestral geographic reconstruction and identifies the contribution of potential predictors of spatial spread [20,35]. Our results show that most candidate predictors of virus spread, such as geographic distance and demographic factors, are not significantly associated with viral spread (Fig 3). However, we do find that air passenger flow is a dominant driver of DENV lineage movement and the inclusion of this factor in the model is supported for all three DENV serotypes. GDP at the location of origin and CPT (container port throughput) at the destination location are negatively associated with DENV-1 lineage movement in Asia, but not for the other two genotypes (Fig 3). Sample sizes for each location are not associated with viral lineage movement, which suggests that our conclusions are not driven by sampling biases.
To understand the spatial circulation of DENV in Asia, we reconstructed the past spatial transmission patterns of serotypes DENV-1–3 over the study period, for each country from which sequences were sampled (Fig 4). A Bayesian stochastic search variable selection procedure was employed to infer a minimum set of location exchange events, while a Markov jump (MJ) analysis was used to quantity viral lineage movement among pairs of locations (see Methods). The phylogeographic analysis in Fig 5 indicates several significant migration links (with BF support > 1000). For DENV-1 these are between Cambodia and Vietnam, Thailand and Laos, and Singapore and China. For DENV-2, the well supported links are between Thailand and Cambodia, China and Indonesia, and Indonesia and Singapore. Finally, for DENV-3 the strongly supported links are between India and Sri Lanka (BF > 800). We also inferred a number of other well supported movements among countries; a full list is provided in S2 Table.
We further find that estimated number of viral lineage migrations (including both importations and exportations) for each country is associated with measures of centrality in the air transportation network (Pearson correlation: r = 0.45, P = 0.05, for degree centrality and state transitions; and r = 0.73, P < 0.01, for betweenness centrality and state transitions; Fig 6A, S4 Fig). However, countries do not all contribute equally to viral lineage dissemination. An analysis of the number of virus lineage movements and the air transportation network analysis for each country indicates increased lineage virus movements from Thailand, India, and Indonesia to other locations [38,39], and a trend towards viral lineage importation from other locations for Vietnam and China. Lineage import and export is approximately equal for Cambodia and Singapore (Fig 6B and S5 Fig). Given the betweenness centrality of Thailand and India in the regional airline network, these countries may act as net sources of viral lineages, while China, Cambodia, Indonesia and Singapore may establish strong links with multiple countries within the network.
There has been an increasing frequency of DENV lineage co-occurrence in China over time, concomitant with its increasing centrality in the Asian air transportation network (Figs 2 & 5). In contrast, Vietnam, a dengue-endemic country, has a low frequency of viral lineage import and export with other locations (Fig 6B), likely due to its lower network centrality. However, we find that some non-endemic countries that do not contribute significantly to viral lineage movement across Asia may still maintain high airline passenger flows, such as Japan, South Korea, and countries in Western Asia. This may be because such areas lack a climate suitable for widespread dengue virus transmission. Further we are not able to infer transmission between countries if no viral sequences are available for those locations. The results here do not appear to be driven by heterogeneous sampling; similar results were obtained using downsampled datasets with a maximum of 5 sequences per country per year (S6 Fig).
Our study provides a temporal description of the patterns of DENV spread across Asia. The spatial dynamics of dengue virus in Asia inferred here suggests that DENV genetic diversity in Asia is dynamic yet spatially structured, with frequent virus lineage movement among countries, and frequent co-circulation of dengue virus lineages. Our analysis suggests that the air transportation network has contributed to the spatial distribution of DENV serotypes in Asia. This can be explained by the movement of viraemic individuals; if the destination and timing of virus introductions coincides with a climatically suitable period then an epidemic can become established in a susceptible recipient population. Further, it is possible that the mobility of viraemic mosquitoes through air transportation networks is non-negligible [40]. For example, one study estimated that 8–20 Anopheline mosquitoes were imported into France per flight in one 3-week period in 1994 [41]. Previous studies have also shown that mosquito species, including Aedes albopictus, can survive long-haul flights [42,43]. Implementing disinfection is proving effective in vector control and disease prevention [40,44], highlighting the importance of air transportation in pathogen importation.
The positive coefficient β for the air traffic predictor in our phylogenetic GLM model indicates that DENV lineage dispersal is associated with air traffic; this result is consistent with previous studies of DENV in Brazil, which concluded that air travel of humans and/or mosquitoes is associated with virus lineage movement [22]. Notably there is strong statistical support for the inclusion of the air travel predictor variable in the GLM models for all three DENV serotypes in our study. Further, we find that origin GDP and destination CPT (container port throughput) are negatively related to DENV diffusion for DENV-1; this result is possibly linked to better public health infrastructure and vector control programmes in wealthier countries [45]. However, there was no support for the inclusion of those two predictors in the analyses of DENV-2–3. This may stem from the larger number of sequences for DENV-1, or from differences in sampling distribution among countries, despite the fact we have used a subsampling procedure in an attempt to mitigate potential bias. Alternatively, observed among-genotype differences may be driven solely by stochastic variation in the process of spatial dissemination; for example, the well-supported migration links inferred using DENV-1–3 sequences were not identical (Fig 5).
Analysis of the air transportation network indicates differences in the roles of air transport nodes in the regional-scale dissemination of DENV lineages. Our data suggest that large transportation hubs account for most of the inferred virus lineage movements within their respective networks. For example, Thailand and India seem to act as central hubs of DENV lineage movement in Asia, demonstrating both inward/outward viral migrations with other countries in the region. In contrast, our results suggests that the low level of network centrality of Vietnam may explain its limited role in regional DENV geographical dissemination. Some countries, such as China and Malaysia, may have had an increasing role in DENV spread through time. We note that dengue is still considered to be an imported disease in mainland China [46]. It might be expected that some DENV endemic countries, with climatic suitability for dengue transmission and high betweenness-centrality in air transportation network, contribute to seeding Asian dengue epidemics. However, it has been shown that the risk factors for the persistence and transmission of dengue are complex [47], once the socioeconomic conditions were taken into account (as shown in Fig 3).
Previous studies based on simulation models have shown that global warming is likely to increase the area of land with a climate suitable for dengue virus transmission by altering the distribution of Aedes aegypti, the main mosquito vector of dengue [48,49]. This may result in more countries being at risk of dengue virus via the air transportation network, resulting in the potential for spread across even greater geographic scales. Temperature increases may lead to lengthened mosquito lifespans and shortened extrinsic incubation periods, which may result in more infected mosquitoes for a longer period of time [50,51].
Our study has several limitations. First, we cannot discern the relative contribution of infected humans versus infected mosquitoes in the spread of DENV, which requires more detailed epidemiological data and pathogen genome sequences from both mosquitoes and patients in future studies. However the information provided by the genetic analyses here can be useful to parameterize future spatially-structured transmission models. For example, the structure and travel flux of the airline transport network, as well as the among-country rates of lineage movement from the phylogeographic model, could be used to inform future simulations [20]. Second, our findings are based on DENV E genes for Asian countries, as this is the only genome region for which sufficient numbers of sequences are available, yet complete virus genomes would provide better phylogenetic resolution [52,53]. Further, our analyses were limited to DENV in Asia and do not consider transmission to or from other regions. Finally, the uneven sampling of DENV sequences in Asia, especially in a few countries after 2000 (e.g. Vietnam and Singapore) necessitated the use of sub-sampling analyses that accounted for the number of samples per location. Results obtained from a more representative data set indicates that our main conclusions are robust to sample sizes.
Future trends in global mobility could potentially accelerate the appearance and diffusion of DENV worldwide. Prevention and control of dengue epidemics requires a better understanding of its mode of geographical dissemination, especially for countries in the tropics. Our study highlights the importance of developing a multivalent DENV vaccine in order to cope with increasing frequency of DENV serotype co-occurrence. The potential impacts of vaccination on dengue epidemics in Asia should be considered [54]. Viral spatial dissemination, together with host age-structure and host-vector interactions are required in mathematical models to inform future vaccine deployment decisions.
|
10.1371/journal.pntd.0001207 | Community Participation in Chagas Disease Vector Surveillance: Systematic Review | Vector control has substantially reduced Chagas disease (ChD) incidence. However, transmission by household-reinfesting triatomines persists, suggesting that entomological surveillance should play a crucial role in the long-term interruption of transmission. Yet, infestation foci become smaller and harder to detect as vector control proceeds, and highly sensitive surveillance methods are needed. Community participation (CP) and vector-detection devices (VDDs) are both thought to enhance surveillance, but this remains to be thoroughly assessed.
We searched Medline, Web of Knowledge, Scopus, LILACS, SciELO, the bibliographies of retrieved studies, and our own records. Data from studies describing vector control and/or surveillance interventions were extracted by two reviewers. Outcomes of primary interest included changes in infestation rates and the detection of infestation/reinfestation foci. Most results likely depended on study- and site-specific conditions, precluding meta-analysis, but we re-analysed data from studies comparing vector control and detection methods whenever possible. Results confirm that professional, insecticide-based vector control is highly effective, but also show that reinfestation by native triatomines is common and widespread across Latin America. Bug notification by householders (the simplest CP-based strategy) significantly boosts vector detection probabilities; in comparison, both active searches and VDDs perform poorly, although they might in some cases complement each other.
CP should become a strategic component of ChD surveillance, but only professional insecticide spraying seems consistently effective at eliminating infestation foci. Involvement of stakeholders at all process stages, from planning to evaluation, would probably enhance such CP-based strategies.
| Blood-sucking triatomine bugs are the vectors of Chagas disease, a potentially fatal illness that affects millions in Latin America. With no vaccines available, prevention heavily depends on controlling household-infesting triatomines. Insecticide-spraying campaigns have effectively reduced incidence, but persistent household reinfestation can result in disease re-emergence. What, then, is the best strategy to keep houses free of triatomines and thus interrupt disease transmission in the long run? We reviewed published evidence to (i) assess the effectiveness of insecticide-based vector control, gauging the importance of reinfestation; (ii) compare the efficacy of programme-based (with households periodically visited by trained staff) and community-based (with residents reporting suspect vectors found in their homes) surveillance strategies; and (iii) evaluate the performance of alternative vector-detection methods. The results confirm that insecticide-based vector control is highly effective, but also that persistent house reinfestation is a general trend across Latin America. Surveillance systems are significantly more effective when householders report suspect bugs than when programme staff search houses, either manually or using vector-detection devices. Our results clearly support the view that long-term vector surveillance will be necessary for sustained Chagas disease control – and that community participation can substantially contribute to this aim.
| Chagas disease still imposes a heavy burden on most Latin American countries, with about 10–12 million people infected by Trypanosoma cruzi [1], [2]. Multinational control initiatives have since the early 1990s drastically reduced prevalence and incidence, mainly through insecticide-based elimination of domestic vector populations (blood-sucking bugs of the subfamily Triatominae) [3] and systematic screening of blood donors with highly sensitive serological tests [1], [2], [4], [5]. In spite of these advances, vector-borne transmission is estimated to cause about 40,000 new infections per year [6]. Reinfestation of treated households by native vectors as the residual effect of insecticides vanishes is the most likely mechanism underlying such persistent transmission [7]. Similarly, outbreaks of acute Chagas disease have been attributed to the contamination of foodstuffs by infected adult (i.e., winged) triatomines that invade premises where food is processed or stored [8]–[12]. In Amazonia and other humid forest ecoregions, where the bugs rarely colonise inside houses, endemic, low-intensity transmission seems also mediated by adventitious, household-invading triatomines [13]–[15]. In addition, there is growing concern that insecticide-resistant vector populations, such as those detected in southern South America [16], [17], may threaten effective disease prevention.
This rapid overview shows why sustained Chagas disease control is believed to require some sort of longitudinal, long-term surveillance system capable of detecting and eliminating household infestation foci [1], [18]. Surveillance typically relies on the periodical inspection of households by trained personnel. Active vector searches are performed with or without the aid of chemical ‘flush-out’ agents such as low-dose pyrethroid dilutions, and infestation foci are eliminated by insecticide spraying when discovered [18].
However, detecting the vectors can be difficult, particularly when only small populations occur within or around households. In fact, vector colonies are expected to become rarer and smaller as control programmes proceed, and managers are progressively less prone to fund costly active surveillance resulting in few detection events. A number of vector-detection devices have been designed in an attempt to enhance surveillance; most consist of boxes that triatomines can use as refuges or of paper sheets or calendars where the typical faecal streaks of the bugs can be identified [19]–[27]. Such ‘sensing devices’ are placed within households or in annex structures and checked periodically for bugs or their traces, supposedly reducing the costs of surveillance while retaining adequate sensitivity [26]–[29].
Finally, and since the early vector control trials, there has been a perception that resident householders may have better chances of discovering bugs in their own homes than a visiting team searching the house for a few minutes every several months [30]–[32]. ‘Community participation’ in entomological surveillance gained extra momentum with the Declaration of Alma Ata [33], [34], which “…encouraged approaches to health care that incorporated community participation and community development” (ref. [34], p. 1). Experiences involving community participation in Chagas disease control have been described in several settings across Latin America [18], [30], [31]; they seem to converge towards an encouraging overall picture, and the Chagas disease example has accordingly been praised in several subjective reviews (e.g., [35], [36]).
However, the effectiveness of these diverse strategies for Chagas disease vector surveillance, including community participation, has not been thoroughly and objectively assessed at the continental scale. With the aim of filling this gap, we systematically reviewed the published evidence on this issue, tackling specifically the following major questions: (i) How common and important is the phenomenon of house reinfestation by triatomine bugs after control interventions?; (ii) How effective are different vector surveillance strategies at detecting infestation/reinfestation foci?; (iii) To what extent have community participation and empowerment been effectively promoted?; and, finally, (iv) Can available strategic options be condensed in overarching recommendations for surveillance that apply across the highly diverse ecological and social-cultural settings where the problem is present?
The review protocol is available upon request from the corresponding author. This review was carried out in the context of a collaborative project led by the Inter-American Development Bank, and was not formally registered. We searched Medline, ISI Web of Knowledge, Scopus, LILACS, and SciELO; the major query argument was “Triatomin* AND (Control OR Surveillance)”. Searches retrieved records from 1948 to 2009, including additional documents identified by searching bibliographies and in the authors' records. This search strategy aimed at recovering documents describing vector control interventions, with or without surveillance, so that post-control reinfestation trends could also be assessed. Only documents describing field interventions aimed at the control and/or surveillance of domestic Chagas disease vectors were included in the full review process. Descriptive (non-intervention) reports, results of research with laboratory or experimental vector populations, expert reviews, and opinion or commentary pieces were either excluded or used only for the introduction and/or discussion.
We were particularly interested in comparing strategies involving institutional (by professional staff) or participatory surveillance. We also compared alternative methods for vector detection, including active searches, vector-detection devices, and community participation. Major outcomes included household infestation/reinfestation indices (or, in some cases, bug catches) and vector detection rates. Inclusion/exclusion of documents was assessed independently by ARdA and FA-F, and discrepancies resolved by consensus. Figure 1 presents the flow diagram of the review process. Data were independently extracted by ARdA and FA-F using predefined data fields inspired by the Guide to Community Preventive Services [37] (www.thecommunityguide.org) and including study quality indicators. FA-F revised data extraction results and resolved inconsistencies by re-checking the original documents. The following items were considered: (1) study classification (study design, intervention components, whether or not the intervention was part of a broader initiative, outcomes); (2) descriptive information, including (2.i) description of the intervention (what was done, how, where and by whom it was done, theoretical basis of the intervention, types of organisation involved, whether or not there was any intervention in a control group), (2.ii) study characteristics (place, time, population, settings, outcome measurement, whether or not there was a measurement of exposure to the intervention), (2.iii) results (primary results, sample and effect sizes), and (2.iv) applicability in settings other than the actual study one (direct and indirect costs, harms and benefits, implementation process, and whether the community participated at each stage of the process – design, pre-implementation, effecting, and evaluation); and (3) study quality, including quality of descriptions, sampling (universe, eligibility and selection of participants, sample size, potential sampling biases), effect measurements, data analyses (statistics, confounders, repeated measures or other sources of non-independence), and interpretation of results (rate of adherence, control and assessment of potential confounders and sources of bias). Relevant references and other details deemed important were also recorded. The protocol required extracting detailed demographic data about intervention and control or indirectly affected populations. Such information was however absent from or incomplete in most studies; this, together with the fact that the outcomes of primary interest refer to households, not individual people, led us to exclude these items from the protocol during the course of the review.
The often important morphological, ecological and behavioural differences among triatomine bug species [3], combined with the likely sensitivity of results to study-specific (methods, research team performance) and site-specific conditions (vector density, household building materials and structure), led us to avoid estimating meta-analytical summary effects from different reports. Inadequate design and/or reporting of several studies were further factors hindering meta-analysis. When enough information was given in the original reports, we nonetheless re-analysed data from studies comparing control strategies (in terms of household infestation rates) and vector detection techniques (in terms of detection rates). Whenever possible, we used McNemar's tests for correlated proportions [38], with odds ratios (OR) estimated as the ratio of discordant results. When independence of observations was likely, or in the absence of complete data on repeated observations, ORs were estimated from standard contingency tables [39]. Approximate OR 95% confidence intervals (95%CI) were calculated by assuming normality of log-odds [39]. The VassarStats online facility (http://faculty.vassar.edu/lowry/VassarStats.html) and Microsoft Office Excel® spreadsheets were used for the analyses.
Database searches retrieved 1,342 candidate documents; elimination of duplicates yielded 858 unique records (Figure 1) in English, Spanish, or Portuguese. Assessment of titles and abstracts yielded five groups: (a) documents apparently describing control and/or surveillance interventions (236 records), (b) non-intervention studies, (c) studies with laboratory or experimental vector populations, (d) subjective reviews and opinion pieces, and (e) reports clearly irrelevant to our review. Evaluation of group (a) documents against inclusion criteria identified 93 reports for full data extraction [Supporting Information, List S1]; of the remaining 143 (plus several additional references), 26 studies [Supporting Information, List S2] were also used for partial quantitative assessments, and the rest were considered as supplementary sources of qualitative information for the introduction and/or discussion.
The spatial and ecological coverage of our review is represented in Figure 2. Only 11 randomised trials [40]–[50] were identified, with just one crudely assessing a community-based intervention [50] and four describing different aspects of the same trial [44]–[47]. Over half of the studies dealt directly or indirectly with different strategies for household-level vector surveillance. Interventions ranged from insecticide spraying (the most frequent) to educational activities, with a few studies describing alternative control approaches such as environmental management [51]–[58] or insecticide-treated materials [48], [49], [59]. Most studies measured intervention effects as reductions in household infestation rates (through entomological surveys) or as vector detection rates (through detection records). While the quality of the descriptions was generally adequate, analytical procedures were often dubious; for instance, albeit many studies describe results in which the same sampling units were assessed more than once (e.g., before-after, time-series) or by more than one method (e.g., vector-detection studies), only a few apply statistical tests suited for repeated measures or other sources of non-independence of observations.
Collaborative efforts involving both academic institutions and official public health agencies were common (∼70% of studies), a typical historical trait of Chagas disease vector control [60]. Even though sustainability was discussed in several documents, detailed assessment of the costs (monetary and not) and potential unintended benefits and harms was rare. Forty-eight reports described some sort of ‘community participation’ in the intervention; however, none of them explicitly stated that participation took place at the design stage, and only three describe a participatory evaluation process [47], [58], [61]. In contrast, local residents helped carry out the intervention in 45 studies, mainly by reporting vectors caught in their homes; in 20, the community was also involved in the pre-implementation phase.
Since Carlos Chagas historic paper [62], vector control has become the cornerstone of primary Chagas disease prevention [60], [63]. Pioneering attempts involved chemical (including cyanide gas) and physical means (including flamethrowers) [64]. The failure of DDT in controlling triatomines was followed by substantial optimism when HCH (lindane) proved successful in early trials in Brazil [65], [66], Argentina [67], and Chile [68]. The effectiveness of insecticide-based control kept improving as new chemicals and better formulations, with longer residual effects and lower toxicity, were introduced [40]–[42], [45], [69], [70]. Synthetic pyrethroids are now widely used and continue to be very efficient [71]–[75]; yet, recent research suggests that resistance may be widespread among some Triatoma infestans populations [16], [17], and insecticides are less effective in peridomestic environments [43], [76]. The top-quality report (in terms of sample size, design, and data treatment) we retrieved shows that peridomestic T. infestans foci reappear quickly after spraying (albeit with lower-density colonies) and that standard deltamethrin application with manual sprayers performs better than more sophisticated techniques [43].
Table 1 summarises the results of major reports on Chagas disease vector control [5],[18],[44],[57],[61],[63],[71]–[73],[77]–[113]. Overall, these studies unequivocally show that household insecticide spraying has successfully reduced infestation rates throughout Latin America, but also that reinfestation of dwellings by native vector species is common, spatially widespread, and temporally persistent. In many cases, the elimination of introduced populations was closely followed by the occupation of vacant niches by ‘secondary’ vector species, suggesting that the former had displaced the latter upon introduction [114], [115].
The ultimate measure of vector control effectiveness is the reduction of disease incidence. This is usually assessed through serological surveys [116]–[118], with an emphasis on the younger age classes. Domestic triatomine control has resulted in significantly lower seropositivity rates in every country and setting where this has been studied, but residual/re-emerging transmission is not uncommon [6], [18], [63], [97]–[99], [102]–[107], [119]–[126]. Infection rates in vectors [127] and non-human reservoir hosts [74], [128] also decrease sharply in areas under entomological control-surveillance, and this is crucial for reducing household-level disease transmission risk [129].
For the purposes of our quantitative appraisal, we defined ‘community participation’ in Chagas disease vector surveillance as simply the involvement of local residents in reporting the presence of suspect bugs in their households. This narrow definition is justified by (i) the need to use some measure of effect size that is (at least qualitatively) comparable across studies, (ii) the fact that vector detection is the primary purpose of entomological surveillance, (iii) the fact than most ‘participatory’ experiences are limited to stimulating bug notification, and (iv) the principle of parsimony, whereby simpler approaches to surveillance, if they are shown to work, enjoy better chances of effectively translating into policy and practice. Table 2 shows the main results of studies quantitatively comparing the effectiveness of vector notification by residents with either active bug searches by control programme staff (the standard approach) or different vector-detection devices (e.g., ‘sensor boxes’) [32], [85], [107], [130]–[132].
With a few exceptions, notification by residents performs obviously much better than active bug searches at detecting infestation foci, although the effect seems to be somewhat smaller in the peridomestic environment [32], [132] (Figure 3). Because notification costs less than active searches, these results are strong indication that it is probably much more cost-effective [20], [116], [133]. Vector-detection devices also seem to be largely outperformed by notification; the evidence is more limited in this case, but comparisons between detection devices and active searches (next subsection) suggest that notification by residents is also superior.
Several ‘passive’ vector surveillance methods have been devised and tested over the years. As defined here, they differ from the traditional, ‘active’ surveillance approach in that control programme agents do not search the whole residence to determine whether it is infested; instead, they rapidly check for bugs (or their traces) in a ‘detection device’. Table 3 summarises the main results of major comparative studies [20]–[22], [26]–[28], [130], [132], [134]–[140]. In general, the sensitivity of vector-detection devices does not seem to be superior to that of active searches, but (i) both methods appear to complement each other, with only one of them revealing infestation in many instances (see also ref. [141]), and (ii) the costs of the passive approach are, in general, lower (but see ref. [28]). Several studies with small sample sizes favour sensing devices, whereas the results of larger trials tend to show that they perform equally or worse than active searches (Figure 4). The evidence in relation to vector-detection devices remains therefore inconclusive, and further research is needed; below (Conclusions and outlook) we provide methodological suggestions to this end.
In the long run, Chagas disease prevention will depend on keeping households free of T. cruzi vectors [60], [116], [142]. Insecticide-based control campaigns have been extremely successful, but there is compelling evidence that persistent reinfestation of a fraction of treated households is the pattern to be expected across Latin America; reinfestation, in turn, can result in disease transmission re-emergence [18], [105], [106], [143], [144]. These well-supported findings clearly substantiate the view that long-term vector surveillance will be critical for the interruption of Chagas disease transmission [5], [7], [18], [35], [142], [145], [146].
Entomological surveillance primarily aims at detecting (then eliminating) household infestation foci; it thus allows for monitoring reinfestation trends in areas under control [5], [92], [94], [95], [147]–[151]. This is of fundamental importance for both (i) eliminating residual foci of introduced species targeted for local eradication and (ii) keeping reinfestation by native species at levels below disease transmission thresholds [73], [115], [152], [153]. We note, however, that ‘native’ vector species may be equally or more efficient than introduced ones at transmitting T. cruzi, and that even the most notorious ‘primary’ vectors, T. infestans and Rhodnius prolixus, are native (and reinfest treated households) [18], [143], [154]–[158] in their original ranges. Thus, entomological surveillance has a major role to play in most of Latin America even after introduced vector populations have been eliminated; in areas under surveillance, rapid diagnostic tests could be used to discover residual or re-emergent transmission foci [142].
But in order to attain these goals, vector detection must be as effective as possible, and the evidence we have reviewed shows that available vector-detection techniques all work far from perfectly. What would be, then, the best strategy to meet the permanent challenge of detecting reinfestation? Our appraisal yields strong support to the view that notification of suspect vectors by residents is the most sensitive among the several detection approaches tested to date – and that it is also probably the cheapest. Furthermore, the difference in performance seems to widen as vector population density declines, which is the typical situation in post-control settings.
Such an austere ‘participatory’ strategy signals the minimum degree of community involvement required to effectively enhance surveillance: residents are just asked to report suspect insects found in their homes, and a response is mounted by professional staff, often related to decentralised health services [142], [154], [159], [160], to eliminate infestation when needed [18], [145], [161]. An educational/communication component tailored to the social-cultural background of the community is obviously required to stimulate notification [4], [35], [162], [163], but our review suggests that very simple interventions can be effective enough. Perhaps the main challenge here is to sustain community awareness in the face of even rarer infestation events; continuous education, a clearly defined channel for communication between residents and control agents, and an opportune response to any notification (including those involving insects other than triatomines) are probably the key to long-term success [35], [73], [152], [159], [164]–[166].
This is not to say that more sophisticated approaches would not perhaps bring further benefits to people living under risk conditions. For instance, we found that most community-based experiences in Chagas disease vector surveillance are merely utilitarian, with little or no participation of the community in the design, planning, and evaluation of interventions. Effective involvement of all stakeholders along the whole process would no doubt foster true empowerment, and this could in itself result in improved health and living standards [33], [34], [167]–[171]. Still, we underscore that, in the absence of adequate resources for comprehensive community-based programmes, stimulating vector notification by residents may suffice to boost the efficiency of entomological surveillance across highly diverse ecological and socio-economic settings.
Finally, our review revealed that there is plenty of room for improvement of both methodological and reporting standards in the Chagas disease control/surveillance literature. In many cases, the results were reported incompletely and/or confusingly, sometimes precluding data extraction; in several instances, the data in the text, tables, and figures were incongruent. Indeed, just a few of the reviewed studies followed high-quality designs (e.g., with some sort of randomisation) and used sound analytical approaches, particularly in relation to the non-independence of observations; these reports tended to rely on small sample sizes and/or have limited spatial scope. Apart from the obvious need for using adequate design and analytical procedures, several guidelines for good reporting practices are readily available (e.g., the STROBE statement [172]); researchers and journal editors share the responsibility of improving the standards of published reports on Chagas disease control and prevention.
Indeed, we believe that the main limitations of our review relate to the quality of the original reports, even if the breadth of our appraisal probably lightens the effects of individual study drawbacks. We did not test formally for publication bias, but deem it unlikely that any major study was overlooked; the possibility that such a bias exists should however be kept in mind when interpreting our results, particularly in relation to vector-detection devices. In an attempt to overcome possible study-level biases, we made every effort to extract and re-analyse the data in each document, without taking reported results at face value, but this does not alleviate design or data collection bias. However, we are confident that our main findings (that reinfestation by triatomines is common and widespread and that householder involvement in vector reporting enhances surveillance) are not bias-induced artefacts. We also note that our assessment focused on the initial stage of surveillance – the detection of infestation foci. The responses triggered by detection events, the monitoring of infestation trends, and the analysis and dissemination of epidemiological data are also essential components of disease surveillance [173], but their appraisal was beyond the scope of this review.
Entomological surveillance is and will remain crucial to contain Chagas disease transmission; yet, the zoonotic nature of the parasite's life cycle implies that eradication is unfeasible [1]. The enduring challenge of household reinfestation by locally native vectors can only be met by means of horizontal strategies – and these work better when the community takes on a protagonist role. Even very simple forms of participation, such as encouraging vector notification by residents, can substantially enhance the effectiveness of surveillance. Control programmes should therefore incorporate community-based approaches as a strategic asset from inception; such approaches must include a timely, professional response to every notification, and would very likely benefit from a strengthened focus on community empowerment.
It must finally be emphasised that, in practice, vector detection failures are unavoidable, particularly when bug population density is low [174]. It may then be argued that infestation rates are virtually always underestimated and that, because these rates are the foremost indicator used in decision-making [175], imperfect detection can seriously misguide Chagas disease control programme management. We consequently suggest that a critical area for future research relates to the reliable estimation of vector detection probabilities. This is somewhat more difficult in the absence of a ‘gold-standard’ technique, but by no means unworkable: repeated-sampling approaches [176]–[178] readily yield detection probability estimates (with confidence intervals) that can in addition be modelled as a function of covariates – such as, for instance, alternative detection methods, different fieldwork teams, different vector species, or physically diverse ecotopes. These approaches have been successfully applied in wildlife [179] and disease ecology studies [180], [181], and can also help enhance Chagas disease vector research [182].
|
10.1371/journal.ppat.1006814 | The potent effect of mycolactone on lipid membranes | Mycolactone is a lipid-like endotoxin synthesized by an environmental human pathogen, Mycobacterium ulcerans, the causal agent of Buruli ulcer disease. Mycolactone has pleiotropic effects on fundamental cellular processes (cell adhesion, cell death and inflammation). Various cellular targets of mycolactone have been identified and a literature survey revealed that most of these targets are membrane receptors residing in ordered plasma membrane nanodomains, within which their functionalities can be modulated. We investigated the capacity of mycolactone to interact with membranes, to evaluate its effects on membrane lipid organization following its diffusion across the cell membrane. We used Langmuir monolayers as a cell membrane model. Experiments were carried out with a lipid composition chosen to be as similar as possible to that of the plasma membrane. Mycolactone, which has surfactant properties, with an apparent saturation concentration of 1 μM, interacted with the membrane at very low concentrations (60 nM). The interaction of mycolactone with the membrane was mediated by the presence of cholesterol and, like detergents, mycolactone reshaped the membrane. In its monomeric form, this toxin modifies lipid segregation in the monolayer, strongly affecting the formation of ordered microdomains. These findings suggest that mycolactone disturbs lipid organization in the biological membranes it crosses, with potential effects on cell functions and signaling pathways. Microdomain remodeling may therefore underlie molecular events, accounting for the ability of mycolactone to attack multiple targets and providing new insight into a single unifying mechanism underlying the pleiotropic effects of this molecule. This membrane remodeling may act in synergy with the other known effects of mycolactone on its intracellular targets, potentiating these effects.
| Buruli ulcer is a necrotizing skin disease caused by an environmental mycobacterial pathogen. The pathogenesis of this neglected tropical disease involves the production of a toxin, mycolactone, which spreads through the tissues, away from the infecting organisms. Mycolactone has pleiotropic effects on fundamental cellular processes, resulting in pronounced cytotoxicity and immunosuppressive effects that together drive progressive ulceration. The molecular mechanisms underlying its cellular effects have been partly deciphered, but multiple cellular targets have been identified. A literature survey revealed that most of the identified targets of mycolactone are membrane receptors residing in particular domains of the plasma membrane. Despite its lipid-like nature, mycolactone has been shown to be intracellular, implying that it can cross the plasma membrane. We describe here a surprising membrane-reshaping effect of mycolactone due to effects on lipid domain formation. By reversing lateral lipid segregation, mycolactone may disrupt the formation of domains with well-established roles in the regulation of cellular signaling pathways. This remodeling of the cell plasma membrane may underlie the molecular events enabling mycolactone to attack multiple targets.
| Buruli ulcer (BU) is the third most common human mycobacterial infection in the world, after tuberculosis and leprosy [1,2]. BU is a neglected tropical disease of the skin and subcutaneous tissue caused by an environmental pathogen, Mycobacterium ulcerans (M. ulcerans). This disease, which can affect all age groups and both sexes, is commonest in West Africa and parts of Australia, but has been reported in over 30 countries worldwide [3,4]. These painless ulcers affect at least 5,000 patients per year and are thought to be heavily underreported [5]. Infection with M. ulcerans results in persistent severe necrosis with no acute inflammatory response. The ulcer begins as a painless nodule or papule on the skin, which, if left untreated, progresses to massive ulceration that may cover 15% of the skin of the patient, resulting in significant morbidity [5,6]. BU is not lethal, but patients may suffer lifelong disfigurement, functional impairment and disability unless the infection is recognized and treated at an early stage.
M. ulcerans pathogenesis is mediated by a necrotizing immunosuppressive toxin, mycolactone (S1 Fig). This lipid-like polyketide macrolide has been identified as the main virulence factor produced by M. ulcerans and is responsible for the skin lesions and tissue necrosis [6,7]. After its production [8,9], this diffusible toxin is excreted in vesicles derived from the bacterial membrane and enriched in extracellular matrix, which acts as a reservoir of the toxin [10]. In vitro, mycolactone has been shown to localize in the cytosol of cultured murine fibroblasts, through non-saturable and non-competitive uptake in the presence of excess mycolactone [11,12]. Mycolactone has also been reported to accumulate in a time- and dose-dependent manner in the cytoplasm of human epithelial cells and lymphocytes, but not in the plasma or nuclear membranes of the cell [13]. These findings suggest that mycolactone can diffuse across cell membranes by non-cell-specific passive diffusion to reach its intracellular targets [11,14].
Mycolactone A/B (a 3:2 ratio of Z-/E-isomers of the C4’-C-5’ bond in the long “Southern” polyketide side chains, S1 Fig), which is produced by the most virulent strains of M. ulcerans, has been shown to have pleiotropic effects on fundamental cellular processes, such as cell division, cell death and inflammation, depending on toxin dose and exposure time [14,15]. Exposure to pure mycolactone is cytotoxic for many cell lines, but the dose and exposure required for cell death are highly variable [15]. Early studies on cell lines suggested a role for mycolactone in cell-cycle arrest in the G1/G0 phase and apoptosis [7,16]. However, recent studies have suggested that anoïkis, due to cytoskeleton rearrangements, leading to changes in cell adhesion and detachment, is a much more likely mechanism of cell death in vivo [3,17]. By inducing changes to the cytoskeleton and disrupting tissue structure, this toxin compromises cell structure and homeostasis through the impairment of extracellular matrix biosynthesis [18]. In addition to its cytotoxicity, mycolactone has immunosuppressive activity, resulting in a lack of local inflammation despite extensive tissue damage, together with inhibition of the local immune response [19–23]. At low concentrations, this molecule has been found to be a powerful analgesic, due to its stimulatory effect on the angiotensin receptor [24]. These effects may account for the painlessness of BU lesions. The precise molecular mode of action of mycolactone in eukaryotic cells remains unclear, but a number of cellular targets have been identified. A literature survey revealed most of these targets to be membrane receptors residing in ordered plasma membrane nanodomains known to modulate the functionalities of membrane proteins [25,26].
Mycolactone can impair the migration of naïve T cells to peripheral lymph nodes [27], where they make contact with antigen-presenting cells during T-cell receptor activation. This alteration of T-cell homing is accompanied by a decrease in L-selectin receptor (CD62-L) levels. The downregulation of this receptor normally involves proteolytic cleavage upon stimulation, but the cleavage of L-selectin seems to involve membrane microdomains, which act as a signaling platform [28]. Similarly, the chemokine receptors involved in T-cell inflammatory responses also reside in membrane domains and, the depletion of cholesterol from membranes decreases chemokine binding and abolishes chemokine receptor signaling [29].
Another effect of mycolactone A/B is hyperactivation of the Src-family kinase, leading to the depletion of intracellular calcium and a downregulation of T-cell receptor (TCR) expression, limiting the T-cell response to stimulation and potentially contributing to apoptosis [3,14]. This hyperactivation is initiated by the relocalization of Lck in the microdomains of the plasma membrane, triggered by the action of the toxin [30]. Mycolactone has been reported to inhibit angiotensin II binding, in a dose-dependent manner, and to elicit signaling through human type 2 angiotensin II receptors (AT2Rs), leading to a potassium-dependent hyperpolarization of neurons, accounting for the painlessness of BU lesions [24]. AT2R, like AT1R, is a G protein-coupled receptor (GPCR). Microdomains (both lipid rafts and caveolae) have been reported to be involved in regulating GPCR signaling, by affecting both signaling selectivity and coupling efficacy [31,32].
Mycolactone has recently been shown to modulate Wiskott-Aldrich syndrome protein (WASP) and neural WASP (N-WASP), two members of a family of scaffold proteins that transduce various endogenous signals in dynamic remodeling of the actin cytoskeleton [17]. In immune cells, WASP regulates ordered lipid domain dynamics during immunological synapse formation, which involves clustering of the microdomains of the plasma membrane for optimal T-cell activation. WASP, which is recruited to lipid domains immediately after TCR stimulation, is required for the movements of these microdomains [33]. By disrupting WASP autoinhibition [17], mycolactone can hijack actin-nucleating factors, leading to uncontrolled activation of the ARP2/3-mediated assembly of actin, and a deregulation of lipid domain dynamics. Similarly, mycolactone provokes a disruption of the protein C anticoagulant pathway, with a depletion of thrombomodulin (TM) receptors at the surface of endothelial cells [34]. Nevertheless, the receptors of the protein C activation and activated protein C (APC) signaling pathways are colocalized in the lipid microdomains of endothelial cells [35,36].
Finally, it has recently been reported that mycolactone inhibits the function of the Sec61 translocon [37–39], a transmembrane channel located in the endoplasmic reticulum (ER) membrane [40]. This ubiquitous complex is responsible for cotranslational protein translocation, a universally conserved process in the biosynthesis of secretory and membrane proteins that operates for most of the 30–50% of mammalian proteins carrying a canonical signal peptide [41]. In investigations of transmembrane proteins (TNF), monotypic proteins (COX-2) and conventionally secreted proteins (IL-6), Hall et al. showed that mycolactone prevents ER protein translocation, with the proteins concerned being translated in the cytosol, where they are marked for rapid destruction by the proteasome. In this way, mycolactone causes a selective ~30% decrease in membrane-associated proteins and prevents the production of the vast majority of N-glycosylated proteins [37,38]. Cholesterol and sphingolipid levels are lower in the ER than in the plasma membrane and other organelles, but it has been suggested that ER membranes nevertheless contain lipid domains [42,43]. The fractionation of rough ER integral membrane proteins with 0.18% Triton X-100 (similar to the treatment of cytoplasmic membranes with 1% Triton X-100, which has successfully revealed the presence of lipid domains in the cytoplasmic membrane) showed that the 0.18% Triton X-100 fraction contained mostly ER-resident proteins, including, in particular, the Sec61alpha subunit [44], the central transmembrane component of the sec 61 ER translocon targeted by mycolactone [38,39,45].
Thus, mycolactone has diverse complex effects on a range of cells and tissues, and the underlying mechanism unifying its pleiotropic effects seems to be its action through microdomain-associated proteins [14]. In this study, we aimed to characterize in more detail the effects of pure mycolactone on biological membranes, focusing, in particular, on the effects of this toxin on microdomain segregation. Indeed, no molecular-scale description of the effects of this toxin on the cell plasma membrane before it reaches its cellular targets, most of which are located in the ordered plasma membrane nanodomains, has ever been reported.
We investigated the capacity of mycolactone to interact with membranes and its effects on lipid organization when crossing the membrane, with several biophysical techniques, including Langmuir monolayers, which we used as an in vitro model of cell membranes, together with fluorescence and Brewster angle microscopy. Langmuir monolayers consist of supramolecular lipid films that form at an air-buffer interface. They can mimic biological membranes and are, thus, attractive membrane models, because the thermodynamic relationship between monolayer and bilayer membranes is direct, and monolayers overcome, independently of their lipid composition, the limitations associated with the regulation of lateral lipid packing encountered in model bilayer systems [46]. They are widely used in studies of peptide or membrane probe/lipid interactions [47–52], and in studies of membrane-protein association [53–61]. Brewster angle microscopy (BAM), which was used for the in situ characterization of Langmuir monolayers, provides additional information about membrane morphology and lipid organization at the air-water interface [62–64]. In our system, we used a lipid composition closely resembling that of the plasma membrane, including among others, 33% sphingomyelin (SM) and 19% cholesterol (Chol), which was considered to be a biologically normal concentration (natural membranes contain 5–50 mol% cholesterol [65,66]). Using this experimental approach, we demonstrated marked effects of mycolactone on membranes, and were able to visualize, for the first time, the capacity of this molecule to disrupt membranes at the molecular level.
We investigated the interaction of mycolactone with biological membranes and evaluated the possible influence of lipid composition (i.e., with or without cholesterol), by exploring the binding properties of the toxin with a model membrane reconstituting lipid monolayers at the air/water interface. These so-called Langmuir monolayers are half-membrane models [46], and they can be used not only to characterize protein–membrane interactions, but also to determine the mechanism of action of bioactive molecules on cell membranes. We chose this model for study on the basis of its simple experimental design, the possibility of changing lipid composition easily and its suitability for evaluations of the membrane insertion capacity of membranotropic molecules [47–61,67].
We characterized the surfactant properties of mycolactone, by evaluating its interfacial behavior at the air/buffer interface and in the absence of lipids.
Experiments without lipids at the air/buffer interface can be used to determine: i) the concentration at which amphiphilic molecules saturate the lipid-free interface (i.e., surface saturation concentration) and ii) the concentration range minimizing aggregation and, therefore, useful for experiments. It is widely accepted that the analytical concentrations to be injected into the subphase for subsequent molecule/lipid interaction analyses should be based on such pre-evaluations and lower than the surface saturation concentration, to prevent artifacts due to molecule aggregation [51,67–69].
The surface saturation concentration was determined by tensiometry [70,71]. Various mycolactone concentrations, from 60 nM to 6 μM, were injected into the PBS subphase. For each concentration, the adsorption of mycolactone at the air/buffer interface was monitored by continuous surface pressure measurement until the equilibrium value, πe, was reached. The curve of πe as a function of mycolactone concentration rapidly increased to reach a plateau at 34 mN/m (Fig 1). At this surface pressure, the interface was saturated with mycolactone molecules, regardless of the concentration of the toxin in the subphase. The surface saturation concentration of mycolactone was then determined at the start of the plateau, and was found to be 1 μM.
Brewster angle microscopy (BAM) images recorded at πe with a final mycolactone concentration of 0.6 μM (a), 1.2 μM (b) or 6 μM (c) (Fig 1, Inset) confirmed the ability of this molecule to accumulate in a concentration-dependent manner at the air/PBS (pH 7.4) interface, and to form a very thick film (3.7 ± 0.3 nm thick) at very high concentrations (6 μM). At a concentration of 0.6 μM, below the apparent surface saturation concentration of 1 μM, mycolactone formed a homogeneous interfacial film, but with some bright nuclei also visible (spots, Fig 1, Inset a). These bright nuclei resulted from aggregate formation, as demonstrated by dynamic light scattering (DLS) for mycolactone solutions in PBS pH 7.4 in S1 Appendix.
Thus, mycolactone displayed surfactant properties at a nude interface. To prevent the association of molecules into aggregates in the next experiments, we used a low concentration of mycolactone, 60 nM. At this concentration, mycolactone interacts with lipids as a monomer (see S1 Appendix).
The aim of this study was to analyze the interaction of mycolactone with biological membranes. We studied two membrane models: i) a monolayer consisting of a lipid mixture resembling that of the plasma membrane (given in mol%) [72–75]: 39% POPC, 33% SM, 9% POPE, 19% Chol (mixture 1), and ii) a monolayer with the same lipids but without cholesterol (given in mol%): 48% POPC, 41% SM, 11% POPE (mixture 2). Cholesterol is known to regulate lipid segregation in plasma membranes [25,26]. Mycolactone receptors have been reported to be located in ordered plasma membrane microdomains. We therefore investigated the effects of this particular membrane lipid on the ability of mycolactone to bind to membranes.
We first studied the interfacial properties of the two monolayers alone, and the impact of cholesterol on lipid organization in particular, at 20 and 25°C.
Whatever the temperature, the π-A isotherms of mixture 1 (Fig 2A) showed the monolayer to be in liquid-condensed (LC) phase throughout compression. The beginning of the steep rise started at a molecular area of 70 Å2, and the monolayer was compressed up to a lateral pressure of πcoll = 45 mN/m, corresponding to collapse. The molecular area at collapse, Acoll, was 31 Å2. This area was smaller than expected for two fatty acyl chains of phospholipids; the area per CH2 chain in a close-packed configuration is approximately 20 Å2 [76]. This discrepancy can be explained by the condensing effect of the cholesterol. The molecular area of a pure expanded monolayer of POPC (the major component of mixtures 1 and 2) at a lateral pressure of πcoll = 40 mN/m is ~40 Å2 at 20 or 25°C, consistent with the Tm value (-4°C) of POPC (S2 Fig). The addition of 19% cholesterol to the POPC monolayer, resulted in the same Acoll for the 81% POPC/19% cholesterol mixture at 25°C, but this area decreased to ~32 Å2 at 20°C. This suggests that the presence of 19% cholesterol lead to extensive condensation of the POPC monolayer in the liquid-expanded state (S2 Fig). Cholesterol has been shown to dehydrate lipid bilayers, resulting in lipid condensation [77]. The much lower level of condensation observed in the presence of mixture 1 (~37 Å2 at π = 40 mN/m) in terms of the area of POPC (~40 Å2) may be due to the presence of 9% POPE in mixture 1, at least partly preventing the condensing effect of cholesterol. BAM images recorded during compression revealed that the small lipid domains (shown in light gray) present at the start of compression (Fig 2A, image A, white arrows) increased in size and coalesced (Fig 2A, image B) to form a homogeneous interfacial film at the end of compression (Fig 2A, image C), regardless of temperature. This observation is consistent with the behavior of a condensed monolayer. Finally, the monolayer was homogeneous at 30 mN/m (Fig 2A, Image C), the lateral surface pressure reported for biological membranes [78].
A different pattern was observed for the isotherms of mixture 2 (Fig 2B), with the monolayers displaying a phase transition at both temperatures. Surface pressure began to increase at a higher molecular area, 98 Å2. BAM images taken at 20°C revealed the presence of holes (grayscale level identical to the background) in a continuous lighter phase (Fig 2B, image A, white arrows) for surface pressures below 3 mN/m. These holes gradually disappeared during compression until a short plateau was reached at about 5 mN/m. Beyond this point, the monolayer was homogeneous (Fig 2B, image B) and in a liquid-condensed state until collapse (πcoll = 45 mN/m; Acoll = 25 Å2). At 25°C, the phase transition, which could be attributed to the liquid-expanded/liquid-condensed (LE/LC) transition phase of monolayers incorporating SM [79,80], was attenuated. Consequently, the monolayer remained homogeneous (absence of holes at low surface pressures) throughout compression (Fig 2B, image C), until collapse (πcoll = 44 mN/m; Acoll = 30 Å2). In the absence of cholesterol, no lipid domains were observed in the monolayers. The apparent condensing effect observed for mixture 2 (S2 Fig) relative to the pure monolayer of POPC in the expanded state is due to the presence of 41% SM, a high-melting lipid (Tm = 41.4°C). The shift in Acoll values observed when the temperature was increased from 20 to 25°C could be explained, in all cases, by the disordering effect of the higher temperature on acyl chain packing, tending to fluidize the monolayer.
Superimposition of the isotherms recorded at 20 and 25°C (Fig 2C) highlighted the effect of cholesterol on the condensation state of the monolayer. At surface pressures below 10 mN/m, the isotherms of mixture 1 (with cholesterol) were shifted towards lower molecular areas than those of mixture 2 (without cholesterol). By contrast, at high surface pressures (above 25–30 mN/m), the isotherm of mixture 1 at 20°C was shifted towards larger areas than those of mixture 2 at the same temperature. This clear difference between the two mixtures was consistent with the modulation of membrane fluidity by cholesterol, through modification of the ordering of lipid acyl chains [81]: cholesterol tends to condense fluid phases (i.e., it increases the lipid chain ordering of the liquid-crystalline disordered phase) and to fluidize condensed phases (i.e., it decreases the lipid chain ordering of the solid-ordered phase) [82–85]. A comparison of the four isotherms also revealed that, in the absence of cholesterol at 20°C, the monolayer was extremely condensed. This condensation state may be directly due to the presence of 41% SM in mixture 2. Indeed, sphingolipids generally form a solid gel phase and are fluidized by sterols, which interact preferentially with them in the membrane [66,86]. Furthermore, the domains observed in mixture 1 (Fig 2A, image A) were probably characteristic of the liquid-ordered phase resulting from a ternary mixture of a high chain-melting lipid (like SM) and a low chain-melting lipid (like POPC) with cholesterol, and preferential interactions between Chol and SM [65,79,82,87–92].
We analyzed the interaction of mycolactone with monolayers at a working surface pressure of 30 mN/m, to mimic the lateral pressure of biological membranes [46,78]. As a control, and to decipher the effect of mycolactone more effectively, we checked the stability over time of the mixed monolayers in the presence of ethanol, the solvent used for mycolactone. For this purpose, we injected a volume of ethanol equivalent to that used for mycolactone solution (4.45 μL) into the subphase underneath the stabilized monolayer at 30 mN/m. We then recorded changes in surface pressure over a period of about seven hours. At 20°C, the monolayers were highly stable, with only small surface pressure variations (± 2 mN/m) over time (S3 Fig). At 25°C, a greater variation of surface pressure was observed (from ‒2 to ‒5 mN/m), possibly due to subphase evaporation.
We used BAM images for simultaneous characterization of the morphology and lipid organization of the mixed monolayers at 20°C (Figs 3A and 4A, rows a) and 25°C (Figs 3B and 4B, rows a). All the monolayers were homogeneous after one hour of relaxation, just before injection. After injection, the changes in monolayer morphology differed between temperatures and membrane lipid compositions.
At 20°C, the mixed films displayed a phase segregation that differed according to the presence or absence of cholesterol. In the presence of cholesterol (Fig 3A, row a), lipid organization gradually changed, after about 3 h, with the formation of circular domains of an expanded fluid phase (dark phase) trapped within a more condensed phase (white phase). The same change in monolayer morphology was obtained without the injection of ethanol (S4 Fig, row a). This segregation, observed at 20°C, and leading to a new thermodynamic equilibrium with no loss of stability, could therefore be attributed to preferential interactions between cholesterol and the high-melting lipid SM in the mixed monolayer [65,82,83,87], with an expulsion of low-melting lipids such as POPC/POPE, resulting in the formation of round domains of fluid phase, as already reported for ternary mixtures of PC/SM/Chol [81,88,89]. In the absence of cholesterol, ordered domains appeared earlier, from the start of the experiment, and progressively grew in the form of “stars” (bright clusters, Fig 4A, row a). These domains resembled the condensed domains observed in the liquid-expanded/liquid-condensed (LE/LC) transition phase during the compression of a pure monolayer of SM on a PBS subphase (pH 7.4) at 20°C (S5 Fig). These findings suggest that SM molecules retain their ability to segregate over time in the mixed monolayer, but only in the absence of cholesterol.
Conversely, no segregation occurred at the higher temperature. Indeed, at 25°C, in the presence (mixture 1 –Fig 3B, row a) or absence (mixture 2 –Fig 4B, row a) of cholesterol, the two monolayers remained homogeneous throughout the entire experiment.
We investigated the membrane-binding properties of mycolactone and evaluated the effect of this interaction on lipid organization in mixed films, by injecting a solution of mycolactone in ethanol into the subphase at a final concentration of 60 nM, beneath the monolayers of mixture 1 (with cholesterol) or mixture 2 (without cholesterol), compressed at an initial surface pressure πi of 30 mN/m.
Upon injection, regardless of lipid composition and temperature, the interaction of mycolactone with the monolayer resulted in a rapid increase in surface pressure up to ~36 mN/m within the first 15–20 minutes (Fig 5). After a stabilization period of about 1–1.5 h, π gradually decreased over time. The absence of cholesterol clearly did not affect the ability of mycolactone to penetrate into the monolayer; it simply delayed the decrease in surface pressure.
BAM images were taken before and after mycolactone injection, throughout the adsorption period (Figs 3 and 4, rows b). In all cases, monolayers were homogeneous at the initial surface pressure of 30 mN/m. Mycolactone injection modified lipid segregation in the monolayers independently of temperature, but differently according to the presence or absence of cholesterol in the monolayers.
For mixture 1 at 20°C (Fig 3A, row b), lipid organization changed rapidly over the first 15 min towards the formation of circular domains in a more condensed state (light gray phase), trapped within a less condensed phase (dark gray phase). This reorganization is essentially the opposite of the organization observed with the pure monolayer (Fig 3A, row a). A similar pattern was observed if the mycolactone was injected at the apparent saturation concentration of 1 μM (S4 Fig, row b). The time required for mycolactone to reverse the segregation pattern in membranes (15 minutes) corresponds to the time required for toxin penetration into the monolayer until stabilization. At 25°C (Fig 3B, row b), this segregation pattern occurred 3 h after injection, whereas no segregation was observed for the control monolayer at 25°C. Highly luminous structures (Fig 3, row b, t = 15 min or 85 minutes at 20°C and t = 30 min at 25°C), similar to those observed for mycolactone at the air/buffer interface (Fig 1, image b) were also observed. This feature indicated the presence of the toxin within the monolayer in the presence of cholesterol.
For mixture 2 at 20°C, no star-shaped domains were visible, contrasting with observations for the monolayer alone. Very small domains (small bright dots) became visible much later, 4 h after mycolactone injection (Fig 4A, row b, white arrows). At 25°C, no significant change in the morphology of the monolayer relative to the control was observed upon mycolactone injection (Fig 4B). Again, only the presence of bright objects corresponding to mycolactone at different time points (Fig 4B, rows b) attested to the interaction of the toxin with the monolayer. In the absence of cholesterol, the presence of the toxin within the monolayer therefore prevented SM molecules from aggregating, thereby fluidizing the condensed and extremely rigid mixture 2 monomolecular film (Fig 2B). At 25°C, the monolayer was fluid enough to prevent SM aggregation, and this attenuated the potential fluidizing effect of the mycolactone.
We analyzed the influence of lipid organization on the membrane-binding properties of the toxin further, by investigating the effect of initial surface pressure on the interaction of mycolactone with the monolayers. For this purpose, we monitored the maximal increase in surface pressure Δπmax immediately following toxin injection at various πi values, ranging from 5 to 30 mN/m. This relationship has been widely used to assess lipid-protein interactions and to distinguish between electrostatic and hydrophobic interactions [53,55,57,58,60,61,68,93].
The Δπmax = f(πi) plot shown in Fig 6 was used to evaluate the binding parameters of mycolactone on both types of Langmuir monolayers. Linear extrapolation to an increase in surface pressure of zero (Δπmax = 0) can be used to determine i) the maximum insertion pressure (MIP), reflecting the influence of initial lipid packing density on the ability of the molecule to penetrate into the monolayer, and ii) the synergy factor "a" [50,56,93,94]. This factor, first described by Salesse et al. [56,94], provides insight into the mechanisms governing the interaction with lipid monolayers. A positive a value indicates favorable interactions, as further demonstrated by MIP values exceeding the estimated membrane lateral pressure (~30 mN/m). A null synergy factor reveals a stationary state, with no favoring or disfavoring of membrane binding. A negative synergy factor indicates unfavorable binding to the monolayer, corresponding to a repulsion of the molecule as a function of the compactness of the monolayer. Here, MIP and a provided useful information about the effect of lipid composition on the ability of mycolactone to interact with membranes.
For mixture 1, pressure variation profiles were similar at the two temperatures, with a linear decrease as a function of initial surface pressure πi. MIP and the a synergy factor were above 30–35 mN/m and positive, respectively, at both temperatures. These findings are consistent with strong insertion/penetration into the interfacial film and favorable interactions between mycolactone and the monolayer (Fig 6A and 6B). Furthermore, both MIP and a values were higher at 20°C (MIP = 45.9 ± 2.8 mN/m; a = 0.58 ± 0.02) than at 25°C (MIP = 38.7 ± 1.1 mN/m; a = 0.32 ± 0.02), suggesting that, in the presence of cholesterol, decreases in temperature leading to a rigidification of the monolayer may favor the interaction of mycolactone with the mixed film.
By contrast, the curve profiles for mixture 2 differed considerably between temperatures. At 20°C (Fig 6C), the plot obtained was split into two distinct phases: an initial plateau, for which Δπmax remained constant at πi values below 17.5 mN/m, with a synergy factor of 0.93 ± 0.09, and a second phase in which Δπmax decreased at πi values greater than 17.5 mN/m, associated with an MIP value of 34.8 ± 1.6 mN/m and a synergy factor close to 0 (a = 0.01 ± 0.07). Such ‘biphasic’ behavior was recently reported by Hädicke and Blume for the binding of small cationic peptides to anionic phospholipid monolayers [95], and may be related to the physical state of the monolayer. As shown by these authors, incorporation, through hydrophobic interactions, into the loosely packed monolayer in the LE phase can lead to a constant Δπ value, depending on the nature of the lipids making up the monolayer. By contrast, in the LC phase, Δπ and πi display an inverse linear relationship, due to lipid condensation. For mixture 2, the monolayer displayed a LE/LC phase transition, as shown on the isotherms (Fig 2B), with the presence of holes in the loosely packed monolayer, as revealed by BAM (Fig 2B, image A). We therefore suggest that behavior similar to that proposed for small hydrophobic peptides may account for the unusual results obtained with mycolactone. Indeed, this toxin is also a small hydrophobic molecule (MW: 743.021), and, when it penetrates into a loosely packed monolayer in the LE phase by hydrophobic interactions, it triggers no increase in surface pressure because the monolayer is too weakly compressed (loosely packed) and the molecular area of the toxin is too small to cause lipid condensation at the interface. In addition, mycolactone was probably able to fill the space, i.e., the holes observed in the monolayer, due to its own surface activity, leading to an absence of surface pressure variation (Δπmax remained constant) as long as the monolayer was weakly compressed.
Beyond 17.5 mN/m, the monolayer was sufficiently tightly packed to attain its condensed state (observed on the isotherm Fig 2B), yielding a negative slope of the Δπmax = f(πi) plot (Fig 6C, part 2), with a synergy factor close to 0 (a = 0.01 ± 0.07). This value indicates that, even in a stationary state in which mycolactone was able to penetrate the monolayer at 20°C, no specific interactions (either favorable or unfavorable) occurred between mycolactone and lipids. The decrease in the ability of the molecule to penetrate the membrane was therefore entirely due to the physical condensation of the monolayer as a result of the increase in lipid packing density during compression [56,94].
At 25°C (Fig 6D), the curve profile and the MIP (41.8 ± 2.2 mN/m) were similar to those obtained for mixture 1 at the same temperature, but the a value (0.45 ± 0.02) was different. At the higher temperature (25°C vs. 20°C), the monolayer was more fluid, as revealed by the shift of the π-A isotherm towards larger molecular areas due to the disordering effect of the higher temperature on acyl chain packing (Fig 2B), and favorable interactions occurred between the toxin and the monolayer. As previously observed for π-A isotherms (Fig 2A and 2B), the effect of temperature on monolayer fluidity was more pronounced for mixture 2. This difference may account for the difference in synergy values.
At 25°C, the interaction of mycolactone with monolayers seems to be governed by greater monolayer fluidity. However, the presence of cholesterol in the monolayer enabled the mycolactone to penetrate into a more condensed monolayer. Indeed, if mycolactone insertion were regulated solely by monolayer fluidity (as observed at 25°C), then its insertion should increase with temperature, which was found to be the case in the absence (Fig 6C and 6D), but not in the presence of cholesterol (Fig 6A and 6B). Conversely, the binding parameters (MIP and synergy) were highest at 20°C in the presence of cholesterol (Fig 6A). As the synergy factor measures sensitivity to lipid acyl chain packing, we can conclude that the insertion of mycolactone into the monolayer was favored by the presence of cholesterol at lower temperatures, which favored increases in monolayer rigidity.
We investigated the effects of mycolactone on lipid segregation in the monolayer in the presence of cholesterol, by performing the same experiments (πi = 30 mN/m, 20°C) by fluorescence microscopy (FM), with mixture 1 labeled with TopFluor Cholesterol probe. This molecular probe can be used to study intracellular cholesterol dynamics, because its diffusion in the plasma membrane is free and unhindered [96]. In this context, the fluorescent domains observed were, thus, those containing the TopFluor Cholesterol molecule. In our study, the use of this fluorescent marker made it possible to track the localization of cholesterol in the monolayer and its distribution in domains.
Lipid segregation in the control monolayer was observed 3 h after ethanol injection (4.45 μL), with the appearance of circular dark domains trapped within a light phase (Fig 7, row a). Segregation occurred more rapidly in the presence of the toxin (in 1h15), but with a pattern opposite to that in the control, with the formation of circular fluorescent domains within a dark phase (Fig 7, row b). In ternary mixtures, SM is known to interact preferentially with cholesterol to form domains of liquid-ordered (Lo) phase, corresponding to a phase intermediate between the liquid-condensed phase (Lc) and the fluid liquid-expanded (Le) phase [91,92]. It can therefore be inferred from our measurements that the green fluorescent areas correspond to domains of liquid-ordered phase, whereas the dark areas correspond to domains of fluid phase [96]. Thus, these FM experiments yielded results identical to those obtained with BAM for the pure monolayer or after the injection of mycolactone (Fig 3A), in support of our conclusion. A similar correlation between the results of BAM and FM was reported in another recent study [79].
By modifying the interactions between SM molecules, and, probably, between SM and cholesterol, through physical insertion in the monolayer (the toxin impeded the segregation of SM in mixture 2 at 20°C without specific interaction, a = 0), mycolactone reversed the segregation of the Lo phase in the monolayer. Thus, mycolactone probably interacts preferentially with the Lo phase, which is intermediate between the highly condensed phase of SM and the fluid phase of POPC. This conclusion is consistent with the results obtained for the two mixtures at 20°C, and with the synergic interaction observed only in the presence of cholesterol (a>0 for mixture 1 at 20°C–Fig 6A). This preferential interaction with the Lo phase may also explain why the presence of cholesterol enhanced the penetration capacity of mycolactone at lower temperatures, which were associated with lower levels of fluidity (highest MIP and a value for mixture 1 at 20°C–Fig 6A).
The adsorption kinetic (π-t) curves showed a progressive decrease in surface pressure after the injection of mycolactone, at both temperatures (Fig 5). This suggests that the toxin affects monolayer stability and may have a detergent-like effect. Indeed, detergents are amphiphilic molecules with surfactant properties that can solubilize lipids, depending on membrane phase and composition [97].
We tested this hypothesis, by performing interaction assays with a final concentration of 60 nM Tween 20 or Triton X-100 (i.e., non-ionic detergents) at 20°C, with a monolayer of mixture 1 compressed at a πi of 30 mN/m. Tween 20 is known to solubilize lipid membranes regardless of their aggregation state [98], whereas Triton X-100 solubilizes the liquid-disordered (Ld) phase but not the liquid-ordered phase (Lo) [81,88,97,99]. The kinetic curves recorded were then compared with that obtained with mycolactone (Fig 8A).
When Tween 20 was injected beneath the monolayer, the surface pressure π increased to a plateau value of about 34 mN/m, gradually decreasing thereafter (Fig 8A). Investigations of monolayer morphology by BAM (Fig 8C, row b) revealed that, upon interaction, lipid organization shifted towards the formation of circular condensed domains (white phase) trapped within a fluid phase (dark phase), as in the case of mycolactone (Fig 8C, row a). However, this lipid reorganization occurred 2h40 after detergent injection, later than for the toxin (Fig 8C, row b).
The surface pressure π remained constant after Triton X-100 injection (Fig 8A). The absence of an effect on the surface pressure stability of mixture 1 was therefore compatible with an absence of Lo phase solubilization (a property of Triton X-100). However, BAM images revealed that Triton X-100 provoked the same pattern of lipid segregation as mycolactone (condensed domains in a fluid phase), but with the same time-shift (2h50) as for Tween 20 (Fig 8C, row c). Similar changes in the morphology of monolayers containing SM and cholesterol upon interaction with Triton X-100 have already been reported [100], and this detergent, which can induce Lo/Ld phase segregation in a typical raft-like ternary mixture, was recently described as a potent membrane-reshaping agent [97,99].
All these findings reveal, therefore, that the presence of 60 nM detergent or toxin in the subphase modifies lipid segregation in the POPC/SM/POPE/Chol monolayer in a manner opposite to that in the control.
In these experiments, both the detergents and the toxin were injected into the subphase at the same final concentration (60 nM). However, the solubilizing action of a detergent depends on its critical micellar concentration (CMC) and on the detergent/membrane ratio at which it is used [101]. The CMC of Tween 20 is 50–60 μM [102], and that of Triton X-100 is 0.2 mM, at 25°C [98,103]. For mycolactone, the apparent saturation concentration of 1 μM determined by tensiometry (Fig 1A) may be considered equivalent to an apparent CMC. Under these conditions, the two detergents and mycolactone may behave differently at an effective (active) concentration of 60 nM. Thus, to maintain a constant ratio between the effective concentration injected in the subphase and the CMC, we further investigated the effect of each detergent on monolayer stability with an “effective concentration/CMC ratio” of 0.06 (i.e., 60 nM divided by 1 μM, as for mycolactone).
To respect this new experimental constraint, Tween 20 or Triton X-100, at final concentrations of 3.6 μM and 12 μM, respectively, were injected in the subphase of the mixture 1 monolayer (Fig 8B). The use of these new detergent concentrations had no significant effect on interaction kinetics. The main difference concerned the formation of condensed domains, which occurred more rapidly at this constant “effective concentration/CMC ratio” of 0.06 than following the injection of a 60 nM solution: 1h30 for Tween 20 (Fig 8C, row d) and 1h40 for Triton X-100 (Fig 8C, row e). However, these times remains longer than the 15 min for lipid segregation triggered by mycolactone with the same constant ratio of 0.06 (Fig 8C, row a). At an “effective concentration/CMC ratio” of 0.06, at which each molecule acts as a monomer, the same effects on lipid morphology were observed by BAM for Tween 20, Triton X-100 and mycolactone. Under these conditions, the toxin penetrated and destabilized the monolayer just like the detergents, but more efficiently, acting as a reshaping agent [97,99]. As shown in this study, mycolactone preferentially binds to monolayers containing cholesterol, and this interaction induces a destabilization of the Lo phase by fluidizing the monolayer, modifying the preferential interactions between SM and cholesterol, like a detergent, but not like Triton X-100 [88,97]. Similar results were recently reported for glycyrrhizin, a molecule of the saponin class extracted from plants and recognized as a natural detergent, which causes membrane perturbations after its migration toward SM/sterol-enriched membrane domains [79].
Mycolactones form a family of highly related macrocyclic polyketides identified as the primary virulence factors responsible for Buruli ulcer (BU), a neglected tropical disease of the skin and subcutaneous tissue caused by the environmental human pathogen Mycobacterium ulcerans [4]. Despite the wealth of research describing the pathogenic mechanism [14,34,37], there is still no molecular explanation of the necrosis seen in the ulcers, over and above cytopathic activity, and the immunomodulatory or analgesic properties of mycolactone [13,24]. The effects of the toxin on the cell plasma membranes they cross have never been described.
In this study, we investigated, for the first time, the membrane-binding properties of mycolactone, with Langmuir monolayers, which were used as membrane models for the fine analysis of membrane binding kinetics. We chose to use this system because its experimental design is simple and it can be adapted for investigations of the molecular insertion properties of membranotopic molecules. We studied monolayers with a lipid composition of 39% POPC, 33% SM, 9% POPE and 19% cholesterol (% mol), which was considered representative of the plasma membrane [72–75]. These monolayers were compressed at an initial surface pressure (πi) of 30 mN/m, a value considered representative of the lateral pressure of biological membranes [46,78]. The results obtained provided new insight into the mechanism of membrane interaction and the effect of the toxin on membrane lipid organization.
By studying the behavior of the pure toxin at an air/buffer interface, we found that mycolactone, like detergents, has surfactant properties, with an apparent surface saturation concentration of 1 μM, after which, the toxin is no longer in the monomer form (Fig 1). Experiments were conducted at a final concentration of 60 nM, to prevent artifacts during monolayer investigations. At this concentration, mycolactone interacts with the membrane as a monomer. The concentrations of mycolactone naturally present in the lesions during the course of the disease and leading to progressive ulceration are not always known. Determinations of the concentrations of mycolactone A/B in the various untreated pre-ulcerative nodules and plaques, ulcers and edematous lesions in M. ulcerans-infected human skin (biopsies) have revealed considerable variability [104]. The median concentration in all types of lesions or within the lesion itself (center or periphery) varied from 35 nM (i.e., 26 ng/mL) for pre-ulcerative lesions or the periphery of the lesions, to 596 nM (i.e., 443 ng/mL) in ulcers or 1.2 μM (i.e., 895 ng/mL) in edematous lesions. Mycolactone can also be detected in ulcer exudates obtained non-invasively from wound swabs, at concentrations of 67–270 nM (i.e., 50–200 ng/mL) [105]. Finally, it has been shown that the toxin concentration rapidly increases following inoculation with M. ulcerans in a mouse model of disease, from 385 ± 142 nM (i.e., 286 ± 105 ng/mL) on day 3, to 1.28 ± 0.29 μM (i.e., 948 ± 215 ng/mL) on day 7 and 4.85 ± 0.63 μM (i.e., 3603 ± 478 ng/mL) on day 62 [106]. Analyses of the ability of low biological concentrations (60 nM) of mycolactone to interact with the membrane provide information about the initial effects of the toxin on the plasma membrane, the first barrier that the toxin must cross to reach its intracellular targets. Presumably, these effects are progressively amplified during disease development, with the gradual increase in mycolactone concentration.
Mycolactone can bind to membranes regardless of their lipid composition (Fig 5). However, the presence of cholesterol promotes toxin insertion into the monolayer (Fig 6), highlighting the key role of this sterol in the interaction of the toxin with the membrane. Cholesterol modulates membrane fluidity and influences the organization of other lipids by changing their ordering, available area and the formation of domains of characteristic composition [107]. We also found that mycolactone had a strong effect on lateral lipid segregation in the membrane and the formation of distinct domains in monolayers containing cholesterol (Figs 5 and 7).
In the monolayer with a composition resembling that of the plasma membrane (mixture 1), preferential interactions between cholesterol (19%) and one high-melting lipid, SM (33%), which were mixed with the low-melting lipids POPC (39%) and POPE (9%) [65,90,108], led to Lo/Ld segregation (Figs 3A and 7) mimicking the lateral heterogeneity of cell membranes, with the coexistence of ordered and non-ordered lipid domains [66,79,81–83,87–89,92,109,110]. This lateral heterogeneity, with the coexistence of Ld and Lo phases, would compartmentalize cellular membranes and play a key role in the lateral segregation of various classes of membrane proteins to facilitate the various cellular functions and processes occurring at the membrane [81,92,108,111–113].
We show here that mycolactone acts as a reshaping agent at very low concentrations and that, in its monomer form, it disturbs lipid segregation in monolayers (Figs 3A, 7 and 8). A similar effect has been described for penetratin, a cell-penetrating peptide known to cross cell membranes. Penetratin recruits specific lipids locally for the formation of fluid membrane patches dispersed within ordered domains [114]. The reshaping induced by mycolactone may have a direct effect on the cellular functions modulated by lipid domains, by affecting the formation of these domains. This hypothesis is supported by the promotion of mycolactone/lipid interactions within the monolayer by cholesterol and by results concerning the membrane-fluidizing effect of the toxin preventing SM aggregation (in the absence of cholesterol, Fig 4A).
Using a fluorescent derivative with a biological activity one tenth that of mycolactone, Snyder & Small showed, in 2003, that mycolactone was localized in the cytosol of murine fibroblasts cultured in vitro [11,12]. Similar results have been obtained with human epithelial cells and lymphocytes exposed to a 14C-labeled form of the toxin; the mycolactone accumulates in the cytoplasm, and not in cell plasma or nuclear membranes [13]. Overall, these results are consistent with non-cell-specific passive diffusion of the toxin through the plasma membrane to reach its intracellular targets. We show here that mycolactone, in its monomer form (i.e., 60 nM final concentration), can modify membrane lipid organization by reversing lipid segregation. Thus, if it crosses the plasma membrane at concentrations below its apparent CMC, mycolactone may disturb the natural distribution of lipids in the membrane and the formation of lipid nanodomains, with consequences for metabolic pathways involving ordered membrane domains.
As mentioned in the introduction, most of cellular targets of mycolactone are membrane protein receptors residing in ordered plasma membrane nanodomains, where their functionalities can be modulated. Mycolactone can bind and modulate the activity of two members of a family of scaffold proteins, Wiskott-Aldrich syndrome protein (WASP) and neural WASP (N-WASP), which transduce various endogenous signals through a dynamic remodeling of the actin cytoskeleton [17]. However, sphingolipid-cholesterol domains have been shown to be the preferred platforms for membrane-linked actin polymerization mediated by in situ phosphatidylinositol 4,5-bisphosphate (PIP2) synthesis and tyrosine kinase signaling through the WASP-ARP2/3 pathway (PIP2 stimulates de novo actin polymerization by activating the pathway involving WASP and the actin-related protein complex ARP2/3 [115]). Mycolactone binding to WASP involves a lysine-rich basic region (BR) [17] that has also been implicated in the activation of WASP/N-WASP by PIP2 in both the allosteric and the oligomerization domains [116]. The Lo phase reversion triggered by the membrane insertion of mycolactone may therefore play an important role in modulating the WASP-Arp 2/3 pathway. Could mycolactone replace PIP2 in the disorganized sphingolipid-cholesterol microdomains and bind to WASPs?
Mycolactone also triggers a depletion of thrombomodulin (TM) receptors on the surface of endothelial cells, leading to disruption of the protein C anticoagulation pathway. This depletion has been observed in vitro in human dermal microvascular endothelial cells (HDMVECs) exposed to a very low dose (2 ng/mL i.e., 2.69 nM) of mycolactone and, in vivo, in the subcutaneous tissues of punch biopsies, and is strongly associated with the fibrin deposition commonly observed in BU skin lesions [34]. Nevertheless, TM receptors for protein C activation and activated protein C (APC) signaling pathways are colocalized in the lipid microdomains of endothelial cells [35,36]. This localization in the same domain is the key requirement for APC signaling pathways in endothelial cells [35,36]. Again, the change in lipid segregation, in addition to the observed depletion of TM receptors, due to the interaction of mycolactone with the plasma membrane during its passage into the cell, may disturb lipid microdomain formation in the membrane, thereby modifying the various signaling pathways requiring lipid membrane domains as signaling platforms [111]. Could the TM depletion observed in endothelial cells exposed to mycolactone [34] also be due to changes in lipid segregation in the plasma membrane, to which this receptor is targeted for functional activity?
Finally, many of the pathogenic effects of mycolactone could potentially be explained by a blockade of protein translocation [14]. Nevertheless, no unifying molecular mechanism underlying the pleiotropic actions of mycolactone has yet been identified. It remains, unclear, for example, how mycolactone blocks ER translocation at the molecular level? Baron et al. recently showed that mycolactone targets the α subunit of the Sec61 translocon, thereby strongly blocking the production of secreted and integral membrane proteins [38]. However, it is now widely accepted that the signal sequence of secreted proteins is clipped off during translation by a signal peptide peptidase (SPP). After chaperone-assisted folding, the mature protein is then released into the lumen of the ER immediately after its synthesis [41]. However, it has been suggested that SPP cleavage may be regulated through the control of substrate entry into the microdomains of the membrane containing SPP [117,118]. During the fractionation of rough ER integral membrane proteins with 0.18% Triton X-100, most ER-resident proteins, and the Sec61alpha subunit in particular, were found to be present in the microdomain-like fraction [44]. Little is known about the role of cholesterol in basic ER functions, but recent studies have clearly suggested that cholesterol may act with ER membrane proteins to regulate several important functions of the ER, including the folding, degradation, compartmentalization, and segregation of ER proteins, and sphingolipid biosynthesis [119]. McKenna et al. have provided biochemical evidence that mycolactone induces a conformational change in the transmembrane pore-forming Sec61alpha subunit of the translocon [39]. Hence, taking into account the potentially important contribution of cholesterol to the effect of mycolactone on membrane lipid segregation, we cannot exclude the possibility that the ER membrane may also be reshaped by mycolactone, abolishing the functions of the ER membrane-resident proteins. In the case of the Sec61 translocon, this hypothesis is consistent with the induction of a stabilized closed conformation of the Sec61alpha unit (forming the central gated protein-conducting channel across the ER membrane) [39,45] mediated by lipid redistribution in microdomains to facilitate the interaction of mycolactone close to the luminal plug of Sec61alpha, as recently suggested by Baron et al. [38].
The redistribution of lipid microdomains from the plasma membrane to mitochondria has recently been demonstrated in other contexts and diseases and is robust to this hypothesis [120]. The alteration of ER lipid microdomains initiates lipotoxicity in pancreatic β-cells, disturbing protein trafficking and initiating ER stress, thereby contributing to type 2 diabetes [121]. In the same way, the disruption of lipid microdomains stimulates phospholipase D activity in human lymphocytes, and this activation conveys antiproliferative signals in lymphoid cells, by impairing the transduction of mitogenic signals [122]. The gastrointestinal symptoms observed in patients suffering from Niemann-Pick type C disease are the consequence of changes in the composition of membrane lipid domains, resulting in impaired trafficking and an apical sorting of the major intestinal disaccharidases to the plasma membrane with a decrease in their functional capacities [123,124]. Edelfosine, an alkylphospholipid analog (APL) belonging to a family of synthetic antitumor compounds, induces apoptosis in several hematopoietic cancer cells, by targeting various subcellular structures at the membranes [125,126]. By recruiting death receptor and downstream apoptotic signaling molecules to ordered lipid domains, it displaces survival signaling molecules from these membrane domains. Edelfosine-induced apoptosis in solid tumor cells is mediated by an ER stress response and evidence has been obtained in vitro and in vivo to suggest that edelfosine treatment induces a redistribution of lipid domains from the plasma membrane to mitochondria, suggesting a raft-mediated link between the plasma membrane and mitochondria [120]. All these examples suggest that membrane reshaping can occur in different diseases, and that the disturbance of lipid segregation observed here with mycolactone is not an isolated case.
In summary, all the cellular targets of mycolactone are membrane-bound proteins, and, with the exception of the Sec61 translocon, all are known to be regulated by ordered microdomains, which provide a platform for the assembly of signaling complexes and prevent cross-talk between pathways [25,110]. By disturbing lipid segregation in membranes containing cholesterol, mycolactone affects many cell functions and signaling pathways. This membrane remodeling may occur in synergy with the previously demonstrated effects of mycolactone on its intracellular targets, possibly even potentiating these effects. It is tempting to speculate that microdomain remodeling in membranes underlies the molecular events via which mycolactone affects multiple targets, but further studies are required to confirm this.
Ultrapure water was obtained from a PURELAB option Q7 system (VEOLIA WATER STI, France). Phosphate-buffered saline (PBS, 2.8 mM KCl, 140 mM NaCl and 10 mM phosphate, pH 7.40 ± 0.05 at 25°C) was prepared by dissolving tablets purchased from AppliChem GmbH (Darmstadt, Germany) in ultrapure water.
We obtained 1-palmitoyl-2-oleoyl-sn-glycerophosphocholine (POPC), 1-palmitoyl-2-oleoyl-sn-glycerophosphoethanolamine (POPE), cholesterol (Chol) from ovine wool (≥98%) and 23-(dipyrrometheneboron difluoride)-24-norcholesterol or TopFluor Cholesterol from Avanti Polar Lipids (Alabaster, Alabama, USA). Sphingomyelin (SM) from chicken egg yolk (≥95%) was purchased from Sigma-Aldrich (Saint-Quentin Fallavier, France). All chemicals were used as received. The solvents were of analytical grade (Sigma-Aldrich, Saint-Quentin Fallavier, France). The lipid mixtures were prepared at a concentration of 1 mg/mL in chloroform (or chloroform/methanol, 9:1 v/v, when containing SM) and stored at -20°C under argon to prevent lipid oxidation.
Polyethylene glycol sorbitan monolaurate (Tween 20) and polyethylene glycol tert-octylphenyl ether (Triton X-100) were also purchased from Sigma-Aldrich (Saint-Quentin Fallavier, France).
Mycolactones A/B were purified from M. ulcerans extracts as previously described [6,127]. Briefly, S4018, an African strain of Mycobacterium ulcerans obtained from a patient in Benin, was grown in Middlebrook 7H10 agar supplemented with oleic albumin dextrose catalase growth supplement. The bacteria were resuspended in chloroform-methanol (2:1, v/v) and cell debris was removed by centrifugation. Folch extraction was performed by adding 0.2 volumes of water. The organic phase was dried and phospholipids were precipitated with ice-cold acetone. The acetone-soluble lipids were loaded onto a thin layer chromatography plate and eluted with chloroform-methanol-water (90:10:1, v/v/v) as the mobile phase. The yellow band with a retention factor of 0.23 was scraped off the plate, filtered, evaporated, resuspended in absolute ethanol and then stored in amber glass tubes in the dark. Its concentration was determined by measuring absorbance (λmax = 362 nm, log ε = 4.29), and its purity (>98%) was evaluated with a Shimadzu Ultra-Fast Liquid Chromatograph (UFLC XR system with a CBM-20A controller, a CTO-10AS Prominence column oven, LC-20AB pumps, an SPD M20A diode array detector (Shimadzu, Japan)) and a reverse C18 column (Zorbax 23 Eclipse XDB-C18, 9.4×250mm, Particle Size: 5 μm (Agilent, USA)).
Monolayers were prepared on a KSV 2000 Langmuir-Blodgett trough (3 multi-compartments, KSV NIMA, Biolin Scientific, Finland), with a symmetric compression system. The rectangular trough had a volume of 80 mL and a surface area of 119.25 cm2. A Wilhelmy plate attached to an electronic microbalance was used to measure the surface pressure (π), with an accuracy of ± 0.5 mN/m. The trough was cleaned with successive baths of dichloromethane, ethanol and ultrapure water, and filled with a filtered PBS solution. The subphase buffer was maintained at the desired temperature (20°C or 25°C) throughout the experiment, with an Ecoline RE106 low-temperature thermostat (LAUDA, Germany). It was not possible to work at a higher temperature, due to subphase evaporation, which can falsify surface pressure measurement during the run. Lipid mixtures in chloroform were gently spread at the air/liquid interface of the PBS subphase. The solvent was allowed to evaporate off for 15 minutes, and the monolayer was then slowly compressed by two mobile barriers at a constant rate of 0.045 nm2.molecule-1.min-1 until an initial surface pressure (πi) of 5 to 30 mN/m was reached. A lag time of about 1 h was then applied to allow the monolayer to relax and stabilize. The surface area was then kept constant by stopping the movement of the mobile barriers.
Mycolactone (1 mg/mL in ethanol) was injected (4.45 μL) into the subphase just beneath the lipid monolayer at a final concentration of 60 nM. The changes in surface pressure induced by the interaction of mycolactone with the monolayer were recorded continually, as a function of time, with a computer-controlled Langmuir film balance KSV NIMA (Biolin Scientific, Finland), until the equilibrium surface pressure (πe) was reached, indicating the end of the adsorption process. All measurements were repeated at least three times for each set of conditions, with a satisfactory reproducibility, and the mean values are reported here.
We used a monolayer with a lipid concentration closely resembling that of the plasma membrane, according to several authors [72–75]. This monolayer contained 39% phosphatidylcholine (POPC), 33% sphingomyelin (SM), 9% phosphatidylethanolamine (POPE) and 19% cholesterol (Chol) (in mol%). For evaluation of the influence of cholesterol on both the membrane-binding capacity and effects of mycolactone on phospholipid membrane organization, we also analyzed monolayers with a different composition devoid of cholesterol but with the molar ratios of the other lipids maintained. This second monolayer contained 48% POPC, 41% SM, and 11% POPE.
The surface pressure increase (Δπ in mN/m) after the mycolactone injection corresponds to πe - πi. The curve of surface pressure increase (Δπ) as a function of time (t) recorded during the adsorption of mycolactone onto lipid monolayers corresponds to the adsorption kinetics of the molecule.
The parameters characterizing the binding of mycolactone to different lipid membranes were further determined, as previously described [50,56,93,94]. Briefly, Δπ was plotted against different initial surface pressures πi to determine: i) the critical surface pressures πc, also known as the maximum insertion pressure (MIP), which is calculated by extrapolation of the linear regression line to the x-axis (Δπmax = 0) and, ii) the synergy factor, a, measured by adding 1 to the slope obtained from the linear regression of Δπ as a function of πi. The uncertainty on MIP and the synergy factor, a, were determined as previously described [50,94]. The uncertainty on MIP was calculated with a 95% confidence interval from the covariance of the experimental data for the linear regression [50]. The uncertainty on synergy was calculated as previously described [94]. These experimental errors were directly determined with the free binding parameters calculator software (http://www.crchudequebec.ulaval.ca/BindingParametersCalculator) developed by Salesse’s group.
Brewster angle microscopy (BAM) was used to characterize the lipid domain morphology of monolayers at the air/water interface [62,64]. Monolayer morphology was determined before and after mycolactone injection, with an EP3SW Brewster angle microscope (Accurion, Germany) equipped with a 532 nm laser, a polarizer, an analyzer and a CCD camera. BAM image size was 483 × 383 μm2. For ultrathin films, reflectance depends on both the thickness and refractive index of the monolayer. The different views of the interfacial film were reconstituted with EP3viewer BAM software (Accurion, Germany), based on the brightness of the BAM pictures. For a constant refractive index, reflectance is directly linked to the thickness of the interfacial film.
Langmuir films were generated in a custom-built cylindrical Teflon trough with a quartz window, containing 25 mL of filtered buffer, connected to a peristaltic pump. The system was mounted on the stage of a Zeiss Observer Z1 microscope (Carl Zeiss Vision, Germany) for fluorescence microscopy (FM) experiments.
Samples were prepared for FM by replacing 0.5 mol% of the cholesterol with 0.5 mol% of the sterol fluorescent probe, TopFluor Cholesterol [96]. The images were acquired at excitation and emission wavelengths of 495 and 507 nm, respectively. Images were processed and analyzed with dedicated Zeiss software (Axio Vision 4.8).
We used a Hamilton syringe to spread a few microliters of a 1 mg/mL phospholipid solution in chloroform onto the buffer subphase until the desired πi was reached. One hour later, after the solvent had evaporated and the lipid monolayer had stabilized at the desired initial surface pressure (πi), the mycolactone A/B solution was injected, with a Hamilton syringe, into the subphase through the lipid monolayer, with gentle stirring. During the time course of the experiment, changes in surface pressure (π) were also recorded simultaneously and continuously with a KSV NIMA computer-controlled Langmuir film balance.
|
10.1371/journal.pgen.1005288 | In Vivo Senescence in the Sbds-Deficient Murine Pancreas: Cell-Type Specific Consequences of Translation Insufficiency | Genetic models of ribosome dysfunction show selective organ failure, highlighting a gap in our understanding of cell-type specific responses to translation insufficiency. Translation defects underlie a growing list of inherited and acquired cancer-predisposition syndromes referred to as ribosomopathies. We sought to identify molecular mechanisms underlying organ failure in a recessive ribosomopathy, with particular emphasis on the pancreas, an organ with a high and reiterative requirement for protein synthesis. Biallelic loss of function mutations in SBDS are associated with the ribosomopathy Shwachman-Diamond syndrome, which is typified by pancreatic dysfunction, bone marrow failure, skeletal abnormalities and neurological phenotypes. Targeted disruption of Sbds in the murine pancreas resulted in p53 stabilization early in the postnatal period, specifically in acinar cells. Decreased Myc expression was observed and atrophy of the adult SDS pancreas could be explained by the senescence of acinar cells, characterized by induction of Tgfβ, p15Ink4b and components of the senescence-associated secretory program. This is the first report of senescence, a tumour suppression mechanism, in association with SDS or in response to a ribosomopathy. Genetic ablation of p53 largely resolved digestive enzyme synthesis and acinar compartment hypoplasia, but resulted in decreased cell size, a hallmark of decreased translation capacity. Moreover, p53 ablation resulted in expression of acinar dedifferentiation markers and extensive apoptosis. Our findings indicate a protective role for p53 and senescence in response to Sbds ablation in the pancreas. In contrast to the pancreas, the Tgfβ molecular signature was not detected in fetal bone marrow, liver or brain of mouse models with constitutive Sbds ablation. Nevertheless, as observed with the adult pancreas phenotype, disease phenotypes of embryonic tissues, including marked neuronal cell death due to apoptosis, were determined to be p53-dependent. Our findings therefore point to cell/tissue-specific responses to p53-activation that include distinction between apoptosis and senescence pathways, in the context of translation disruption.
| Growth of all living things relies on protein synthesis. Failure of components of the complex protein synthesis machinery underlies a growing list of inherited and acquired multi—organ syndromes referred to as ribosomopathies. While ribosomes, the critical working components of the protein synthesis machinery, are required in all cell types to translate the genetic code, only certain organs manifest clinical symptoms in ribosomopathies, indicating specific cell-type features of protein synthesis control. Further, many of these diseases result in cancer despite an inherent deficit in growth. Here we report a range of consequences of protein synthesis insufficiency with loss of a broadly expressed ribosome factor, leading to growth impairment and cell cycle arrest at different stages. Apparent induction of p53-dependent cell death and arrest pathways included apoptosis in the fetal brain and senescence in the mature exocrine pancreas. The senescence, considered a tumour suppression mechanism, was accompanied by the expression of biomarkers associated with early stages of malignant transformation. These findings inform how cancer may initiate when growth is compromised and provide new insights into cell-type specific consequences of protein synthesis insufficiency.
| The protein translation machinery encompasses interrelated processes of ribosome biogenesis [1] as well as protein synthesis [2]. Mutations in genes that encode components of this machinery are implicated in a growing list of inherited and acquired disorders termed ribosomopathies. All aspects of cell growth require protein synthesis and deficiency in machinery biogenesis or function can be anticipated to have systemic effects with reduced growth caused by translation insufficiency. This was observed in the Drosophila minutes that were initially identified by diminutive size, and are now known to possess mutations in ribosome related genes [3]. Nevertheless, ribosomopathies present as clinical syndromes with select organ failure, often including the bone marrow [4,5]. The mechanisms dictating which organs are affected by any given ribosomopathy are unknown. Susceptibility to organ failure may reflect specific cell type expression levels or threshold requirements for translation [6]. Developmental requirements during organ expansion [6,7] and functional requirements during cued response to extrinsic signals may add other levels of complexity.
Most ribosomopathies are cancer predisposition syndromes. They can be associated with increased risk of hematological malignancies, and solid tumours have also been reported [4]. Numerous studies have linked defects in translational control and ribosome gene dosage to aberrant growth [8,9]. However, studies have primarily discussed cancer progression in the context of increased ribosome biogenesis and/or translation. What precipitates malignancies in a growth-disadvantaged context such as that of a ribosomopathy remains poorly understood.
A number of consequences have been noted with loss of the highly conserved ribosome-associated protein SBDS and its orthologs in various model systems with a common thread of deregulated protein synthesis. There are several lines of evidence indicating that SBDS functions in ribosome metabolism [10,11], specifically with eukaryotic initiation factor 6 (EIF6) and elongation factor Tu GTP binding domain containing 1 (EFTuD1) protein [12,13]. EIF6 is required for binding and maturation of the 60S ribosomal subunit [14,15] and has been shown to block ribosome subunit joining for formation of the 80S ribosome [16,17]; hence EIF6 is considered to limit translation initiation [18]. Gain of function mutations in the yeast ortholog of Eif6 rescued the severe slow-growth phenotype of SBDS-null yeast strains (sdo1Δ) and EFTuD1-null yeast strains (ria1Δ) [13,19]. The removal of EIF6 from the 60S ribosomal subunit was shown to require the GTPase activity of EFTuD1 [19]. Further, genetic and protein interactions between homologs of EFTuD1 and SBDS have been demonstrated [11,12,20,21]. The current working molecular model is that SBDS acts with EFTuD1 to promote EIF6 removal from the 60S ribosomal subunit [12,22,23].
Shwachman-Diamond syndrome (SDS) is a recessive ribosomopathy caused by biallelic loss-of-function mutations in SBDS [24]. SDS is a multisystem disorder presenting typically within the first year of life with failure to thrive, chronic infection and low blood counts [25]. Exocrine pancreatic dysfunction and blood lineage cytopenia (most often neutropenia) are defining features [26]. Other clinical findings include skeletal defects, decreased brain volume and cognitive impairment [27–30]. SDS is associated with high risk of hematological malignancies (up to 30%) [31]; more recently, early onset solid tumours have also been observed, notably including pancreatic carcinoma [32–34]. The exocrine pancreas has amongst the highest requirements for translation in the body as the site of reiterative digestive enzyme production [35]; SDS pancreatic dysfunction is characterized by severe digestive enzyme deficiency [36].
Studies in both patient-derived cell lines [21,37–39] and animal models of SDS [12,13,21,23,40,41] have demonstrated a role for SBDS in ribosome maturation and ribosome subunit joining. Furthermore, decreased global translation was demonstrated in mouse embryonic fibroblasts with disease-associated mutations of Sbds [41] and in human embryonic kidney 293 cell lines depleted for SBDS by siRNA [42]. It remains to be determined how SDS-related disruptions in translation manifest as acute dysfunctions in select organs.
Senescence is a permanent cell cycle arrest associated in vivo with tumour suppression and aging. In the context of tumours, senescence is considered to act as a rapid response to aberrant growth, particularly downstream of oncogene induction (e.g. RAS activation) [43]. Engagement of tumour suppressors including p53, CDKN2A (p16INK4A), pRB [43], TGFβ and CDKN2B (p15INK4B) [44,45], can initiate this permanent arrest of the cell cycle that is associated with quiescent cells that secrete inflammatory cytokines (senescence-associated secretory phenotype) and express senescence-associated β–galactosidase activity (SAβG) [43].
Here we sought to identify in vivo mechanisms underlying pancreas dysfunction, in comparison to other organs, in SDS. We used constitutive and targeted mouse models to establish the timing and type of organ responses to Sbds mutation. Specifically, we show the dependence of many responses on p53 and that SDS-related translation insufficiency induces a senescent cell cycle arrest through the induction of Tgfβ and p15Ink4b in the murine exocrine pancreas. Our study provides new insights into organ selectivity and tumorigenic potential in ribosomopathies.
Mouse models with disease-associated missense (R126T) and null (–) alleles, SbdsR126T/R126T and SbdsR126T/–, displayed severe growth impairment and did not survive birth (S1 Table; [46]). Models demonstrated complete penetrance and consistent genotype-phenotype correlations, with more severe and earlier onset of disease phenotypes in the SbdsR126T/–embryos compared to SbdsR126T/R126T embryos (Fig 1). Heterozygous carriers of either the Sbds−or SbdsR126T alleles were indistinguishable from wildtype, consistent with a recessive mode of inheritance for SDS. Embryos were visibly smaller by two weeks gestation and at E18.5 were, on average, 38% (SbdsR126T/–) and 56% (SbdsR126T/R126T) of age-matched controls by mass (Figs 1A and S1A). Embryo length was also reduced (S1C Fig).
SDS mouse models recapitulated several features observed in human disease. Mutations in Sbds are associated with defects in hematopoiesis [47]. In the fetal period, the liver is the primary site of definitive hematopoiesis. In the SbdsR126T/R126T model near birth (E18.5), histopathology indicated decreased granulocytes in portal areas of the liver as well as pronounced bone marrow hypocellularity, with increased severity in the SbdsR126T/–model (Fig 1B).
SDS is also characterized by decreased ossification and delayed bone growth [27,48]. No gross skeletal defects were apparent in the constitutive models [41]; however, ossification was reduced in the metacarpals at late gestation (Fig 1C). In severe cases, asphyxiating thoracic dystrophy has been observed in SDS [49,50], presumably due in part to the skeletal dystrophy. Beyond this, lung pathology has not been specifically reported in SDS patients. We did observe a severe decrease in saccule expansion in the late fetal lung, despite presence of lung developmental stage biomarkers (S2A Fig).
A defining morphological feature of SDS is a small, fat-replaced pancreas [25,27,30]; we previously showed that pancreatic growth impairment, dysfunction and lipomatosis manifest only in the postnatal period [51].
Translation insufficiency impacts all tissues, and all ribosomopathies are associated with poor overall growth. To further investigate the observed decreases in granulocytes in the liver and hypoplasia of the bone marrow compartment we assessed the abundance of hematopoietic progenitors in the fetal liver of the SDS mouse models. Primary myocult cultures derived from E14.5 fetal livers revealed markedly decreased levels of all myeloid lineage progenitors in both the SbdsR126T/R126T and SbdsR126T/–models (S3 Fig). Decreased levels of granulocytes in the SbdsR126T/–model were also determined by flow cytometry of fetal liver cells (E16.5, S4A Fig). Unlike ribosome deficiency models with dominant inheritance [52–54], erythrocyte levels prior to birth (E18.5) were normal in both mouse models (S4B Fig), consistent with observations in SDS patients [31].
In contrast to other organs, Sbds mutations resulted in severe proliferation defects with pyknotic nuclei and apoptosis (detected by TUNEL staining) in the developing brain by E11.5 in both SbdsR126T/–and SbdsR126T/R126T models (Fig 2A). At E14.5, TUNEL staining was very prominent in the intermediate zone and bromodeoxyuridine labeling further indicated poor growth of neuronal progenitors in the ventricular zone of the developing cortices (Fig 2B). By E18.5, the brain showed multifocal lesions of necrotic neurons (S5 Fig). We did not observe an increase in TUNEL staining in other tissues at E18.5, including the liver and bone marrow, beyond what was observed in controls (S6 Fig).
As mentioned above, the constitutive SDS models did not survive birth. Using a conditional knockout allele (CKO) in conjunction with a pancreas-specific Cre driver (Ptf1aCre) to circumvent lethality, we previously showed that biallelic loss-of-function mutations in Sbds result in a very small pancreas (53% of controls, relative to body mass [51]) with severe atrophy of the acinar component of the adult pancreas. This phenotype included a dramatic depletion of zymogen granules, the specialized vesicles that house digestive enzymes in pancreatic acinar cells. Furthermore, in contrast to the developing brain, poor pancreatic growth was not explained by apoptosis [51].
Given the loss of zymogen granules and acinar cell hypoplasia with a persistent absence of cell death markers, we considered that a senescent cell-cycle arrest might explain atrophy of the SDS pancreas. Several acinar cells of the SDS pancreas were positive for SAβG activity by 20 days of age, becoming more prominent by 30 days of age (Fig 3A). With this evidence of senescence, we next investigated the nature of this response by performing transcript analyses with reverse-transcriptase real-time quantitative PCR of a curated cellular senescence panel of genes. Pancreas samples from littermate control-mutant pairs were compared at two time points prior to the pronounced SAβG activity (Fig 3B, S2 Table). SDS pancreas transcripts showed a suite of changes that were, consistent with the literature, indicative of a senescent-associated cell cycle arrest and secretory program [55–57]. We then further investigated targets of the p53/p21Cip1 and Tgfβ/p15Ink4b networks with additional samples and time points (Fig 3C). We detected markedly increased expression of p15Ink4b (Cdkn2b) along with Tgfβ together with low Myc expression at 15 and 25 days of age (Fig 3C). Increased expression of p21Cip (Cdkn1a) occurred at the early time point of 15 days. Consistent with low Myc levels being permissive for p15Ink4b induction by Tgfβ [58], decreases in Myc transcript levels were noted already at one-week of age, preceding increases in Tgfβ and p15Ink4b (Fig 3C). An increase in p53 transcript expression (3.70 fold, relative to controls), a known mediator of the senescence response [43], also coincided with the onset of SAβG activity (Fig 3C).
Protein expression analyses of control-mutant littermate pairs from several litters paralleled the transcript changes in the SbdsP–/R126T pancreas with changes in Myc and Tgfβ signalling (Fig 4A). Steady-state protein levels of Myc and Tgfβ were consistently reduced and higher in mutants, respectively (Fig 4A). Tgfβ signalling is propagated by phosphorylation of the Smad proteins by Tgfβ receptors [59]. We noted less Smad3 phosphorylation, but more Smad2 phosphorylation in mutants than in controls (Fig 4A). We also observed increased transcript levels for Tgfβ receptors, TgfbrII and TgfbrIII, which can be upregulated during increased Tgfβ signalling [60] (Fig 3C).
Expression of several factors implicated in the senescence-associated secretory phenotype [55,56,61] beyond Tgfβ, were also elevated. These included extracellular matrix proteins fibronectin (Fn1), osteonectin (Sparc) and collagen (Colla1) as well as innate immunity genes (e.g. Irf5, Irf7 and Nfkb1) and insulin growth factor binding proteins (Igfbp5 and Igfbp7) (Fig 3B; S2 Table), consistent with a senescence program.
Notably, indicators of replicative- and oxidative stress-induced senescence (e.g. Sod1 and Akt1, respectively [43]) were not elevated (S2 Table). Further, that expression of proto-oncogenes Akt1, Hras and Kras as well as Myc trended downwards or were reduced refuted an oncogene-induced senescence response (Fig 3B; S2 Table).
Tgfβ is a known driver of epithelial to mesenchymal transition [59] so we also considered that this process may be occurring in the SDS pancreas. We did observe indicators of dedifferentiation in the mature SDS pancreas (see below); however E-cadherin (Cdh1) transcript levels were not significantly reduced at young ages (Fig 3C).
To determine if the molecular signature of the pancreas senescence represented a common response to Sbds-ablation, we investigated whether cyclin inhibitors, Tgfβ, and Myc transcript level changes were evident in tissues of the constitutive SDS model (SbdsR126T/R126T). A marked increase in p21Cip transcript levels in the brain at E14.5 was observed when apoptosis was detected (Fig 3D). At this same early time point, Tgfβ expression was not altered in either mutant fetal brain or liver even though both organs manifested phenotypes (Fig 3D). Tgfβ expression was low in bone (Fig 3D) and unchanged in lung at E18.5 (S2B Fig). By E18.5, Tgfβ expression was elevated in the SDS mouse brain (Fig 3D), likely a late response to brain damage [62]. No changes in p15Ink4b or Myc expression levels were observed in fetal liver, lung, bone or cartilage tissues (Fig 3D; S2B Fig). These findings are consistent with Tgfβ/p15Ink4b-mediated senescence being a specific response of the pancreas to Sbds deficiency.
p53 is a known driver of senescent cell cycle arrest [57,63] and increased levels of p53 have been reported in SDS patients [64]. Moreover, studies of ribosomal gene haploinsufficiency have implicated p53 as a key factor in response to ribosome dysfunction [54,65]. We observed increased steady-state levels of p53 protein in the SDS mouse pancreas by immunoblotting (3 weeks of age, Fig 4B). Immunohistochemistry for p53 (15 days of age, prior to the detection of SAβG staining, Fig 4C) specifically highlighted nuclei of acinar cells, but not islet cells (Fig 4C). To determine if the senescence in the SDS pancreas is p53-dependent we bred the SDS pancreas model to a Trp53–/–mouse.
Complete genetic ablation of p53 alleviated the phenotypes of the SDS pancreas. SbdsP–/R126T;Trp53–/–animals demonstrated a notable improvement in pancreas mass as compared with SbdsP–/R126T;Trp53+/–animals (Fig 5A). Growth improvement was also evident at the histological level as acinar hypoplasia and fat infiltration did not occur in Sbds/Trp53 double mutants in direct contrast to single Sbds mutants (Fig 5B). The molecular signature associated with senescence in the SDS model pancreas was no longer detected; specifically Tgfβ, and p15Ink4b transcripts were not elevated and Myc transcript levels were not decreased (Fig 5C) at 25 days of age. Further, elevated SAβG activity was not detected at 32 days of age (S7 Fig).
By one month of age, the architecture of the acinar epithelium in Sbds/Trp53 double deficient pancreata appeared disordered with many apoptotic cells evident by two months of age (Fig 5B). The morphology was consistent with early stages of acinar-ductal metaplasia. By 60 days of age, we had already noted that some acinar cells in the SDS model pancreas were positive for transcription factors Hes1 and Pdx1, both of which are associated with dedifferentiation (Fig 5B) [66]. With complete ablation of p53, staining of acinar cells with these dedifferentiation markers became widespread (Fig 5B). In contrast, we did not detect changes in islet structure, nor did islets contain apoptotic cells, consistent with our previous observation that mutations in Sbds specifically impact the acinar compartment of the pancreas [51].
The absence of p53 further revealed translation-insufficiency as a consequence of Sbds loss-of-function. A long established feature of ribosomal deficiency includes small cell size [7], a phenotype noted for the acinar cells of the Sbds/Trp53-double deficient pancreata. Quantification of micrographs of doubly deficient pancreas tissue revealed a nuclei count increase per acinar area compared to Trp53–/–controls (with Sbds), indicating more cells per area (and hence a decreased cell size; Fig 6A). Correspondingly, a smaller mean acinus diameter was also evident in the double mutant micrographs (Fig 6B).
Despite this indication of ribosomal deficiency, Sbds/Trp53 double mutants demonstrated a substantial rescue of digestive enzyme expression and zymogen granule abundance. In fact, SDS pancreas lysates showed qualitatively different protein expression patterns that became similar to controls when p53 was absent (Fig 6C). Although amylase expression remained low, increases in protease (carboxypeptidase) expression at three weeks (Fig 6D) as well as restoration of zymogen granules by one week (Fig 6E) in the absence of p53 supported improvement in exocrine function.
We previously suggested that loss of Sbds results in a moderate decrease in 80S monosome peak levels compared to littermate controls [51]. Corresponding increases in free ribosomal subunit levels were not apparent as would be expected if ribosome production was maintained, whereas quantification of the 80S monosome peak levels in the polysome profiles of mutant pancreata showed that the modest decrease normalized to that of controls with ablation of p53 (Figs 6F and S8).
In contrast to senescence and its molecular signature that included Tgfβ, p15Ink4b and Myc, p53-dependence was not specific to the pancreas as loss of p53 impacted many phenotypes of the SDS mouse model. Although the lethality and growth impairment with reduced mass in the constitutive SDS mouse embryo was not improved (S1 Table; S1 Fig), loss of p53 had a restorative effect on blood progenitor levels (Fig 7A; S3 Table) and led to reduced apoptosis in the early SDS mouse brain to non-detectable levels (Fig 7B). As in the pancreas, polysome profiles are perturbed in SbdsR126T/R126T fetal livers, however loss of p53 resulted in only modest effects, with 80S monosome levels remaining far short of control levels (Figs 7C and S8).
Mutation of factors implicated in ribosome metabolism and translation lead to dramatic consequences for growth [3,8,9,67,68]. Our findings support the classification of SDS as a ribosomopathy [12,69]. Constitutive and targeted mouse models with SDS-associated Sbds alleles demonstrated severe growth impairment at both the organismal and organ levels. Mitogens were decreased in the SDS pancreas including low levels of the proto-oncogene Myc, a key regulator of exocrine pancreas expansion and acinar cell maintenance [70]. Moreover, arrest at the cell cycle level was evident with increased expression of cyclin inhibitors, senescence in the pancreas, and decreased BrdU-incorporation in the developing brain. With respect to direct evidence of a perturbation in ribosome metabolism, polysome analyses indicated a decrease in the proportion of 80S monosome levels in Sbds-ablated mutants relative to age-matched controls, with a notable difference in magnitude between the pancreas and liver. Although a subunit joining problem has been proposed previously in the context of Sbds mutations [12,21,37], our results are more consistent with an overall decrease in ribosome biogenesis, at least in the SDS pancreas. Finally, as observed in other ribosomopathies [54,65,71,72], we observed stabilization of p53 protein and increased Trp53 transcript levels in the SDS pancreas.
Constitutive ablation of Sbds in the mouse resulted in deficits in the hematopoietic and skeletal compartments, consistent with disease [25]. We previously demonstrated that targeted ablation of Sbds in the pancreas recapitulated all known SDS phenotypes of that organ [51]. Here we further identified a severe brain phenotype in constitutive models, with decreased proliferation in undifferentiated cells as well as pervasive cell death in differentiating neurons. These neural cell losses likely contribute to the perinatal lethality of the constitutive models. SDS is associated with cognitive impairment; imaging indicates reduced brain volume in patients [29,73] and approximately 20% of children with SDS meet criteria for intellectual disability [28] amongst other neurodevelopmental/behavioural concerns [27,28,74,75]. Neural cell death in the mouse occurred by p53-dependent apoptosis, consistent with neurological phenotypes observed in other ribosomopathy models [76–78].
Ribosome biogenesis and translational control have a significant impact on cell-cycle progression [79–81], therefore it is not surprising that cells with aberrant ribosome biogenesis and/or translation exhibit cell cycle arrest. However, while SDS-associated genotypes resulted in apoptosis in the fetal brain, senescence was observed in the postnatal pancreas. Given reports that Sbds loss may result in irregularities of the mitotic spindle [82,83], we considered that loss of Sbds could involve a DNA damage-induced senescence response. However, transcripts for DNA damage response factors Atm, Chek1 and Chek2 (S2 Table) were not upregulated. The senescence cell cycle arrest involved the p53/p21Cip1 and Tgfβ/p15Ink4b networks. Tgfβ is a key member of the senescence-associated secretory phenotype, with roles in both the establishment and maintenance of senescence [57]. Select activation of the Smad effector proteins to relay the Tgfβ signalling cascade is context specific [84]; our findings indicated Smad2 is involved in mediating Tgfβ senescence in the SDS pancreas.
Senescence is considered a hallmark of premalignant tumours [85]. A Tgfβ/p15Ink4b-mediated senescent response has been observed in the context of tumour suppression in hepatocellular carcinoma human cell lines [45] and in lymphomas [44]. Senescence can, over time, promote malignant transformation in neighbouring cells due to the chronic secretion of inflammatory cytokines that are part of the senescence-associated secretory phenotype [86]. Although we have not observed tumour formation up to 14 months in our mice, acinar cells did express markers of dedifferentiation and in light of recent reports of early onset, aggressive pancreatic cancers in SDS patients [32,34], we argue that senescent cells present in the SDS pancreas could contribute to malignant transformation.
That senescence and the underlying pathway involving Tgfβ were not invoked in other organs of the constitutive SDS model highlighted the diversity of outcomes with loss of Sbds. At the same time, both pancreatic senescence and neural apoptosis were abrogated with genetic depletion of p53, implicating p53 as a key mediator of the response to Sbds loss in these two tissues. In the pancreas, we detected stabilization of p53 expression, detectable by 15 days of age, specifically in nuclei of acinar cells. Moreover, we discovered that the characteristic pancreatic phenotypes in SDS are extensively p53-dependent, including organ morphology and the shutdown of the zymogen granule proteome which was evident at the transcription level.
A recent zebrafish model of SDS, generated via morpholino-mediated knockdown of homolog sbds (sbds-MO), demonstrated deficits in pancreatic progenitor proliferation that were phenocopied by ablation of ribosomal constituent proteins, highlighting hypersensitivity of the pancreas compartment to mutations in ribosome-associated genes [40,87].
The absence of p53 did not constitute the rescue of overall growth or perinatal survival of the SDS mouse model highlighting p53-dependent and p53-independent aspects of SDS pathology (as were reported for the zebrafish model [40]). Specifically, the apparent improvements in the Sbds/Trp53 pancreas double mutant phenotypes, were accompanied by a decrease in acinar cell size (a hallmark of translation insufficiency) supporting that protein synthesis remained compromised. Moreover, the profound 80S monosome loss in fetal liver cells (Fig 7) was not recovered with ablation of p53. We conclude that p53 is responding to Sbds deficiency by initiating cell cycle arrest (apoptosis or senescence) with some benefits. However, how the SDS-translation insufficiency triggers p53 activation, or how p53 activation achieves the apparent changes in phenotypes is not clear, perhaps through disturbed production or threshold shift of some critical checkpoint factor(s). With regard to the synthesis of specific proteins, we did note absence of recovery of amylase protein synthesis despite some resurgence of amylase transcript levels in the double mutant (Figs 5C and 6D).
In the pancreas, loss of Sbds is accompanied by cell cycle arrest and reduction of the zymogen granule transcriptome, leading to organ failure. Genetic ablation of Trp53 attenuated the response with rescued growth and increased staining for dedifferentiation markers. It remains to be determined what alternative network(s) may signal the apoptosis that was subsequently observed in absence of p53. Overall, our findings indicate a cellular imperative to shut down cells with disrupted ribosome metabolism, consistent with reports of protective cell shutdown in other ribosomopathy models [6,71].
Can the study of the Sbds-deficient models inform a key question of what dictates organ hypersensitivity to ribosome dysfunction? Robust and ubiquitous expression argues against Sbds expression levels being the limiting factor directly underlying the varied organ responses in SDS [24,46]. The responses of organs to the SDS-translation deficiency varied in both timing and molecular signature. However, despite these differences, many aspects of the brain, blood and pancreas pathologies are all downstream of p53. Our study suggests that the perceived organ paucity in ribosomopathies stems in part from a disparity in molecular responses to translation dysfunction, likely downstream of p53 activation. Such responses, with dependence on cell type, can thus result in vastly different tissue outcomes.
All animal experiments were carried out under the guidelines of the Canadian Council on Animal Care, with approval of procedures by The Animal Care Committee of the Toronto Centre for Phenogenomics, Toronto, AUP #0093. The generation of constitutive SDS and SDS pancreas mouse models was described elsewhere [46,51]. Heterozygous carriers of either the missense (R126T) or null (–) mutation were indistinguishable from wildtype littermates. All mouse lines were maintained on a C57BL/J6 background and no gender effects were observed. Excision of the floxed CKO allele was achieved by breeding with the Ptf1aCre mouse [88]. The p53 deficient strain B6.129S2-Trp53tm1Tyj/J (The Jackson Laboratory) was bred onto Sbds mutant lines for loss of p53 function studies. For embryonic staging, the morning of a vaginal plug was counted as embryonic day (E) 0.5. Mice were euthanized by decapitation, cervical dislocation or CO2 inhalation. Genotyping of DNA from tail samples was performed with the REDExtract-N-Amp Tissue PCR Kit (Sigma) using primers as previously described [51].
Flash frozen tissues were lysed in Polysome Buffer (100 mM KCl, 5 mM MgCl2, 10 mM Tris–HCl pH9.0, 1% Triton X–100 and 1% sodium deoxycholate in diethylpyrocarbonate—treated water) on ice using a polytron. Insoluble cell debris was pelleted by centrifugation at 2,500 X g for 15 min at 4°C. Cyclohexamide (0.1 mg/mL) and heparin (1 mg/mL) were added to the supernatant, and equal amounts of RNA (determined by A260 using a Nanodrop Spectrophotometer) were loaded onto a 10–50% sucrose gradient (100 mM KCl, 5 mM MgCl2, 10 mM Tris-HCl pH9.0). Sucrose gradients were subjected to ultracentrifugation (151,000 X g for 2 hours at 4°C) prior to fractionation using a density gradient fractionation system (Brandel). UV absorbance (A254) was recorded using PeakTrak software (Teledyne Isco). Area under the curve (AUC) was calculated using Adobe Photoshop CS5.1 as described [89]. Individual peak/compartment areas were expressed relative to the total AUC of the profile.
For paraffin embedding, organs were dissected and fixed overnight in ice-cold 4% paraformaldehyde prior to processing into paraffin blocks. Sections with thickness of 5 μm were used. Safranin O (counterstained with Fast Green) staining was performed by the pathology core at the Toronto Centre for Phenogenomics. For immunohistochemistry, antigen retrieval was achieved by boiling in citrate buffer (10 mM sodium citrate, pH6.0), endogenous peroxidases were blocked with 6% H2O2, and non-specific epitopes were blocked with 5–10% goat serum. Antibodies used are given in S4 Table; antibody binding was visualized using diaminobenzidine reagent (Sigma). For senescence-associated β–galactosidase activity staining assays, fresh tissue was embedded and frozen in Tissue-Tek O.C.T. Compound (Sakura Finetek) as per supplier instructions. Frozen tissues were sectioned as 8 μm slices. Senescence staining was performed at pH5.5 as previously described [90]. Apoptosis was detected on paraffin sections by TUNEL assay either using the In Situ Cell Death Detection Kit (Roche) as per supplier’s instructions (fluorescein visualization) or by the pathology core at the Toronto Centre for Phenogenomics (diaminobenzidine visualization). 5–bromodeoxyuridine (50 μg/g, BD Biosciences) was injected in staged pregnant females 24 hours prior to embryo dissection at E14.5. 5–bromodeoxyuridine incorporation was detected using the BrdU In Situ Detection Kit (BD Biosciences). For cell size measurements, nuclei and acini from at least 3 non-overlapping micrographs taken at 40X magnification from 4 biological replicates were counted and measured.
One week old pancreata were dissected and fixed in 2% glutaraldehyde in 0.1 M sodium cacodylate buffer (pH7.3). Fixed samples were processed and sectioned for electron microscopy by the joint Advanced Bioimaging Centre of The Hospital for Sick Children and Mount Sinai Hospital in Toronto.
Single cell suspensions from E14.5 embryo livers were prepared by grinding and filtering tissue through a 40 μm cell strainer (BD Biosciences). Cell suspensions were stained with conjugated antibodies against cell surface antigens with a FACSCalibur system (BD Biosciences) as previously described [91]. Antibodies used were Gr-1, c-kit (FITC-conjugated), Mac-1, and Ter119 (BD Biosciences). Flow cytometry data were analyzed using FlowJo software (Tree Star, Inc.).
Single cell suspensions from E16.5 embryo livers were prepared by grinding and filtering tissue through a 40 μm cell strainer (BD Biosciences). The number of cells per liver was determined by manual counting using a hemocytometer. Suspended cells (1X105 in 0.3 ml Dulbecco’s Modified Eagle Media) were mixed with 3 ml of methylcellulose media (Stem Cell Technologies) containing recombinant murine stem cell factor, recombinant murine IL-3, recombinant human IL-6 and recombinant human erythropoietin (Stem Cell Technologies), split into thirds and plated on three 35 mm tissue culture plates. Cells were incubated for 7 days at 37°C, 5% CO2 and ≥95% humidity. Colonies of each cell type were identified and counted using a light microscope according to supplier’s instructions. Counts for all three plates of each cell type were averaged and presented as counts per 80,000 cells plated. At least five embryos of each genotype were investigated.
Total RNA was isolated from RNAlater (QIAGEN) stabilized pancreas tissue (N = 3–4 for each genotype at each time point) or flash frozen tissues (brain, lung, liver, cartilage, bone; N = 4 for each genotype at each time point) using the RNeasy Mini Plus Kit (QIAGEN) according to manufacturer’s instructions with the addition of 5% β–mercaptoethanol in the homogenizing Buffer RLT Plus. For bone, cartilage and lung, homogenized tissues were first treated with Trizol (Life Technologies) before application to the RNeasy spin columns. Quality control and real-time quantitative PCR was performed as previously described [51]. Results are presented relative to the expression of the optimal control gene (four genes tested for each sample) for that tissue and time point as determined by GeNORM analysis [92]. A significant change was defined as a ≥2 fold difference with a P–value <0.05. Oligonucleotide primers are given in S5 Table. Expression levels of 84 cellular–senescence associated genes were assayed using the SABiosciences Cellular Senescence RT2 Profiler PCR Array (QIAGEN) with total RNA isolated from pancreata of mice at 15 and 25 days of age. A significant change was defined, as per supplier’s instructions, as a ≥3 fold difference with a P-value of <0.05. Selected gene results were confirmed by real-time quantitative PCR of independently prepared cDNA samples with distinct primer sets (with the exception of Cdkn2b where QIAGEN array primers were used).
Pancreas tissue (~30 mg) from 20 day old mice (prior to fat infiltration) was homogenized in RIPA buffer (150 mM NaCl, 1% NP-40, 0.5% sodium deoxycholate, 0.1% sodium dodecyl sulfate, 50 mM Tris-HCl, pH7.5) using a polytron over ice. Insoluble components were pelleted by centrifugation (17,000 X g at 4°C). Equal amounts of protein (determined by Lowry assay, BioRad) in 2X Laemmli buffer were separated by 12% SDS–PAGE and either stained with silver salts or Coomassie brilliant blue, or blotted using the Trans–Blot Turbo Transfer Pack with the Trans-Blot Turbo Transfer System (BioRad). Trans-Blot Turbo nitrocellulose membranes (BioRad) were blocked in 5% (w/v) powdered skim milk (5% (w/v) goat serum for Novacastra CM5 p53 antibody) prior to overnight incubation with primary antibodies followed by species appropriate horseradish peroxidase-conjugated secondary antibodies (S4 Table). Bound antibodies were visualized with Amersham ECL Prime Western Blotting Detection Reagent (GE Healthcare Life Sciences) on the ChemiDoc MP Imaging System using Image-Lab 4.1 Software (BioRad).
All statistical tests were carried out using R statistical software (R Foundation, from http://www.r-project.org). For T-tests, Welch’s correction was used to adjust for non-constant variance. Wilcoxon Rank Sum Test and Kruskal-Wallis analysis of variance were used where data did not show normal distribution. Bonferroni adjusted critical values were used to declare significance, adjusting for the number of comparisons per analysis. Raw P-values are reported.
|
10.1371/journal.pcbi.1000291 | Accurate Path Integration in Continuous Attractor Network Models of
Grid Cells | Grid cells in the rat entorhinal cortex display strikingly regular firing
responses to the animal's position in 2-D space and have been
hypothesized to form the neural substrate for dead-reckoning. However, errors
accumulate rapidly when velocity inputs are integrated in existing models of
grid cell activity. To produce grid-cell-like responses, these models would
require frequent resets triggered by external sensory cues. Such inadequacies,
shared by various models, cast doubt on the dead-reckoning potential of the grid
cell system. Here we focus on the question of accurate path integration,
specifically in continuous attractor models of grid cell activity. We show, in
contrast to previous models, that continuous attractor models can generate
regular triangular grid responses, based on inputs that encode only the
rat's velocity and heading direction. We consider the role of the
network boundary in the integration performance of the network and show that
both periodic and aperiodic networks are capable of accurate path integration,
despite important differences in their attractor manifolds. We quantify the rate
at which errors in the velocity integration accumulate as a function of network
size and intrinsic noise within the network. With a plausible range of
parameters and the inclusion of spike variability, our model networks can
accurately integrate velocity inputs over a maximum of ∼10–100
meters and ∼1–10 minutes. These findings form a
proof-of-concept that continuous attractor dynamics may underlie velocity
integration in the dorsolateral medial entorhinal cortex. The simulations also
generate pertinent upper bounds on the accuracy of integration that may be
achieved by continuous attractor dynamics in the grid cell network. We suggest
experiments to test the continuous attractor model and differentiate it from
models in which single cells establish their responses independently of each
other.
| Even in the absence of external sensory cues, foraging rodents maintain an
estimate of their position, allowing them to return home in a roughly straight
line. This computation is known as dead reckoning or path integration. A
discovery made three years ago in rats focused attention on the dorsolateral
medial entorhinal cortex (dMEC) as a location in the rat's brain where
this computation might be performed. In this area, so-called grid cells fire
whenever the rat is on any vertex of a triangular grid that tiles the plane.
Here we propose a model that could generate grid-cell-like responses in a neural
network. The inputs to the model network convey information about the
rat's velocity and heading, consistent with known inputs projecting
into the dMEC. The network effectively integrates these inputs to produce a
response that depends on the rat's absolute position. We show that such
a neural network can integrate position accurately and can reproduce
grid-cell-like responses similar to those observed experimentally. We then
suggest a set of experiments that could help identify whether our suggested
mechanism is responsible for the emergence of grid cells and for path
integration in the rat's brain.
| Since the discovery of grid cells in the dorsolateral band of the medial entorhinal
cortex (dMEC) [1], several ideas have been put forth on how
grid-cell activity might emerge [2]–[7]. The theoretical ideas
suggested so far fall into two categories. In continuous attractor models (see [8]–[15] and [2],[4],[7] for the grid cell system), which are the focus of
this work, grid cell activity arises from the collective behavior of a neural
network. The network's state is restricted to lie in a low-dimensional
continuous manifold of steady states, and its particular location within this
manifold is updated in response to the rat's velocity. In the second
category of models [5],[6],[16],[17], grid-cell activity arises independently in
single cells, as a result of interference between a global periodic signal and a
cell-specific oscillation, whose frequency is modulated by the rat's
velocity.
These ideas differ radically from each other, but they share a common assumption
about the nature of the input feeding into dMEC, namely, that the input conveys
information primarily on the rat's velocity and heading. Within all these
models, grid cell activity must then arise from precise integration of the
rat's velocity.
Grid cell firing exhibits remarkable accuracy: The periodic spatial tuning pattern
remains sharp and stable over trajectories lasting 10's of minutes, with an
accumulated length on the order of hundreds of meters [1]. Experiments performed
in the dark show that grid cell tuning remains relatively accurate over ∼100
meters and ∼10 minutes even after a substantial reduction of external
sensory inputs. However, in these experiments olfactory and tactile cues were not
eliminated, and grid cell responses may have been informed by positional information
from such cues. Therefore, the duration and length of paths over which coherent grid
responses are maintained without any external sensory cues is not known. For
position estimation on the behavioral level, we searched for but found no clear
quantitative records of the full range over which rats are capable of accurate
dead-reckoning. Behavioral studies [18]–[21]
document that rats can compute the straight path home following random foraging
trajectories that are 1–3 meters in length, in the absence of external
sensory cues.
How do theoretical models measure up, in estimating position from input velocity
cues? The theta-oscillation model of grid cells [5],[6],[16],[17], under idealized
assumptions about internal connectivity, velocity inputs, and neural dynamics, is
not able to produce accurate spatial grids over the known length- and time-scales of
behavioral dead-reckoning if the participating theta oscillations deviate from pure
sine waves. This is because the model is acutely vulnerable to subtle changes in the
phase of the underlying oscillations. In reality, theta oscillations are not
temporally coherent: cross-correlograms from in vitro intracellular
recordings [17],[22],[23] and in vivo extracellular
recordings [24],[25] show that the phase of the theta oscillation in
the entorhinal cortex typically decoheres or slips by half a cycle in less than 10
cycles or about 1 second, which corresponds to a distance of only 1 meter for a run
velocity of 1 m/s. This means that the model grid cells will entirely lose track of
the correct phase for the present rat position within that time.
For continuous attractor models, we previously showed [3] that due to rotations and
non-linear, anisotropic velocity responses, a detailed model [2] integrates velocity
poorly, and does not produce a grid-cell firing pattern even with idealized
connectivity and deterministic dynamics. Another model [7] generates grid
responses in a small periodic network, but it includes no neural nonlinearities or
variability in neural responses, and depends on real-time, continuous modulation of
recurrent weights by the velocity inputs to the network.
Conceptually, the existence of an integrating apparatus seems pointless if it is
completely dependent on nearly continuous corrections coming from an external source
that specifies absolute position. Thus, it seems reasonable to require that
theoretical models of path integration in dMEC, if using faithful velocity inputs,
have the ability to reproduce stable grid cell patterns for trajectories lasting a
few minutes.
Our aim, therefore, is to establish whether it is possible for model grid cells to
accurately integrate velocity inputs. We restrict our analysis specifically to
continuous attractor networks. As will become clear, the precision of velocity
integration can strongly depend on various factors including network topology,
network size, variability of neural firing, and variability in neural weights. Here
we focus on three of these factors: boundary conditions in the wiring of the network
(periodic vs. aperiodic), network size, and stochasticity in neural activity
We quantify path integration accuracy in both periodic and aperiodic recurrent
network models of dMEC, and demonstrate that within a biologically plausible range
of parameters explored, such networks have maximum attainable ranges of accurate
path integration of 1–10 minutes and 10–100 meters. Larger, less
noisy networks occupy the high end of the range, while smaller and more stochastic
networks occupy the low end. We end with suggestions for experiments to quantify
integration accuracy, falsify the continuous attractor hypothesis, and determine
whether the grid cell response is a recurrent network phenomenon or whether it
emerges from computations occurring within single cells.
In our model, each neuron receives inhibitory input from a surrounding ring of local
neurons. The entire network receives broad-field feedforward excitation (
Methods
). If the inhibitory interactions are sufficiently strong, this type of
connectivity generically produces a population response consisting of a regular
pattern of discrete blobs of neural activity, arranged on the vertices of a regular
triangular lattice [3],[4],[26], Figure 1A. Ignoring boundary
effects for the moment, all possible phases (translations) of the pattern are
equivalent steady states of the pattern formation process, and therefore form a
continuous attractor manifold.
To reproduce the regular single-neuron (SN) lattice patterns observed in experiment,
the pattern formed in the neural population must be coupled to the rat's
velocity. This coupling is arranged in such a way (Figure 1B and
Methods
) that it drives translations of the pattern within the neural sheet, in
proportion to the movements of the rat in real 2-d space, Figure 1C.
Briefly, velocity coupling involves distributing a set of direction labels () to the neurons in any patch of the network (Figure 1B). The direction label signifies that (1) the neuron receives input from a
speed-modulated head-direction cell tuned to that direction, and (2) the
neuron's outgoing center-surround connectivity profile is centered not on
itself, but is shifted by a few neurons along a corresponding direction on the
neural sheet. The neuron tends, through its slightly asymmetric connectivity, to
drive network activity in the direction of the shift. However, another neuron with
the opposite direction preference will tend to drive a flow in the opposite
direction. If all neurons have equal inputs, the opposing drives will balance each
other, and the activity pattern will remain static. If, however, the rat moves in a
particular direction in space, the corresponding model dMEC cells will receive
larger input than the others, due to their head-direction inputs, and will succeed
in driving a flow of the network pattern along their preferred direction. This
mechanism for input-driven pattern flow is similar to that proposed in a model of
the head-direction system [14]. Figure 1C demonstrates how a flow of the population pattern will drive
activity at spatially periodic intervals in single neurons.
To obtain spatially periodic responses in single neurons over long, curved,
variable-speed trajectories, additional conditions must be met, as we discuss below.
We present results from two topologically distinct networks: one with aperiodic, and
the other with periodic, connectivity.
We simulate dynamics in a network of neurons driven by velocity inputs obtained
from recordings of a rat's trajectory (see
Methods
). The network contains 1282 (∼104) neurons
arranged in a square sheet. Neurons close to each edge of the sheet form
connections with neurons on the opposite edge, such that the topology of the
network is that of a torus. Figure
2A shows the population activity in the network at one instant of the
run.
A grid cell response, as reported in experimental papers, is obtained by summing
the firing activity of a single neuron over a full trajectory. Unlike the
population response, which is an instantaneous snapshot of full neural
population, the single-neuron response is an integrated measure over time of the
activity one cell. In the rest of this paper, SN response refers to the
accumulated response of single neurons over a trajectory.
In the periodic network, the SN response, accumulated over the ∼20 minute
trajectory, and plotted as a function of the true rat position, shows coherent
grid activity, Figure 2B.
The network accurately integrates input velocity, as can verified directly by
comparing the cumulative network pattern phase to the rat's true
position, Figure 2C. The
total error, accumulated over ∼260 m and 20 minutes, is <15 cm,
compared to a grid period of about 48 cm. This corresponds to an average
integration error of less than 0.1 cm per meter traveled and less than 0.01 cm
per second traveled. The range of rat speeds represented in the input trajectory
was 0–1 m/s, showing that this network is capable of accurate path
integration over this range of speeds.
A deterministic periodic network of only 402 (∼103)
neurons also performs well enough to produce coherent SN grids over the same
trajectory, Figure S1.
The presence of a clear spatial grid in the SN response to velocity inputs alone
is a good indication of the accuracy of integration. If the rat's
internal estimate of position were to drift by half a grid period, the neuron
would fire in the middle of two existing vertices rather than on a vertex. As
the rat traveled over its trajectory, the neuron would fire at various
“wrong” locations, with the resulting SN response becoming
progressively blurred until no grid would be discernible. This would happen even
if the population pattern remained perfectly periodic throughout.
Therefore, the following properties are equivalent: (1) Coherent grids in the SN
responses, (2) Accurate path integration of the full trajectory over which the
SN responses are visualized, with errors smaller than the grid period. An
example of this equivalence is given in Figure 2A and 2C, which show sharp SN
patterning and a very small integration error.
Next, because the population pattern phase accumulates errors whenever the
pattern slips relative to rat motion, another equivalent condition for accurate
path integration is (3) Linear relationship between network flow velocity and
input velocity over the input velocity range, independent of direction.
These equivalent conditions for accurate integration apply to both periodic and
aperiodic network models of grid cells (discussed next).
It is unclear whether a torus-like network topology, in which neurons along
opposite edges of the network are connected to form periodic boundary
conditions, exists in the rat's brain. Even if such connectivity
exists, it may require, at an earlier stage of development, an initially
aperiodic network (see Discussion). Hence
it is interesting to consider whether a network with non-periodic boundaries can
produce grid-cell like SN activity. The difficulty here is that as the
population pattern flows in response to velocity inputs, it must reform at the
boundaries of the neural sheet. Newly forming activity blobs must be created at
accurate positions, and the process must not interfere with the
pattern's flow.
A central result of the present work on aperiodic networks is that such networks
can, in fact, accurately integrate velocity inputs. With an appropriate choice
of architecture and inputs and with deterministic dynamics, an aperiodic network
can produce SN responses that are as accurate as in the periodic case above.
This is illustrated in the example of Figure 2D–F. At the aperiodic
boundaries, the same dynamics that governed the initial pattern formation
process also cause the pattern to continually regenerate as the pattern flows
(Figure 1C, bottom). The
phases or locations of the renewing blobs at the boundary are consistent with
the rest of the network pattern, in part because their placement is influenced
by inhibition from the neighboring active neurons in the network interior.
For the two types of networks from the previous section, the structure of the
state-space is schematically illustrated in Figure 4. The state-space illustration is
instrumental in synthesizing the findings of the preceding section –
in particular: Why does the pattern not rotate in the periodic network? Why is
the pattern pinned at low input velocities in the aperiodic network? Why does
network size matter more for aperiodic than for periodic networks? We assume
that the dynamics minimize an energy functional, whose local minima correspond a
set of fixed points (attractors) (This assumption is precisely correct in the
absence of a velocity-driven shift mechanism, since the connectivity matrix is
then symmetric [27],[28].)
Consider first the periodic network. Starting from a steady
state of the dynamics, and rigidly translating the stable population pattern,
produces an equivalent steady state with exactly the same energy. The set of all
such states forms a continuous manifold of attractor states, related to each
other by continuous translation. This manifold can be visualized as the trough
of the energy surface, Figure
4A. Rotating a steady state pattern, on the other hand, produces states
with higher energy. (Rotation can be visualized as follows. Imagine first
cutting open the toroidal periodic network along the edges of the sheet that
were originally glued together to produce a periodic network. On the resulting
sheet, rotate the pattern, and rejoin the cut edges. This procedure will produce
discontinuities in the pattern along the rejoined edges.) Hence the attractor
manifold does not include continuous rotations.
Inputs that induce pattern translation will stably move the network state along
the trough, even if the inputs are small, and the integrated value of the input
will be reflected in the updated network phase. On the other hand, inputs that
attempt to induce rotations will not produce lasting changes in network state,
because these states are unstable and will quickly (over a few hundred
milliseconds or less) decay as the pattern relaxes to its preferred orientation.
Similarly, distorting the pattern by stretching it, adding noise, or by removing
blobs from the pattern will generate an unstable state, which will rapidly decay
to a steady state within the attractor manifold.
In the aperiodic network, translations of a steady state pattern
are similar but not exactly equivalent, because the phase of the activity
pattern relative to the boundary affects the energy of the state. Strictly
speaking then, these states do not form a continuous attractor manifold, Figure 4B. Instead, the
manifold is slightly rippled along the direction of translations. To drive
translations, velocity inputs must be large enough to overcome the ripple
barrier. This explains why below a critical velocity, the pattern is pinned in
our simulations. The ripple amplitude depends on how much influence the boundary
has on the network dynamics. If activity fades to zero sufficiently smoothly
near the boundary the ripple can be small. Pattern translation then corresponds
to motion along a nearly flat direction on the manifold, pinning is confined to
a negligibly small range of velocities, and integration of inputs can be
accurate. A reduction of pinning can be achieved also by increasing the network
size, while keeping the boundary profile fixed, because boundary effects scale
as the ratio of network periphery to network area.
A stable population pattern state can be rotated around the center of a circular
aperiodic neural sheet to obtain another stable state that is identical in
energy to the original one. Hence, rotations correspond to a flat direction in
the energy surface, Figure
4B. Any input that couples even slightly with the rotational mode can
drive rotations in the network pattern. The velocity inputs to the network,
though configured to drive translational pattern flow, can weakly drive
rotations due to boundary effects that couple the translational drive to
rotational modes. In spiking networks, discussed below, rotations can be driven
also by noise.
In the network models described here, the structure of the attractor manifold
(e.g., Figure 4A or 4B) is
completely determined by the matrix of pairwise weights between neurons and the
inputs received by each neuron. Once the weights between all pairs of neurons
and the inputs to each neuron are specified, the matrix does not change if the
locations of the neurons on the cortical sheet are shuffled, so long as the
weights and inputs to each neuron are held fixed (see
Discussion
). Thus, statements about the existence of a manifold of stable network
states and stable SN grid responses, and the predictions that stem from them, do
not depend on topography, even when stated here for expositional simplicity in
terms of topographically arranged population-level patterns.
So far we have considered errors in integration that occur in the absence of
noise. Unlike in the noise-free case, neural noise can induce the population
pattern to flow or rotate even when velocity inputs are absent. To assess how
noise influences the precision of the network's response, we present
results from spiking neural networks with the same connectivity as in the rate
based models. Dynamics in these networks are noisy due to the stochasticity of
discrete spiking events.
For the same network parameters as in Figure 2, and assuming that neural firing is
an inhomogeneous Poisson process, we find that the periodic network continues to
perform well enough to produce coherent SN responses over long trajectories
(Figure 5A and Figure S3).
In the aperiodic network, performance with Poisson spiking neurons is
considerably worse than in the rate based model, enough to destroy the grid-like
SN response over a ∼130 meter, 10-minute trajectory, in particular due
to rotations (Figure S3). Network performance improves, however, if spiking in the
network is more regular than implied by inhomogeneous Poisson statistics. To
quantify this effect, we performed simulations with sub-Poisson statistics (see
Methods
). The variance of neural firing is characterized, in our simulations, by
the coefficient of variation (CV) of the inter-spike interval. With a
sufficiently low CV, aperiodic network dynamics are precise enough to produce a
coherent SN response over a trajectory lasting 10 minutes and ∼130
meters, Figure 5B and Figure S3.
Armed with the proof-of-concept results that a continuous attractor network model
can integrate velocity inputs accurately enough to produce SN grids, we next
seek to explore testable predictions of the continuous attractor hypothesis in
the grid cell system and contrast them with the properties of models in which
the grid responses emerge independently in each cell [5],[6],[16]. Unless explicitly
specified, all proposed tests are intended for conditions in which external,
spatially informative cues have been removed.
The three main contributions of this work are:
So far, the predictions of continuous attractor models are consistent with the full
corpus of grid cell data, and explanatory of many results from experiment,
suggesting, when combined with conclusion (1), that continuous attractor dynamics
are a viable, relevant mechanism for grid cell activity and path integration.
Accurate behavioral dead reckoning is a cascaded result of accurate velocity
input (relative to the rat's motion) and accurate integration of that
input. Our interest in this work was in assessing how well continuous attractor
models of dMEC can integrate their inputs. Thus, we did not focus on potential
inaccuracies (noise or biases) in the velocity inputs themselves. Even if the
network were a perfect integrator, errors in the input would produce an
incorrect position estimate. Such errors are likely to play a role in reducing
the behavioral range over which rats display accurate dead-reckoning.
A strength of attractor networks is that responses are self-averaging over the
full network: if the velocity inputs are unbiased estimators of rat movements,
but are noisy, or if the velocity inputs to the network are not perfectly
balanced in number for all directions, the full network will average all its
inputs, and the net pattern flow will only reflect this average. For accurate
position estimation, however, it is important and therefore likely that inputs
to the network are well tuned.
Another factor that could degrade integration performance is inhomogeneity or
stochasticity in the recurrent network weights. While stochasticity in neural
activity causes the network state to drift along the attractor manifold,
variability in network connectivity modifies the structure of the attractor
manifold itself. If recurrent connectivity deviates significantly from the
translation-invariant form needed to ensure that all translations of the pattern
are accessible without crossing over energy barriers, the activity pattern can
become pinned at particular phases [38], reducing the
fidelity of the network response to small velocity inputs.
Because knowledge about synaptic strengths in the brain is exceedingly limited,
it is unclear what level of variability should be expected in dMEC weights, and
whether this amount is sufficient to cause significant pinning. A question for
theory, not addressed in this work, is to estimate the amount of variability in
the network weights that would be sufficient to reduce the accuracy of
integration below that observed in dead reckoning behavioral experiments. For
experiments, the difficult challenge is to obtain an estimate of variability in
dMEC connectivity.
The network size estimate in our continuous attractor model
(103–104 neurons) may be viewed as a
wasteful proposed use of neurons, but it is broadly consistent with estimates
for the total number of neurons in the entorhinal cortex [39]–[41]. By contrast, independent neuron models [5],[6],[17],
which do not require populations of neurons to produce grid cell responses, make
far more parsimonious use of neurons. In such models, a natural question is to
understand what function may be served by the large number of neurons in dMEC.
Within dMEC, the breakdown of total neural allocation, between neurons per grid
network versus the number of different grid networks, is unknown. dMEC might
consist of a very large number of very small networks with different grid
periods, which is optimal for representational capacity [42]. (For a fixed neuron
pool size, the addition of neurons per grid at the expense of the total number
of different grids causes a large capacity loss [42].) But the dynamical
considerations presented here suggest otherwise, because accurate path
integration in each grid requires many neurons. In contradiction to optimal
capacity considerations, therefore, continuous attractor models predict a large
membership in each grid network, and correspondingly few different grids.
A fascinating question is whether the discrete islands of cells observed in
anatomical and imaging studies of cells in layer II of the human and primate
entorhinal cortex [41], [43]–[46], as
well as indications in rodents for modular structure in dMEC [46],[47] correspond to
separate attractor networks, in which case the number of different grid periods
can be directly inferred.
We have shown that both periodic and aperiodic networks can perform accurate
integration. Which topology is dMEC likely to posses? The models and results of
this work are largely agnostic on this question. However, the aperiodic network
requires fine-tuning of its parameters to perform nearly as well as an untuned
periodic network. Even after fine-tuning, integration in the periodic network
tends to be better, because unlike in the aperiodic case, the population pattern
cannot rotate. Thus, from a functional perspective, periodic boundaries are
preferable over aperiodic ones.
Other constraints on network topology may stem from the developmental mechanism
of the grid-cell network. Such developmental constraints could overrule
potential functional preferences, in determining network topology.
If neural locations in the cortical sheet are scrambled, while preserving the
neural indices and the pairwise weights between neurons, the grid-like patterning in the cortical
sheet will disappear, but there will be no change in the single neuron
triangular lattice response or in any other dynamical property of the network.
The underlying structure of the attractor manifold (e.g., whether or not it is
continuous) is a function of network connectivity, but does not depend on the
layout of neurons on the cortical sheet. Thus, the lack of topography observed
in experiments, in which neighboring neurons have different phases, is not a
problem for the dynamics of continuous attractor models of grid cell activity.
Instead, the problem is one of learning: how does a network wire up so that the
intrinsic structure of the weight matrix resembles center-surround connectivity,
but the neurons are themselves not arranged topographically in space?
A topographic, aperiodic model network would have relatively simple wiring rules
(if we ignore the directional neural labels and corresponding segregation of
head-direction inputs and shifts in the outgoing weights required for the
velocity-coupling mechanism): each neuron would simply have spatially restricted
center-surround interactions with its neighbors. This has prompted the
observation that such a topographic network could serve as a starting point for
the development of a network with a less topographical layout and periodic
boundaries [4]. For instance, the proposal by [4]
for wiring an atopographic and periodic network is based on three assumptions:
(1) that another area, the ‘teacher’, contains an initial
aperiodic, topographic network with population grid patterning and no velocity
shift mechanism, (2) that the network pattern, when subject to intrinsic or
extrinsic noise, tends to translate without rotation, (3) that the network
projects through spatially random connectivity to the naive dMEC, and
activity-dependent activity mechanisms within dMEC cause neurons that are
coactivated by the teacher network, to wire together. However, results from the
present work show that the fundamental features of aperiodic networks pose a
problem for such a scheme.
We showed that the population pattern in a deterministic aperiodic network fully
equipped with a translational velocity shift mechanism and driven by purely
translational velocity inputs, tends to rotate within a few minutes. This is the
short end of the time-scales over which plasticity mechanisms for network
development would act. If the network is entirely driven by noise and lacks a
specific velocity shift mechanism (as in [4]), the problem is
far worse: undesirable rotations become as likely as translations, and the
pattern orientation can decohere in seconds, invalidating assumption (2). Thus,
the precursor network pattern will not be able to entrain a periodic grid in the
target network.
The problem of pattern rotations over the time scale of learning is pertinent for
any effort to produce a periodic network from an initially aperiodic one in the
absence of anchoring sensory inputs and a velocity coupling mechanism.
The concept of low-dimensional continuous attractors has influenced our
understanding of neural systems and produced successful models of a number of
neural integrators [8]–[10],[13],[14],[48],[49].
Yet proof of continuous attractor dynamics (or some discrete approximation to
continuous attractor dynamics) in the brain has remained elusive: experiments in
supposed continuous attractor systems have failed to unearth evidence to
conclusively validate or falsify the continuous attractor hypothesis. The
relative richness (e.g., size, dimensionality of the manifold) of the grid cell
response compared to other possible continuous attractor systems may provide a
more structured and unambiguous testing ground for predictions stemming from the
continuous attractor hypothesis. Testing of these predictions, many based on
cell-cell correlations, is feasible with existing experimental technologies, and
such tests may help to determine whether a low-dimensional continuous attractor
is central to the dynamics of the grid cell system.
The dynamics of rate-based neurons is specified by:(1)
The neural transfer function is a simple rectification nonlinearity: for , and is 0 otherwise. The synaptic activation of neuron is ; is the synaptic weight from neuron to neuron . The time-constant of neural response is
τ = 10 ms. The time-step
for numerical integration is
dt = 0.5 ms.
We assume that neurons are arranged in a 2-d sheet. Neuron is located at . There are neurons in the network, so ranges from (,) to (,). We use in all figures except where specifically indicated. Each neuron also has a preferred direction (W, N, S, E) designated by . Locally, each 2×2 block on the sheet contains one
neuron of each preferred direction, tiled uniformly.
The preferred directions are restricted to N,S,E,W for convenience in modeling; in
the rat, these preferences might span the continuum . The preferred orientation of a neuron is used to (1) determine
the direction in which its outgoing weighs are shifted, and (2) determine the rat
velocity inputs it receives.
The recurrent weight matrix is(2)with(3)The weight matrix has a center-surround shape, but is centered at the
shifted location . Implicit in the form of the weight matrix, where connectivity is
a function of neural separation, is the assumption that neurons are topographically
arranged. This is not a necessary requirement (see
Discussion
), but does greatly facilitate visualization and presentation. In all
simulations, we used , , and where is approximately the periodicity of the formed lattice in the
neural sheet. With , all connectivity is inhibitory; thus, local surround inhibition
alone is sufficient to reproduce gird cell responses, but the network could include
excitatory interactions () without qualitatively affecting the results.
The feedforward input to neuron is(4)where is the unit vector pointing along , and is the velocity vector of the rat, measured in m/s. If (Eq. 2) and (Eq. 4), the network generates a static triangular lattice
pattern, Figure 1A, with overall
intensity modulated by the envelope function (e.g., Figures
2D, 3B, and
3D1–D4).
If are non-zero, they allow rat velocity () to couple to the network dynamics, and drive a flow of the formed
pattern. The magnitudes of both and multiplicatively determine how strongly velocity inputs drive the
pattern, and thus control the speed of the flow of the pattern for a fixed rat
speed. The triangular lattice pattern is only stable for small values of the shift in the outgoing weights, thus we keep fixed so that the outgoing weights are shifted 2 neurons. With fixed, determines the gain of the velocity response of the network. If , we can expect the velocity inputs to drive pattern flow without
destroying the stability of the formed lattice. In the plots shown, . The grid spacing of the SN response is ultimately determined by
two factors: (i) The grid spacing of the population response, which is set by the
shape of the symmetric weight matrix , and (ii) the gain of the network's flow response to a
velocity input, which depends on and .
The envelope function spatially modulates the strength of the inputs to the neurons, and
can scale neural activity without disrupting the lattice pattern. This can be seen
from Equation 1: if the input is uniform, then scaling is equivalent to scaling . It is important to observe that the velocity inputs must also be
modulated by the envelope , Eq. 4, to insure the same flow rate in the faded regions as in
the bulk. This is because the local flow rate is given by the velocity-modulated
component of the feedforward input divided by the total feedforward input.
For the network with periodic boundary conditions, the envelope function is 1
everywhere. For the aperiodic network,(5) is the diameter of the network and (for example, see Figure 2D and Figure 3B,
D1–D4). In Figure
3 (D4), R = 128; in all other figures,
R = 64. The parameter determines the range of radii over which input tapering occurs:
The larger , the more gradual the tapering. In all the aperiodic simulations , except for Figure 3
(A–C and D2, D4), where and Figure 3
(D3), where .
To simulate a Poisson process (CV = 1, where CV
is the ratio of the inter-spike interval standard deviation with the mean), in
each time-step neuron spikes with probability given by (in our simulations, is always much less than , ensuring that ). The synaptic activation is computed from neural spiking: it increments by 1 at time if neuron spiked at , and otherwise decays according to(6)The process for generating spike trains with (for integer-valued ) is similar to that for generating a Poisson train. We first
subdivide each interval into sub-intervals of length each, and simulate on this finer time resolution a fast
Poisson spiking process with rate . We then decimate the fast Poisson process, retaining every
m-th spike and discarding all the other spikes. This
procedure generates a spike train with rate and .
Aperiodic network: initially network activity is low; neurons receive external
input with in addition to a small independent random drive, which leads
to spontaneous pattern formation. Periodic network: we initialize an aperiodic
network with otherwise identical parameters, and after pattern formation apply
periodic boundary conditions. The parameters for the aperiodic network have to
be chosen to be commensurate with the size of the network to avoid excess strain
and the formation of defects when the boundaries are made periodic. We flow both
the periodic and aperiodic network states with unidirectional velocity inputs,
corresponding to a velocity of 0.8 m/s, in three different directions (0,,) for 250 ms each to heal any strain and defects in the formed
pattern. After this healing period, we give as input to the network either real
rat velocity (data obtained by differentiating recorded rat trajectories
– published in [1] – then linearly interpolating
between the recording time-steps and the time-step in our simulations), or a sequence of velocity steps
(described next).
The network is initialized to the exact same initial template state at the
beginning of each step (using a template pattern stored following one run of the
initialization process described above). Each step consists of a constant
velocity input, with one of four directions (0, , , ). The velocity is incremented in steps of 0.02 m/s. We use
only the second half of the 5 s long steps to compute the network's
velocity response.
We track how far the pattern has flowed beyond a lattice period and beyond the
scale of the network by continuously recording the velocity of the blob closest
to the center, and integrating the obtained velocity. We track the orientation
of the lattice by computing its Fourier transform and recording the angles of
the three blobs closest to the origin in Fourier space.
To assign units of centimeters to the accumulated network pattern flow and
compare it to rat position (Figure
2C, 2F, 3C, Figure S1,
and Figure
S3), we must obtain the scale factor relating the network pattern flow
velocity to the velocity of the rat. The scale is determined by optimizing the
match between network flow velocity and the derivative of the rat position
throughout the simulation. The offset is set so that the network drift at time is zero.
|
10.1371/journal.pntd.0005838 | Estimating the burden of scrub typhus: A systematic review | Scrub typhus is a vector-borne zoonotic disease that can be life-threatening. There are no licensed vaccines, or vector control efforts in place. Despite increasing awareness in endemic regions, the public health burden and global distribution of scrub typhus remains poorly known.
We systematically reviewed all literature from public health records, fever studies and reports available on the Ovid MEDLINE, Embase Classic + Embase and EconLit databases, to estimate the burden of scrub typhus since the year 2000.
In prospective fever studies from Asia, scrub typhus is a leading cause of treatable non-malarial febrile illness. Sero-epidemiological data also suggest that Orientia tsutsugamushi infection is common across Asia, with seroprevalence ranging from 9.3%–27.9% (median 22.2% IQR 18.6–25.7). A substantial apparent rise in minimum disease incidence (median 4.6/100,000/10 years, highest in China with 11.2/100,000/10 years) was reported through passive national surveillance systems in South Korea, Japan, China, and Thailand. Case fatality risks from areas of reduced drug-susceptibility are reported at 12.2% and 13.6% for South India and northern Thailand, respectively. Mortality reports vary widely around a median mortality of 6.0% for untreated and 1.4% for treated scrub typhus. Limited evidence suggests high mortality in complicated scrub typhus with CNS involvement (13.6% mortality), multi-organ dysfunction (24.1%) and high pregnancy miscarriage rates with poor neonatal outcomes.
Scrub typhus appears to be a truly neglected tropical disease mainly affecting rural populations, but increasingly also metropolitan areas. Rising minimum incidence rates have been reported over the past 8–10 years from countries with an established surveillance system. A wider distribution of scrub typhus beyond Asia is likely, based on reports from South America and Africa. Unfortunately, the quality and quantity of the available data on scrub typhus epidemiology is currently too limited for any economical, mathematical modeling or mapping approaches.
| Scrub typhus is a mite-transmitted infectious disease that can be life-threatening. Diagnosing this disease is difficult, requiring special techniques that are often not readily available. As the actual impact of scrub typhus on the population and its geographical distribution remains unknown, we searched systematically for available information in medical databases. Scrub typhus is common: more than every fifth person in areas where scrub typhus occurs carry antibodies as a sign of previous contact. All countries with an established surveillance system have recorded an increase in scrub typhus cases over the past 8–10 years, while reports from South America and Africa suggest a wider distribution beyond Asia. Scrub typhus is a serious disease: approximately 6% of cases die if untreated, and 1.5% if treated, but mortality can reach 13% in areas where the usual treatment does not always work well. Death rates of complications are higher, reaching 14% in brain infections, 24% with multiple organ failure, and pregnancies with scrub typhus can have poor outcomes, with high miscarriage rates. Despite many limitations on the amount and quality of available reports, we found that scrub typhus is a severely underappreciated tropical disease, affecting mainly rural populations, but increasingly urban areas as well.
| Scrub typhus is an infectious disease caused by Orientia tsutsugamushi, an obligate intracellular bacteria, transmitted by the bites of chigger mites [1]. In Southeast Asia, scrub typhus is a leading cause of treatable non-malarial febrile illness [2]. The first accounts linking febrile illness with the appearance of “harmful” mites (Japanese: “tsutsuga” mushi) range back to 313 AD in China [3]. Scrub typhus was originally associated with the Asian-Pacific “Tsutsugamushi triangle,” until recent evidence from the Arabian Peninsula, Chile and possibly Kenya suggested a wider global distribution in tropical and subtropical regions [4–7].
The use of improved diagnostic methods, increased medical investigations and awareness have recently contributed to greater recognition of scrub typhus in some countries, such as in Laos, India, southern China, South Korea, and Japan [8]. There is also evidence suggesting that a combination of climate change and expansion of humans into previously uninhabited areas may play a role in both re-emergence and apparent rising incidence of scrub typhus [9–11].
There are no licensed vaccines for scrub typhus, and no systematic vector control efforts in place. Despite increasing awareness in endemic regions, the public health burden and global distribution of scrub typhus remains poorly known.
Although scrub typhus received much attention before and during the Second World War and to a lesser degree during the Vietnam/American war, basic epidemiology is poorly understood with limited data on incidence and burden of disease for patients, their families, societies and the economy. This ignorance is probably due to a combination of factors; clinical presentation is very similar to other causes of fever, diagnostic difficulties contribute to mis-diagnosis and under recognition, and appropriate diagnostic tests are not widely available. Following the discovery of chloramphenicol in the 1940s, the scientific interest dropped rapidly and scrub typhus has since received little global attention [12]. The data quoted by the World Health Organization (WHO) stating that over a billion people are at risk and one million cases are estimated per year is referenced to a paper published 20 years ago in 1997 [13, 14].
Extrapolation based on geographical mite distributions and densities are not helpful due to patchy data, limited by the dynamics of infected mite populations and insufficient characterization of transmitting vectors. With new data and improvements in approaches to estimating the burden of febrile illnesses, it is important to reevaluate the burden of scrub typhus.
Rationale for this study: Scrub typhus is among the leading causes of undifferentiated treatable fever in Asia. The mortality rates appear low at first glance, but considering the numbers of those exposed and/or infected a significant disease burden is expected globally. The following research questions were addressed: What is the estimated global burden of disease for scrub typhus? What data on seroprevalence and minimum incidence for scrub typhus are available by geographical regions? What data on DALYs, YLLs and YLDs are available, and what is the mortality rate of treated scrub typhus?
In this study we summarized the literature relating to the disease burden and economic impact of scrub typhus since the year 2000 in order to estimate the global incidence and burden of this disease.
A literature search of three databases: Ovid MEDLINE (2000-present), Embase Classic + Embase (2000-present) and EconLit (2000-present) was conducted on 11th April 2016 using three search strategies. First search terms: Scrub typhus, Orientia tsutsugamushi, Rickettsia tsutsugamushi, chigger borne rickettsiosis, chigger borne typhus, Orientia tsutsugamushi infection, Rickettsia tsutsugamushi infection, tsutsugamushi disease, tsutsugamushi fever (keyword) AND prevalence, incidence, epidemiology. A second search included the above search for scrub typhus and all variations AND cost, cost analysis, cost of illness, drug costs, economics, health care cost, hospital costs, cost benefit analysis, cost effectiveness analysis, quality adjusted life year. A third search included scrub typhus AND mortality or death on the 1st Oct 2016, for which all currently available data was included (no date restrictions). Data on untreated mortality have been reported [15], and therefore only papers with treated infection were included in estimating mortality. All titles and abstracts were reviewed by 2 authors for inclusion and any disagreements were discussed and inclusion based on the senior author’s opinion. Only English language publications were included.
A total of 190 publications were selected for full article review (Fig 1A and 1B, S1 File). The final number of articles included for full data extraction was 87. The data extraction form was trialed on the first 5 papers and required minor alterations. Due to the limited nature of data available no summary measures were applied. Studies were examined for selection bias and graded as follows:
Papers were also graded on diagnostic tests used:
Of the 87 studies included, 44 (50.6%) gave information on incidence, seroprevalence and/or prevalence in febrile inpatients (denominator = febrile cases per year), whilst health economic or burden of disease data were given in 4 studies (4.6%) and mortality data in 38 (43.7%). The publications with no apparent patient selection bias and use of grade A evidence to diagnose scrub typhus were few (16/87, 18.4%). A total of 143,544 patients with scrub typhus were described in the included papers. Females were reported to be more commonly infected than males 77,204 versus 57,535 (57.3% versus 42.7%, respectively).
Five countries report a passive national surveillance system for scrub typhus.
In South Korea scrub typhus was designated a group III notifiable disease (requiring mandatory reporting and routine monitoring) in 1994. Cases are confirmed by the Korean Centre for Disease Control and Prevention (KCDC) and must show one of the following: an increase in the IFA IgM to O. tsutsugamushi of ≥ 1:16; an increase in the anti-O. tsutsugamushi IFA IgG titre to ≥1:256; a ≥ 4 fold increase in IFA titre. Data from KCDC suggest that the annual minimum incidence increased from 5.7 to 17.7/100,000 people from 2001 to 2012 (>3-fold) (Table 1) [16–18]. Interestingly, the number of patients recorded in urban areas has also increased dramatically, for example, the annual minimum incidence in Ulsan Metropolitan City increased from 2.8/100,000 in 2003 to 59.7/100,000 in 2013 (>21 fold). In Seoul there is evidence of urban scrub typhus, further demonstrating the changing geographical scope and habitat of infected chigger mites [16].
In Japan scrub typhus is a notifiable disease and must be reported to the National Epidemiological Surveillance of Infectious Diseases (NESID) within 7 days of diagnosis by a physician. Confirmed cases are based on: isolation or identification of the organism in the blood; PCR positivity; detection of serum IgM; a ≥ 4 fold increase in IFA titre. Data from NESID show an increase of annual minimum incidence from 0.6/100,000 in 2000 to 3.6/100,000 in 2008 (6-fold) [19, 20].
In Thailand, scrub typhus patients have been reported to the Bureau of Epidemiology for the last 30 years. Data can be viewed online on the homepage available under URL: http://www.boe.moph.go.th/boedb/surdata/disease.php?dcontent=situation&ds=44
Cases are defined based on one or more of the following: isolation or identification of the organism in the blood or tissue sample; PCR positivity; a ≥ 4 fold increase in IFA titre (IgG and/or IgM); ≥ 1:400 IFA in acute serum (IgG and/or IgM); IgM ELISA positivity. Data from the Bureau of Epidemiology noted an increase of annual minimum incidence from 6.0/100,000 in 2003 to 17.1/100,000 in 2013 (2.9 fold) [21].
In China, scrub typhus is a notifiable disease that must be reported to the China Center for Disease Control and Prevention. Cases are defined as those with clinically compatible infection and one or more of the following; isolation or identification of the organism in a blood or tissue sample; PCR positivity; a ≥ 1:160 Weil-Felix test; a ≥ 4 fold increase in IFA titre (IgG and/or IgM). The reported countrywide minimum incidence increased from 0.1/100,000 to 1.1/100,000 people/year from 2006 to 2014 (>11-fold) [22]. The reported incidence rates vary widely by region with the southern provinces more affected. Guangdong Province saw an increase in reported annual minimum incidence from 0.4/100,000 to 3.6/100,000 people from 2006 to 2013 (>8-fold), whereas in 2012 the provinces of Laiwu and Guangzhou City had annual incidences of 5.5/100,000 and 9.9/100,000 people, respectively [10, 23–25].
There are seroprevalence data available from Bangladesh, Indonesia, Laos, Malaysia, Papua New Guinea and Sri Lanka (Table 1). Seropositivity ranged from 9.3%–27.9% suggesting high background exposure levels to O. tsutsugamushi in these countries [26–31]
There are several case series describing the frequency of scrub typhus among patients presenting with fever. In India, scrub typhus was the causative agent in 16.1–96.9% of febrile patients presenting to hospitals (Table 2). However, these studies all suffer from selection bias, as other causes of febrile illness had already been excluded. Studies from Cambodia, Laos, Nepal, and Kenya were subject to less bias as they included complete prospective series of patients presenting with fever to healthcare facilities and demonstrated rates from 1.8–22.3% (Table 2).
Data from specific sub-populations are presented in Table 3. Two studies describe the importance of scrub typhus in women during pregnancy from Laos and the Thai-Myanmar border—with scrub typhus occurring in 3.6–5.4% of febrile patients [32–34]. Maternal infection with scrub typhus during pregnancy was associated with poor maternal and fetal outcomes; 2/9 (22.2%) of cases in Laos and 4/11 (36.4%) in Thailand/Myanmar suffered either abortion or stillbirth.
Among Lao patients with meningitis/encephalitis, 16.0% of those with a diagnosed bacterial cause for their infection had evidence for scrub typhus [35]. However, only 54.8% of these patients received treatment with appropriate antimicrobials during admission and the mortality rate associated with CNS complications was 13.6%. There are no data on morbidity or long-term sequelae available.
National surveillance data from patients in China, Japan, Korea and Taiwan suggest that the age group of 60–69 years was at highest risk of scrub typhus [18, 20, 22, 36]. In Thailand those aged 45–54 years were most commonly infected. In Japan and Thailand males were more at risk of scrub typhus but in all other countries with reports, females are more at risk. In South Korea, China, Taiwan and Thailand farmers were most at risk (38,183/54,558–70% of infections in China from 2006–2014); unfortunately such data are lacking from Japan. Age stratification in untreated mortality revealed increasing risk with increasing age, with the age classes 51–60 and >60 years old associated with a 45.6% and 59.8% mortality rate respectively [15].
The long-term impact of infection with scrub typhus has barely been examined. In Taiwan the hazard ratio of developing acute coronary syndrome was 1.4 (95% CI 1.1–1.8) in those with previous infection with scrub typhus compared to the general population without [37]. A recent case series from India that included patients with unexplained fever and/or multi-system involvement, found 24.4% to have scrub typhus, and 53.1% of patients with scrub typhus had acute kidney injury [38]. A retrospective cohort of severe scrub typhus cases admitted to an ICU in South India, found that respiratory complications requiring mechanical ventilation occurred in 87.9%, and that dysfunction of 3 or more organ systems occurred in 85.2% [39].
Case fatality ratios vary widely between countries, with those countries with easily accessible and established health systems showing lower mortality rates compared to countries with limited facilities (Fig 2 and Table 4). In a previous review, untreated scrub typhus infection was associated with an estimated mortality of 6.0% (median, range 0–70.0%) [15]. This review of treated scrub typhus, which included 39 studies and 91,692 patients found a median mortality of 1.4% (range 0–33.3%).
The burden of disease data for scrub typhus is highly limited. Only one study, from Laiwu Province in China, has calculated the DALYs associated with scrub typhus [24]. This study estimated that 13 DALYs were lost due to scrub typhus across the province (6 in males, 7 in females at a rate of 1.06/100,000). However, in this province no deaths were reported and therefore these data cannot be extrapolated to countries such as India or Laos with evidence of scrub typhus associated mortality. A South Korean study evaluating the net benefit of a scrub typhus prevention program, estimated the cost of scrub typhus (medication and hospital costs and loss of earnings) at $6.6 million per year in 2008 [56]. However, scrub typhus mortality in South Korea was only 0.14% and 75% of patients with a diagnosis were hospitalized. Therefore, these figures cannot be applied to other economically poorer countries where health practice is very different [17].
Scrub typhus represents a major cause of treatable febrile illness across Asia, but its disease incidence remains elusive. Fever remains one of the major reasons to seek healthcare in tropical regions but their causes remain ill-defined [57]. Access to updated evidence on incidence and trends for common causes of febrile illnesses is essential for guiding and informing global, regional, and national health policies. This systematic review collated all currently available literature regarding the disease burden and economic impact of scrub typhus and the result is sobering; there are very few studies and they have great heterogeneity in methodology.
Acquisition of estimates for incidence and mortality proved difficult, as numerators had varying levels of confidence in diagnosis or denominators were either absent, or required further extrapolation. Ideally, data derived from population-based surveillance studies would be graded considerably higher than from hospital-based surveillance, but unfortunately no non-hospital-based surveillance data are publicly available for scrub typhus—unlike for diseases like typhoid where these data are readily available for various countries [58]. Further, the epidemiology of scrub typhus within a country is heterogeneous–the pronounced seasonality of these diseases and the changing urban/rural distribution, with defined areas of high infected mite intensities (mite islands) challenge the common approaches of disease incidence evaluation [59, 60]. Febrile illness surveillance should be performed in multiple representative areas, ideally covering one full calendar year before inferences on national disease incidence can be made [58].
Only 5 countries have established scrub typhus surveillance systems. All of these have shown an increasing minimum incidence of scrub typhus over recent years, with increasing evidence of shift towards urbanized areas. However, the apparent increase in minimum incidence is confounded by local enhanced knowledge of the disease and it remains uncertain whether these data reflect true de novo emerging disease or emerging awareness of a pre-existing disease. Surveillance systems also use diverse diagnostic tests and therefore inter-country comparisons are not always possible. There are no data on whether these surveillance systems have been evaluated to determine an estimate of missed cases, however it is likely that the numbers are conservative estimates. Regardless of these flaws, surveillance systems are an essential part of disease control strategies. Improved febrile disease surveillance providing national data should be initiated in more afflicted countries, as this would result in morbidity and mortality data that could be used to direct healthcare resources, future vaccine demand and delivery and assessment of effectiveness of any control programs.
Clearly, striving towards improved surveillance should be key, with a focus on providing reliable numerators (using diagnostic assays with suitable sensitivities and specificities), and representative denominators (well-defined target populations). Additionally, no ‘multiplier data’ or ‘multiplier studies’ are available—these are considered to improve estimation of incidence by using healthcare utilization surveys and to correct for under-ascertainment in healthcare facility studies [58].
Seroprevalence data was available from 5 countries only–indicating high background exposure levels, and therefore a high probability that larger numbers of unidentified and/or asymptomatic infections occur. Disease seroprevalence data must be interpreted with caution due to unknown antibody dynamics over time and uncertainty as to whether those seropositive became sick or were asymptomatic. In scrub typhus, both humoral and cell-mediated protective immune responses wane over time, but detailed understanding of this remains elusive [61]. Moreover, the population-wide frequencies of patients with reversion to seronegativity and potential disease susceptibility remain unknown, and therefore the actual exposure in these studies is likely to be substantially higher [62].
Scrub typhus is a leading cause of treatable non-malarial febrile illness in prospective fever etiology studies (n = 14). An increasing number of studies have unraveled the major contribution of scrub typhus to the febrile illness burden. However, the large variation of scrub typhus rates in prospective fever studies (median 23.4% IQR 5.2–39.7 ranging from 1–96.9% depending on country and patient selection), reflect a lack of standardization and comparability among study designs and diagnostic modalities used. None of these studies have used modeling or extrapolation to take into account data from healthcare utilization surveys, which may give a more accurate idea of numbers of people with scrub typhus. In addition, recent studies have raised concern on the persistence of O. tsutsugamushi after treatment, especially using bacteriostatic drugs such as tetracyclines and macrolides [63, 64].
Based on very limited data, scrub typhus is likely to have considerable impact on vulnerable populations–the median untreated mortality of scrub typhus in the elderly was ~29%—approximately 5-fold higher compared to the overall population mortality of 6% [15]. In women with scrub typhus during pregnancy, miscarriages occurred in 17% and poor neonatal outcomes in 42% of cases, which is more severe than the consequences of malaria in pregnancy [65]. Further, the mortality in patients requiring a lumbar puncture for scrub typhus CNS complications in Laos was 14% [59]. Scrub typhus is usually an easily treatable disease and the majority of these complications could be prevented by early recognition/diagnosis and increased usage of empirical doxycycline [66].
It is difficult to draw any definitive conclusions from the case-fatality data due to the heterogeneity in studies. They range from national surveillance data to case series of those admitted to ICU. National surveillance data from China, Japan and Korea provide case fatality ratios of 0.068–0.26%. However, the health facilities in these countries are significantly more advanced than other endemic countries. The fever studies from South India provide estimates of case fatality risk, but they vary from 0–33.3%—importantly, these data included patients who presented to hospital and therefore will miss those that do not have severe disease.
DALY data are lacking in all countries except from one area of China, where a rate of 1.06/100,000 people was found, with a zero mortality rate. Case series and studies from Taiwan and India examining long-term complications, imply that the mortality and morbidity from scrub typhus is under-recognized and that possible long term consequences may occur many years later, and may be important contributors to the overall DALY burden [37, 67]. Despite scrub typhus being the foremost cause of treatable febrile illness in Asia it is not evaluated by the Global Burden of Disease studies [68].
This study involved an extensive search of the literature and includes up-to-date and relevant studies. However, there are several limitations; as English is not the native language in the majority of countries where scrub typhus is endemic, there is a potential bulk of relevant literature that is not indexed in the databases used. The risks of publication bias and the heterogeneity of methods and reporting in the articles limit the conclusions. Specific difficulties relating to the diagnosis of scrub typhus suggest that studies reporting data from national surveillance systems are likely to suffer from missing data due to those that do not seek medical attention are misdiagnosed or not reported. The majority of fever studies suffers from selection bias and often relies on suboptimal diagnostic tools.
Reports from Africa, the Middle East and most recently South America, suggest that scrub typhus is more widespread than previously appreciated. The molecular detection of Orientia spp. in rodents from Southern France and Senegal suggest that rodent-mite cycles could maintain the pathogen in nature but whether these Orientia spp. represent human pathogens is unknown [69]. The countries most affected by scrub typhus are currently experiencing profound demographic, economic and ecological changes [70]. Deforestation, growing cities and climate change may lead to migration of rodents carrying infected mites and expand to more urban and non-endemic areas [8, 11, 16]. Recently the impact of an earthquake on exposing the population to the possibly perturbed soil dwelling vectors causing scrub typhus was highlighted in Nepal [71].
Ancestor et al. mapped non-malarial causes of fever, including scrub typhus, in the Mekong region [2]. Kelly et al. developed a vector map of scrub typhus based on literature review to include probable and confirmed cases that included geo-referenced locations [72]. These are useful resources that can be built upon to estimate incidence in areas where data is limited. In scrub typhus the extracted information of studies from the 1940s requires careful consideration to identify what data are clinically relevant today. Derne et al. summarized and mapped the distribution of rickettsia and their vectors in Oceania, confirming the widespread presence and providing a scaffold to build upon [73]. Ideally, concerted efforts in providing well maintained up-to-date mapping of human cases and vector (chigger mite) distribution would contribute substantially to understanding the burden of disease.
Burden of disease studies often use syndromic ‘envelopes’ for certain conditions (for example “diarrhea” or “fever”). Developing a fever ‘envelope’ approach for estimating its burden of disease, in conjunction with detailed fever etiology studies would provide improved, standardized and globally comparable incidence data [74, 75]. The resulting data could be stratified further and would inform on the actual burden of disease, as well as provide valuable baseline data to support economic evaluations and mathematical modeling of future interventions [76]. For example, an incentive for identifying endemic areas of scrub typhus may result in increasing cost-effectiveness of rapid diagnostic test (RDT) use. Testing for frequent bacterial pathogens is likely to be economical, reducing hospitalization rates, and informs not only treatment requirements, but also appropriate antibiotic usage [77].
In the case of dengue, the quality of data available has improved substantially and in 2010 there were an estimated 96 million apparent and 294 million unapparent dengue infections globally [78]. Although dengue and scrub typhus both top the list of fever etiologies in multiple studies in Asia, the more easily-treatable disease is neglected–it is time for more integrated expert collaborative research to provide these urgently needed objective data [57, 78, 79].
These data–despite their limitations–make a case for scrub typhus as an important neglected tropical disease of mainly rural populations, with an increasing urban proportion. In countries with established surveillance systems, the reported incidence is increasing and robust documentation of scrub typhus in Chile suggests a much wider global presence than previously understood. The lack of data on global incidence and disease burden highlights the need for this treatable infection to receive increased attention and research to inform health policy.
|
10.1371/journal.pcbi.1005714 | A computational model of spatio-temporal cardiac intracellular calcium handling with realistic structure and spatial flux distribution from sarcoplasmic reticulum and t-tubule reconstructions | Intracellular calcium cycling is a vital component of cardiac excitation-contraction coupling. The key structures responsible for controlling calcium dynamics are the cell membrane (comprising the surface sarcolemma and transverse-tubules), the intracellular calcium store (the sarcoplasmic reticulum), and the co-localisation of these two structures to form dyads within which calcium-induced-calcium-release occurs. The organisation of these structures tightly controls intracellular calcium dynamics. In this study, we present a computational model of intracellular calcium cycling in three-dimensions (3-D), which incorporates high resolution reconstructions of these key regulatory structures, attained through imaging of tissue taken from the sheep left ventricle using serial block face scanning electron microscopy. An approach was developed to model the sarcoplasmic reticulum structure at the whole-cell scale, by reducing its full 3-D structure to a 3-D network of one-dimensional strands. The model reproduces intracellular calcium dynamics during control pacing and reveals the high-resolution 3-D spatial structure of calcium gradients and intracellular fluxes in both the cytoplasm and sarcoplasmic reticulum. We also demonstrated the capability of the model to reproduce potentially pro-arrhythmic dynamics under perturbed conditions, pertaining to calcium-transient alternans and spontaneous release events. Comparison with idealised cell models emphasised the importance of structure in determining calcium gradients and controlling the spatial dynamics associated with calcium-transient alternans, wherein the probabilistic nature of dyad activation and recruitment was constrained. The model was further used to highlight the criticality in calcium spark propagation in relation to inter-dyad distances. The model presented provides a powerful tool for future investigation of structure-function relationships underlying physiological and pathophysiological intracellular calcium handling phenomena at the whole-cell. The approach allows for the first time direct integration of high-resolution images of 3-D intracellular structures with models of calcium cycling, presenting the possibility to directly assess the functional impact of structural remodelling at the cellular scale.
| The organisation of the membrane and sub-cellular structures of cells in the heart closely controls the coupling between its electrical and mechanical function. Computational models of the cellular calcium handling system, which is responsible for this electro-mechanical coupling, have been developed in recent years to study underlying structure-function relationships. Previous models have been largely idealised in structure; we present a new model which incorporates experimental data describing the high-resolution organisation of the primary structures involved in calcium dynamics. Significantly, the structure of the intracellular calcium store is modelled for the first time. The model is shown to reproduce calcium dynamics in control cells in both normal and abnormal conditions, demonstrating its suitability for future investigation of structure-function relationships. Thus, the model presented provides a powerful tool for the direct integration of experimentally acquired structural data in healthy and diseased cells and assessment of the role of structure in regulating normal and abnormal calcium dynamics.
| The cardiac intracellular calcium (Ca2+) handling system is responsible for the control of cellular and organ contraction associated with the heartbeat [1]. Malfunction of this system can directly affect the ability of the heart to work effectively as a pump–reducing cardiac output and potentially leading to mortality. Moreover, abnormal Ca2+ handling dynamics has been increasingly linked to the development of arrhythmogenic triggers in the myocardium [2–4], through two-way coupling between the electrical and Ca2+ handling systems.
Ca2+ handling in cardiac myocytes is regulated by multiple ion channels, pumps and transporters. During the cellular electrical action potential (AP) associated with excitation, an influx of Ca2+ through opening of the voltage-gated L-type Ca2+ channels (LTCCs—carrying flux JCaL) triggers a significant release of Ca2+ from the intracellular Ca2+ store (the sarcoplasmic reticulum, SR) through opening of the Ryanodine Receptors (RyRs—carrying flux Jrel). This process is referred to as Ca2+-induced- Ca2+-release (CICR) [5]. The Ca2+ released from the SR binds with the contractile myofilaments in the bulk intracellular space—the cytoplasm—facilitating cellular contraction. During relaxation, the Ca2+ concentration in the SR is restored from the cytoplasm through the flux Jup, carried by the SERCA Ca2+ pump; Ca2+ is removed from the cell primarily by the sodium- Ca2+-exchanger (NCX–carrying flux JNaCa) and also through the membrane Ca2+ pump (PMCA—carrying flux JpCa). This completes the cardiac cellular Ca2+ cycle associated with electrical activation and contraction.
The LTCCs are found in clusters along the surface membrane and transverse-tubules (TTs)—invaginations of the sarcolemmal membrane responsible for delivering the AP into the interior of the cell. PMCA and NCX are distributed on the surface sarcolemmal membrane and the TTs. On the SR membrane there are clusters of RyRs [6] and continuously distributed SERCA proteins. Critical for CICR is the spatial arrangement of the TT network and the junctional portion of the SR (jSR) to form a microdomain, termed the dyad, to co-localise LTCCs and RyRs on the two membranes. There are thousands of dyads within a cell, and the bulk of the SR forms a network like structure (nSR) connecting the distributed jSRs.
Employing serial block face scanning electron microscopy (SBF-SEM) we have previously determined the 3-D organisation of the TTs and SR in a large animal model, the sheep, revealing details of the network organisation and jSR morphology and importantly the relationship with the TT network to form dyads [7]. Multiple groups have demonstrated that the TTs and related structures are remodelled in disease conditions [7–14] and we have additionally shown that the SR structure is also perturbed in heart failure [7]. These observations highlight the potential importance of structure-function relationships at the sub-cellular scale in Ca2+ dynamics associated with cardiac arrhythmia and perturbed contraction; the precise nature and impact of these relationships requires further investigation.
Computational modelling is a complementary approach to experimental research, and has been successfully applied to provide insight into numerous cardiac phenomena, such as pacemaker activity (e.g., [15]) and the functional impact of pathophysiological ion channel remodelling (e.g., [16]). Recently, computational models which explicitly account for spatio-temporal Ca2+ dynamics have been developed and successfully reproduced phenomena including Ca2+ transient alternans and spontaneous Ca2+ release [17–41]. Such models are in general idealised (with an idealised structure and dyad distribution), but some have used imaging data to distribute RyRs throughout the cell or cell-portion [17,18,22,39]; Other models have been created of small regions of the cell (e.g. the volume surrounding a single TT or even a single dyad) and have integrated more detailed imaging data [29,31,32,37]. However, a spatio-temporal Ca2+ handling model which accounts for realistic SR structure, TT structure and dyad distribution at the whole-cell scale has not yet been developed and involves significant challenges. Nevertheless, the potential advantages offered by such a model–providing a method to directly study variability in structure-function relationships–remain an attractive prospect.
Therefore, the aim of this study was to develop an approach to overcome the challenges of modelling spatio-temporal Ca2+ dynamics using the experimentally reconstructed 3-D structures for the TT and SR at the whole-cell scale. We demonstrate that the approaches developed are sufficient to capture spatio-temporal calcium dynamics with a realistic network SR structure and membrane fluxes distributed according to the sarcolemma/TTs. We additionally report how this model can be used to reproduce Ca2+ transient alternans and spontaneous release events during perturbed rapid pacing as a demonstration of its suitability for future research, and provide preliminary analysis of the importance of structure-function relationships underlying cardiac cellular dynamics. This model (provided in S1 Code) therefore provides a powerful tool to understand structure-function relationships in physiological and pathophysiological cardiac electro-mechanics.
The development of a computational model to achieve the goal of simulating spatio-temporal Ca2+ dynamics with realistic structure and fluxes requires multiple steps: first, the selected image-based dataset describing structure of the TTs and nSR (Fig 1) must be processed to (a) identify and segment the dyads and (b) construct geometry grids suitable for numerical simulation; second, an idealised model for spatio-temporal Ca2+ dynamics must be updated to include these new structures. A major feature of this approach is processing of the SR reconstruction for simulation.
Details of image acquisition using SBF-SEM (FEI Quanta 250 FEG SEM equipped with a Gatan 3View ultramicrotome) and methods for the 3-D reconstruction of the TT and SR structure have been described fully elsewhere [7]. In brief, tissue was extracted from the sheep left ventricle and immediately fixed and prepared for SBF-SEM [42]. The voxel size of the acquired data was 13.5 nm /pixel in the x-y plane and 50 nm in the z-direction. The region within a cardiac myocyte for which the dyad positional data was generated for this study is indicated in Fig 1A. The staining technique employed allowed clear delineation of the TT and SR features (Fig 1A–1C). Segmentation of the SR subsequently allowed the network SR (nSR) to be distinguished from the jSR ‘patches’. 3-D reconstruction of both the TTs and SR morphology ([43]) within the defined region enabled identification of jSR along TTs forming putative dyads (Fig 1D–1F). The position of each dyad was marked with a sphere (Fig 1) to build up a 3-D distribution grid through 357 consecutive slices.
The resolution at which the images were acquired, 13.5nm ×13.5nm × 50nm, is impractical for computational modelling at the whole-cell scale. Computational grids created for numerical simulation were down-sampled to a resolution of 350nm × 350nm × 350nm following processing at the full resolution.
First, the reconstruction of the SR at full resolution was smoothed and cleaned (Fig 3A). This was the baseline geometry to which the down-sampled geometry was mapped. The primary challenge in modelling SR structure is due to its thin cross section: resolutions necessary to capture detailed SR structure are impractical for whole-cell simulations.
To overcome this challenge, the 3-D network structure of the SR, rather than its full 3-D cross-sectional structure, was identified as the key feature to be captured in the model. Under this reduction, the SR can be approximated by a 3-D network of 1-D strands (Fig 3C). The resolution of this “1-D-strand model” was down-sampled to 350nm × 350nm × 350nm, and an algorithm applied to preserve each SR voxel’s nearest-neighbours (Fig 3C).
The volume of each node in the 1-D strand model corresponds to the total volume of the full resolution reconstructed SR divided by the number of nodes in the 1-D strand map, and is thus not defined by the volume of a single voxel at the discretised resolution (vvox.nsr = 0.00305 μm3). For visualization of small portions of the cell model, the concentration distribution in the 1-D strand model was mapped back onto the full resolution reconstruction; this was impractical for visualization of the whole cell due to the computational memory demands of rendering such a large, high resolution structure.
In this section, a general spatio-temporal Ca2+ handling model is described in the context of an idealised cell structure. Then, approaches to incorporating the reconstructed SR and membrane structures are discussed. Full model equations can be found in S1 Text; in this section, only the fundamental equations are given.
In order to perform preliminary analysis to assess the role of intracellular structure affecting spatio-temporal Ca2+ dynamics, alternative geometries were considered for comparison to the fully detailed structural model: (1) Semi-idealised structure: The cytoplasm geometry was used to describe all of the spatially diffuse spaces and fluxes. Dyads were distributed evenly throughout the volume to match the mean inter-dyad distances in the structurally detailed model, resulting in comparable dyad densities (Ndyads for cross section = 2182 vs 2208 for the full and semi-idealised models, respectively). (2) Altered dyad densities: The density of the dyad distribution (and thus inter-dyad distances) in the fully detailed structural model was altered, by either including additional dyads (at junctions of the SR and TT) or by removing dyads. (3) Altered SR diffusion properties: The diffusion coefficient in the SR was varied (by a factor of half and a factor of two). Furthermore, the connectivity of the SR neighbour map was perturbed by removing some neighbours for a randomly selected set of the SR voxels.
The developed cell model reproduces properties of whole cell electrical and Ca2+ handling dynamics (Fig 6A) during control pacing (basic cycle length, BCL, of 1250ms). The AP duration to 90% repolarisation (APD90) at this cycle length is 345ms and the Ca2+ transient has an upstroke time of 20ms, magnitude of 0.68μM and duration to 90% of peak of 400ms. These properties fall within expected ranges for large mammal ventricular myocytes (e.g. for the sheep–the animal model from which the structural datasets were attained [50,51]). The temporal evolution of force follows the Ca2+ transient (Fig 6Ae) which in turn follows the AP upstroke. Model dynamics are stable over long simulation duration times (Fig 6B) once steady-state is achieved. The model successfully reproduces rate dependence of the diastolic and systolic Ca2+ concentrations, exhibiting an elevation of both as pacing rate increases (S2 Fig).
Spatio-temporal dynamics in the cytoplasmic Ca2+ concentration associated with a single beat are shown in Fig 7 and S1 Video. The model captures the dynamics of a single Ca2+ spark (Fig 7A), illustrating the rapid and non-linear decay of the Ca2+ transient as a function of distance from the centre of the dyad [17,28]: the peak of the Ca2+ concentration at a distance of 1.2 μm from the centre of the dyad was an order of magnitude smaller than at the dyad centre.
A linescan along the longitudinal axis of the cell demonstrates the spatial variation in the rise and decay of the Ca2+ transient (Fig 7B): temporal variation of the initiation of the upstroke of the transient results in significant spatial gradients during the initial phase of excitation; significantly smoother spatial gradients were observed during the decay phase.
Snapshots of spatial Ca2+ concentration in 3-D (Fig 7C) and with enhanced scaling (Fig 8A) reveal the 3-D structure of the Ca2+ gradients associated with normal excitation. Properties of Ca2+ gradients during the different phases of the transient are determined by intracellular structure and the nature of the various fluxes controlling Ca2+ dynamics.
The discrete and non-uniformly distributed dyads, heterogeneity in the dyadic cleft volume and small protein numbers (which enhances stochastic state transitions of the RyRs and LTCCs), combined with the rapid transient morphology of Jrel during excitation led to a rapid upstroke of the Ca2+ transient and significant spatial heterogeneity during this initial excitation phase (Fig 7Ci; Fig 8Ai). Jrel quickly terminates at around the time of the transient peak (S4 Fig)—the Ca2+ influx throughout the cell is significantly reduced and Ca2+ diffuses away from the dyads through the bulk cytoplasm, reducing the peaks surrounding the dyads and smoothing the gradients. Due to significant Ca2+ buffering, diffusion is slow and gradients are not eradicated (Fig 7Ciii; Fig 8Aii,iii). The decay of the Ca2+ transient is slower and more spatially uniform due to the spatially and temporally continuous nature of the effluxes, JMem and JSR, (Fig 7Civ; Fig 8Aiv). Subcellular heterogeneity was also observed during controlled AP clamp conditions.
Comparison to the semi-idealised cell model (which contains a uniform dyad distribution—See Methods: Semi-idealised and perturbed structure models) reveals that the complex 3-D structure of Ca2+ gradients is largely determined by the dyad distribution (compare Fig 8A with 8B). Temporal variation in the activation time of individual dyads results in significant gradients during the first few miliseconds of excitation in the semi-idealised model, but then much more uniform Ca2+ concentration was observed during the main excitation phase, compared to the fully detailed structural cell model (Fig 8). Inclusion of realistic flux distribution in the semi-idealised model enhances spatial heterogeneity but has a much smaller impact than the dyad distribution.
Spatial Ca2+ gradients in the nSR were less pronounced than in the cytoplasm (Fig 9; S1 Video). Nevertheless, gradients were observed during the initial excitation (emptying) phase, due to the distributed dyads/jSRs (Fig 9i and 9ii). The spatial distribution is almost uniform at the time of maximum depletion of the SR (Fig 9iii). During the refilling phase (Fig 9iv), gradients were observed but are more uniform than those during the initial phase (compare Fig 9 panel iii with v), comparable to the behaviour observed in the bulk cytoplasm.
Fluxes acting on the cytoplasm in a single portion of the 3-D cell were analysed (Fig 10). The spatial distribution of Jup (Fig 10A) and JNaCa (Fig 10B) can be clearly seen, and correspond to the nSR and surface sarcolemma/TT structures, respectively. Spatial gradients in the magnitude of both of these fluxes are observed, as a direct result of the gradients in intracellular Ca2+ concentration (note the spatial correspondence between the gradients in Figs 7, 8 and 10).
The effect of voltage inhibition on the activity of JNaCa combined with the initial large peaks of Ca2+ in locations close to the dyads can be clearly seen by spatially distributed peaks in JNaCa in the initial phase of excitation (Fig 10Bi), uniform and small fluxes during the bulk of the excitation phase, wherein the Ca2+ concentration is large (Fig 10Bii), and a more spatially uniform but larger flux during the decay phase of the Ca2+ transient, corresponding to the peak of INaCa (Fig 10Biii).
Varying the SR diffusion parameters primarily affected the spatial-distribution of Ca2+ in the SR: Rapid diffusion in the SR facilitated equilibration of the SR Ca2+ content and reduced gradients; slower diffusion enhanced Ca2+ gradients (Fig 17A). However, the impact on whole-cell dynamics was relatively minimal under the variance in diffusion coefficient of 0.5–2 times the baseline value (0.3 μm/ms) considered in this study.
Reduced connectivity in the SR network led to significant SR Ca2+ gradients during normal excitation, with islands of SR loading appearing during the systolic phase (Fig 17B). This localised early SR loading promoted secondary spontaneous release during the AP–note the locations of secondary release correspond to the islands of SR loading.
In this study, a model of 3-D spatio-temporal Ca2+ dynamics in a large mammal ventricular myocyte was developed which incorporates significant details of intracellular Ca2+ handling structures. This is the first to employ 3-D reconstructions of the TT network and SR for cardiac myocytes in situ within tissue blocks, derived from SBF-SEM (Fig 1). We describe how the structures were processed to form computational grids and mapping functions and incorporated into a 3-D spatio-temporal Ca2+ handling mathematical model, accounting for realistic cytoplasm and nSR structure and fluxes distributed according to the imaging data (Figs 2–5). Our simulations illustrate that the model accurately reproduces excitation characteristics in normal conditions (Figs 6–9) as well as perturbed conditions leading to Ca2+ transient alternans and spontaneous release events (Figs 11–14), and reveals high resolution 3-D Ca2+ gradients and fluxes associated with excitation (Figs 8–10). Furthermore, we provide preliminary analysis which demonstrates the importance of intracellular structures in determining the spatial distribution of dyad recruitment during Ca2+ transient alternans (Fig 13) and Ca2+ wave propagation (Fig 15) as well as the role of SR connectivity in maintaining stable Ca2+ dynamics (Fig 17).
There are numerous exemplary studies in recent years implementing spatio-temporal Ca2+ handling models in multiple dimensions (e.g. [17–41]). These models have been used to mechanistically evaluate physiological and pathophysiological dynamics in the intracellular Ca2+ handling system, such as graded release [20], RyR dynamics [23,40], Ca2+ transient alternans [20,21,25,27,30], Ca2+ waves [19,20,41] and pacemaker activity [38], and account for varying degrees of detail of intracellular structure (realistic RyR distribution; reconstruction of single TT; super-resolution of single dyad). However, no model has yet accounted for the realistic structure of the nSR, nor integrated fluxes according to realistic membrane structure, dyad distribution and nSR structure at the whole cell scale.
The model developed in the present study achieves this goal, providing for the first time a framework to directly integrate imaging data on multiple structures with whole-cell modelling. A major feature of the model is the ability to account for the structure of the network SR at the whole-cell scale. Identifying the network-like structure of the model as the key feature to be captured provides a method to model Ca2+ diffusion throughout the nSR at lower resolutions than required to image this structure, through the reduction to a 3-D network of 1-D strands (Fig 3).
Whereas idealised geometry based models offer ease of data interpretation and are suitable for general analysis of Ca2+ dynamics, the modelling approaches presented in this study provide a method to remove much of the uncertainty inherent to models based on idealised geometries and to directly asses how structure-function relationships are affected by variability in intracellular structure, which may be particularly relevant when considering structural remodelling associated with disease [7]. The present approach provides the first framework which allows multiple structural datasets to be directly integrated with mathematical modelling of Ca2+ dynamics without the assumptions required by idealised models to capture structural variability, which accentuate the inherent uncertainty of these models. This, therefore, potentially significantly increases the confidence of simulation data regarding variability in intracellular structure.
On the other hand, models of specific regions of the cell which incorporate realistic structure can offer significant understanding of local control of EC coupling, but cannot capture whole-cell emergent properties such as the interaction of heterogeneity of these structures throughout the cell. The model presented here therefore complements those previously developed, providing a framework to investigate whole-cell dynamics underlain by real structure and heterogeneity.
We note that, in support of the validation of the model, results in the present study are consistent with those of previous modelling studies where comparable. For example, Izu et al. 2006 [17] found similar results regarding the criticality of inter-dyad distances in maintaining the propagation of Ca2+ waves. Ca2+ spark hierarchy is similar to that shown in Nivala et al. 2012 [19]. The mechanism underlying Ca2+ transient alternans is consistent with the studies of Restrepo et al. 2008 [20], Rovetti et al. 2010 [21] and Alvarez-Lacalle et al. 2015 [25].
The suitability of the model for future research was demonstrated by its application under multiple conditions. The model reproduces whole-cell and spatio-temporal Ca2+ characteristics under control pacing (Figs 7–9), rapid perturbed pacing leading to alternans (Figs 11–13), and rapid pacing leading to SR overload, spontaneous Ca2+ release and the development of single-cell triggered activity (Figs 14–16).
The primary focus of the study was to develop an approach to overcome the challenges inherent in the construction of such a detailed model and to demonstrate the suitability of the model for future research, which was achieved through application of the model in normal and arrhythmic excitation conditions. The intention of the model is for future studies to incorporate multiple datasets, including those describing multiple disease conditions, to fully assess the role of variability in intracellular structure in determining potentially pro-arrhythmic dynamics; such detailed analysis was therefore beyond the scope of this study.
This notwithstanding, the present study also provides for the first time detailed and high resolution (spatial, temporal and concentration) reconstruction of Ca2+ gradients and fluxes in 3-D in both the cytoplasm and network SR (see Results sections: Spatio-temporal Ca2+ dynamics in the bulk cytoplasm; Spatio-temporal Ca2+ dynamics in the network SR; Evaluation of spatial distribution of fluxes in the cytoplasm), as well as preliminary analysis of the importance of structure underlying spatio-temporal Ca2+ dynamics and the role of SR diffusion and connectivity (see Results sections: Intracellular Ca2+ transient alternans; Ca2+ spark hierarchy and spontaneous release events; SR Diffusion Properties).
Comparison with the semi-idealised model revealed similarities and differences between the two levels of detail included in the model. The similarities of whole-cell characteristics and gross spatio-temporal dynamics between the two models during normal structure highlights the suitability of idealised models for general and mechanistic modelling of spatio-temporal dynamics associated with normal structure.
However, multiple results presented here also emphasise the important role of structure and heterogeneity underlying the complex and fine-scale details of 3-D intracellular Ca2+ dynamics, which may be particularly important for disease models of intracellular structural remodelling.
First, the complex structure of 3-D Ca2+ gradients and intracellular Ca2+ fluxes arises primarily as result of dyad distribution and intracellular structure; this complex structure does not emerge in the idealised cell model. Moreover, spatially heterogeneous distributions of the membrane and SR Ca2+ fluxes were observed despite homogeneous distribution of the maximal flux rates, as a direct result of these intracellular Ca2+ gradients.
Second, the degree of disorder associated with alternans dynamics was significantly reduced in the structurally detailed model compared to the idealised model; dyad distribution significantly affects recruitment patterns and therefore spatially constrains the probabilistic nature of recruitment, with real structure reducing the phase variance associated with multiple small-amplitude cycles. Similarly, reduced density of dyad distribution led to significant failure to recruit and very small transients associated with the small amplitude cycle. These results thus add further insight to those gained in previous studies [20,25,27].
Third, the critical inter-dyad distances to maintain Ca2+ propagation are within the distribution of dyad nearest neighbour distances; whereas the overall trend of Ca2+ spark hierarchy is preserved under different dyad distributions, the propagation patterns emerging and dependence on SR Ca2+ concentration were significantly affected by dyad distribution, with whole-cell coordinated single waves requiring relatively short inter-dyad distances.
Finally, the connectivity of the network SR was vital in maintaining normal Ca2+ dynamics; reduced connectivity led to failure of the SR to equilibrate, significant and heterogeneous SR loading and secondary systolic Ca2+ release.
These results highlight the necessity for integrated multi-scale models which can capture realistic structure of both the intracellular and SR spaces as well as heterogeneity in dyad properties and flux distribution. For example, a combination of increased heterogeneity in dyad properties with remodelling of dyad distribution may result in highly unpredictable constraints on dynamics. TT structure, dyad distribution and SR structure have all been shown to be perturbed in both animal models of disease and patients [7–14]. Whereas many previous modelling studies have investigated remodelling in proteins and flux dynamics in the single cell (e.g., [16]), intracellular structural remodelling has only been investigated by a few computational studies and these have been idealised (e.g., [52]). Application and analysis of structurally accurate models such as the one presented in this study will therefore allow mechanistic evaluation of the role of structural remodelling in determining arrhythmogenic electrical and perturbed contractile dynamics at the cellular scale as well as structure-function relationships underlying normal cardiac excitation.
Whereas the present study includes a novel mathematical model describing intracellular Ca2+ dynamics, the primary focus was on the methods to process and integrate structural datasets into modelling frameworks. To demonstrate the generalisability of these approaches, the structural model was integrated with the independent mathematical model of Nivala et al. 2012 [27]. Note that this model contains a different schematic structure (with no subspace) and exhibits a larger Ca2+ transient associated with excitation.
The whole-cell characteristics of the integrated model were comparable to that of the Nivala et al. study, whereas the structure of intracellular Ca2+ gradients is comparable to those previously shown in the present study; Ca2+ diffusion is aided in the Nivala et al. version of the model as a result of the larger transient, and thus the quantitative comparison of the gradients reveals some differences–nevertheless, the main structure remains (S5 Fig). This therefore demonstrates the generalisability of the approaches for integration with independent mathematical models and further supports the suitability for future research.
For the purpose of the methodological development of the model, only a portion of the myocyte was reconstructed at high resolution and processed to form the computational grids and mapping functions used for simulation. For specific studies of variability in Ca2+ dynamics, especially those in disease where intracellular structure may be highly irregular, a larger or full-cell reconstruction would be necessary. However, it should be noted that this does not affect the fundamental model itself nor the methodological approaches for structural data processing.
In the model, the membrane and SR fluxes were spatially distributed according to maps created from the reconstructions. Within this structure, the fluxes were distributed evenly and continuously. Incorporation of immuno-labelling data describing the realistic distribution of the proteins responsible for these fluxes would provide a further degree of accuracy in the cell model as well as a tool for investigation of the functional impact of changes to the distribution of these flux-carrying proteins; arbitrary heterogeneity could have been introduced into the present model but this was avoided due to the inherent uncertainty in such an approach–the purpose of this model being to remove this necessity for future research. In further simulations, we also demonstrate that redistribution of INaCa primarily to the TTs, as has been performed in other studies [24], results in preferential flux in the TTs without significant perturbation to whole-cell dynamics (S6 Fig).
The model includes a restricted buffering subspace which functionally couples neighbouring dyads. The inclusion of this subspace was primarily motivated by considerations for reproducing physiological Ca2+ dynamics. Whereas the presence of pathways between the localised Ca2+ buffers may be physiologically relevant, it is not the intention of this study to make a comment and such functionality requires further investigation experimentally. We, however, note that similar constructs are present in previously developed models from other groups: in both Gaur-Rudy 2011 [28] and Voigt et al. 2014 [26], neighbouring dyadic spaces are directly coupled. The present model contains this functional coupling but also preserves the restricted local spaces of individual dyads. This construct is not critical to the primary novelty of the developed approach–in processing and modelling with the structural datasets–as demonstrated by integration with the independent mathematical model of Nivala et al., which does not contain this subspace (see Discussion: Generalisation of the model).
The dyads/jSRs were treated as point sources (single voxels) in the spatial geometries. Whereas heterogeneity was incorporated through functionality to vary dyadic cleft volume and protein numbers (NRyR and NLTCC), the spatial structure of the dyad, including local RyR distribution, is not accounted for in the present model. A multi-scale integration method could feasibly be implemented to integrate previously developed super-resolution models of single dyads [37,53], though this was beyond the scope of the present study. Similarly, the RyR dynamic model is primarily functional rather than rigorously based on experimental data describing RyR kinetics. Incorporation of a more accurate RyR model, such as the recent induction decay models [40] would also improve the accuracy and suitability for future studies. Furthermore, heterogeneity in the jSR volume and calsequestrin concentration could be included to further analyse the role of these heterogeneities underlying spatial Ca2+ dynamics.
Further possible extensions to the model include segmentation of the mitochondria [39] (the spatial distribution of the mitochondria will affect spatial Ca2+ diffusion because they effectively act as barrier to diffusion) and subsequent incorporation of localised energetics. Furthermore, the contractile system could also be segmented for simulation of spatially localised troponin buffering and force generation, for application to understand reduced contractile force associated with heart failure, for example.
A computational model of 3-D spatio-temporal Ca2+ dynamics has been created which incorporates realistic reconstructions of multiple intracellular structures, namely the network SR structure, cytoplasm volume, and fluxes distributed according to the membrane/TT and SR structures. Understanding the role of intracellular Ca2+ cycling in physiological and pathophysiological cellular dynamics is vital for mechanistic evaluation of perturbed contraction and arrhythmia associated with Ca2+ handling malfunction, and may contribute to improved treatment strategies for prevention, management and termination of life-threatening conditions. The methodological framework and model reported here provide a powerful tool for future investigation of structure-function relationships at the whole cell scale underlying physiological and pathophysiological intracellular Ca2+ handling, beyond the insight gained in this study. Full model code is provided (S1 Code) to facilitate realisation of the potential of future applications of these approaches.
|
10.1371/journal.pgen.1005514 | Basolateral Endocytic Recycling Requires RAB-10 and AMPH-1 Mediated Recruitment of RAB-5 GAP TBC-2 to Endosomes | The small GTPase RAB-5/Rab5 is a master regulator of the early endosome, required for a myriad of coordinated activities, including the degradation and recycling of internalized cargo. Here we focused on the recycling function of the early endosome and the regulation of RAB-5 by GAP protein TBC-2 in the basolateral C. elegans intestine. We demonstrate that downstream basolateral recycling regulators, GTPase RAB-10/Rab10 and BAR domain protein AMPH-1/Amphiphysin, bind to TBC-2 and help to recruit it to endosomes. In the absence of RAB-10 or AMPH-1 binding to TBC-2, RAB-5 membrane association is abnormally high and recycling cargo is trapped in early endosomes. Furthermore, the loss of TBC-2 or AMPH-1 leads to abnormally high spatial overlap of RAB-5 and RAB-10. Taken together our results indicate that RAB-10 and AMPH-1 mediated down-regulation of RAB-5 is an important step in recycling, required for cargo exit from early endosomes and regulation of early endosome–recycling endosome interactions.
| When cargo is internalized from the cell surface by endocytosis, it enters a series of intracellular organelles called endosomes. Endosomes sort cargo, such that some cargos are sent to the lysosome for degradation, while others are recycled to the plasma membrane. Small GTPase proteins of the Rabs family are master regulators of endosomes, functioning by acting as molecular switches. As cargo moves through the endosomal system, it must pass from the domain controlled by one Rab-GTPase to the domain controlled by another. Little is known about how transitions along the recycling pathway are controlled. Here we analyze a group of protein interactions that act along the early-to-recycling pathway. Our work shows that RAB-5 deactivation mediated by TBC-2 and its recruiters RAB-10 and AMPH-1 is important for cargo recycling. This work provides mechanistic insight into how Rab proteins controlling different steps of trafficking interact during endocytic recycling.
| Endocytic recycling, the return of proteins and lipids from endosomes to the plasma membrane, plays a key role in many essential cellular processes including nutrient uptake, cell migration, cytokinesis, synaptic plasticity, immune response, and growth factor receptor modulation [1]. In polarized epithelial cells an additional layer of complexity in the endocytic pathway contributes to formation and/or maintenance of the specialized apical and basolateral domains [2,3]. Both the apical and basolateral membranes deliver cargo to early endosomes, often referred to as apical early endosomes and basolateral early endosomes [3–5]. Basolaterally derived and apically derived cargo can reach common recycling endosomes, from which cargo is sorted for delivery to the basolateral plasma membrane or to apical recycling endosomes [3–5]. The apical recycling endosomes are thought to send their cargo to the apical plasma membrane. Small GTPases of the Rab superfamily play key roles in membrane transport, with at least one Rab protein regulating each transport step. In polarized epithelial cells Rab11 is primarily associated with the apical recycling endosomes and is thought to function in the transport of cargo from the apical recycling endosomes to the plasma membrane [3,6,7]. Rab8 has also been implicated in apical recycling in the intestinal epithelia of mice and worms [8].
Our attention was first brought to bear on the basolateral recycling pathway of C. elegans intestinal epithelia because of the accumulation of grossly enlarged basolateral vesicles in mutants lacking the recycling regulator RME-1/EHD [9]. In the case of rme-1 mutants, these enlarged vesicles accumulated recycling cargo and were positive for the endosomal recycling regulator ARF-6, but lacked early endosome marker RAB-5, suggesting that RME-1 functions at a late recycling step [9–11]. Pulse-chase data in mammalian cells showed that loss of mRme-1/EHD1 likewise resulted in a block in recycling endosome to plasma membrane transport [12,13]. Similarly rab-10 mutants first caught our attention because they displayed enlarged basolateral vesicles in the C. elegans intestine that accumulated recycling cargo [10]. However, in this case the enlarged endosomes were positive for RAB-5, indicating an earlier block in basolateral recycling, at the level of early endosome to recycling endosome transport [10]. We extended this work, identifying two RAB-10 effectors that function with RAB-10 in basolateral recycling, EHBP-1 and CNT-1 [11,14]. EHBP-1 strongly labeled the tubular elements of the recycling pathway, was required for strong RAB-10 endosomal recruitment, and may link endosomes to the cytoskeleton [14,15]. CNT-1/ACAP is recruited to endosomes by RAB-10 and regulates the activity of ARF-6, acting as part of a small GTPase regulatory loop [11]. In turn ARF-6 regulates PI5-kinase, controlling PI(4,5)P2 levels on basolateral recycling endosomes, and the recruitment of downstream PI(4,5)P2 lipid binding proteins such as RME-1 [11,16].
C. elegans RAB-10 and human Rab10 are now known to contribute a wide range of endocytic recycling pathways. Like its C. elegans homolog, mammalian Rab10 functions in basolateral recycling in polarized MDCK cells, where Rab10 localized to basolateral sorting endosomes and the common recycling endosome [17]. C. elegans RAB-10 is also required for the postsynaptic recycling of glutamate receptors in interneurons [18], and dense-core vesicle secretion of neuropeptides by motor neurons [19]. Mammalian Rab10 is required for toll-like receptor 4 recycling in activated macrophages [20], membrane insertion of plasmalemmal precursor vesicles during neuronal polarization and axonal growth [21,22], and insulin-stimulated glucose transporter recycling in adipocytes [23]. Expression of human Rab10 in the C. elegans intestine rescues rab-10 mutant defects, indicating a high degree of functional conservation, suggesting that further elucidating RAB-10 function in C. elegans will provide mechanistic insight into RAB-10/Rab10 function in many or all of these related processes [10].
Countercurrent cascades of Rab GEFs and Rab GAPs have been proposed to mediate Rab conversion, a process by which Rab proteins interact, helping to establish vectorial transport of cargo along membrane trafficking pathways [24]. In such cascades early acting Rab-GTPases recruit effectors that activate later acting Rab-GTPases, and in turn later acting Rab-GTPases recruit effectors that inactivate early acting Rab-GTPases [24]. However little is known of how such cascades contribute to endocytic recycling. Here we show that RAB-10 recruits the RAB-5 GTPase-activating-protein TBC-2 to endosomes in a step necessary for early endosome to recycling endosome transport. This negative feedback from RAB-10 to RAB-5 is required for the exit of recycling cargo from early endosomes. We also show that the BAR-domain protein AMPH-1 is a binding partner of TBC-2 important for recruitment of TBC-2 to endosomes, functioning as part of the transition of cargo from the early to recycling endosome compartments.
We have previously reported several proteins that function with RAB-10 in basolateral recycling in the C. elegans intestine, some of which we first identified via a yeast two-hybrid screen that used a predicted constitutively GTP-bound form of RAB-10(Q68L) as bait [14]. In this same yeast two-hybrid screen we also identified a RAB-10(Q68L) interacting clone encoding full-length TBC-2, a GAP for the earlier acting endosomal GTPase RAB-5 [13,22,25]. The interaction between RAB-10(Q68L) and TBC-2 was positive in both Leu2 and β-galactosidase expression assays (Fig 1A). Using successive truncations of TBC-2 we narrowed the RAB-10 binding site to a 42 amino acid region of TBC-2 (amino acids 279–321) (Fig 1A, 1B and 1E). We noted several runs of highly charged residues in this region, which may represent hydrophilic surface features, and tested their importance for binding to RAB-10 in groups of 5 by alanine scanning. The interaction was abolished when alanine substitutions were imposed at TBC-2 positions aa283-287, aa288-292, and aa294-298 (Fig 1C). These mutations could disrupt binding of TBC-2 to RAB-10 by directly removing surface features involved in the binding interface, or could disrupt the local structure of this region of TBC-2 interfering with binding. Taken together, our results indicate the presence of a predicted coiled-coil domain of TBC-2 that interacts with RAB-10, a key regulator of the basolateral endocytic recycling process. Since TBC-2 is known to act as a GAP for early endosome master regulator RAB-5, these results suggest a negative feedback loop from RAB-10 to RAB-5, potentially acting as part of a RAB cascade in the basolateral recycling pathway.
Intestinally expressed GFP-tagged TBC-2 labels abundant cytoplasmic puncta with the typical size and shape of endosomes (~250–500 nm diameter). If TBC-2 is a physiologically relevant binding partner for RAB-10, we would expect to find RAB-10 and TBC-2 on the same population of endosomes in vivo. Previous qualitative work indicated some localization of TBC-2 to early and late endosomes, but the extent of localization, and its relationship to recycling endosomes, remained unclear [22,25]. To quantitatively test the subcellular localization of TBC-2 we conducted a series of co-localization studies in the intestinal epithelial cells where RAB-10 is known to function, using a set of previously established RFP markers for RAB-10 and a variety of endocytic compartments. The degree of colocalization was measured using Pearson’s correlation coefficient, a statistical measure of the degree of linear dependence of the GFP and RFP signals [26]. Consistent with our binding data, we detected the greatest correlation coefficient of GFP-TBC-2 with RFP-tagged RAB-10(+) and constitutively active RFP-RAB-10(Q68L) (Fig 2A–2A‴ and 2B–2B‴ and Fig 2D). The greater degree of correlation of TBC-2 signal with RAB-10(Q68L) signal is consistent with a model where RAB-10 helps to recruit TBC-2 onto endosomes. GFP-TBC-2 signal also correlated very well with a previously characterized RAB-10 effector, CNT-1 (CNT-1-mCherry) (Fig 2C–2C‴ and Fig 2D), which is also required for the recycling process [11]. These results are consistent with TBC-2 acting with RAB-10 and CNT-1 in the basolateral endocytic recycling.
GFP-TBC-2 signals also showed lesser, but significant, correlations with early endosomal marker tagRFP-RAB-5 (S2A–S2A'' and S2D Fig) and late endosomal marker tagRFP-RAB-7 (S2B–S2B'' Fig and S2D Fig). We also noted that the GFP-TBC-2 signal displays hardly any correlation with that of EHBP-1-mCherry, another RAB-10 interacting protein that labels tubular aspects of the basolateral recycling endosome network (S2C–S2C'' Fig). Collectively, our results indicate that TBC-2 is enriched on a subpopulation of endosomes, where it could function with RAB-10 and RAB-5 to confer effective transport of cargo during the endocytic recycling process.
To further test the idea that an interaction with RAB-10 is important for TBC-2 function in vivo, we examined the effect of a rab-10 loss-of-function mutant on the endosomal localization of GFP-TBC-2 in the intestinal epithelia. In the rab-10 mutant background GFP-TBC-2 became very diffusive, losing its typical punctate endosomal localization, indicating a requirement for RAB-10 in TBC-2 endosomal recruitment (Fig 3A and 3B). Western blot analysis also showed that GFP-TBC-2 levels are reduced in rab-10 mutants, suggesting that TBC-2 is less stable in the absence of RAB-10 (Fig 3E). We extended this analysis further, testing a form of TBC-2 impaired for RAB-10 binding (QRNNE 288–292 AAAAA) for function in vivo. In previous work we showed that TBC-2 is required for the normal recycling of model cargo hTfR-GFP (human transferrin receptor–GFP)[27]. In the absence of TBC-2, hTfR-GFP accumulates in enlarged intracellular structures (Fig 4A, 4B and 4F). While expression of full length wild-type TBC-2 efficiently rescued the localization of hTfR-GFP in a tbc-2 null mutant background (Fig 4A–4C and 4F), we found that expression of the interaction defective form of TBC-2 failed to rescue the localization of hTfR-GFP in a tbc-2 null mutant background (Fig 4E and 4F).
In many cases peripheral membrane proteins of the endosome require multiple protein and/or lipid interactions to direct their localization. Recent work using phage-display to identify the binding preferences of all C. elegans SH3 domains suggested a link between TBC-2 and AMPH-1, a BAR-domain and SH3-domain protein that is the only C. elegans member of the Amphiphysin/BIN1 protein family [28,29]. TBC-2 amino acid sequence 146–160 was identified as the fourth best match for the AMPH-1 SH3-domain binding consensus in the entire predicted C. elegans proteome [28]. Previous work from our laboratory has shown that AMPH-1 participates in the basolateral recycling pathway [28,29]. Thus we sought to further examine this potential interaction. We detected interaction of full-length TBC-2 with the AMPH-1 SH3 domain in a yeast 2-hybrid assay (Fig 1D). Importantly, the interaction was abolished when key residues in the consensus sequence, prolines P150 or P153, or arginine R155, were mutated to alanine (Fig 1D and 1E). Despite losing their ability to interact with AMPH-1, the P150A, P153A, and R155A mutant forms of TBC-2 protein retained the ability to interact with RAB-10(Q68L) in the same two-hybrid assay, indicating that the mutant forms of TBC-2 were stable (S1A Fig). We conclude that the AMPH-1 SH3 domain has the potential to bind to the predicted target sequence in TBC-2 (Fig 1D and S1A Fig).
If an interaction between AMPH-1 and TBC-2 is important in vivo, we might expect to observe a change in TBC-2 localization in an amph-1 mutant background. Indeed, when we examined the subcellular localization of intestinally expressed GFP-TBC-2 in an amph-1 deletion mutant, we found that the normal punctate endosomal distribution of GFP-TBC-2 was severely disrupted (Fig 3A, 3C and 3D). Instead, GFP-TBC-2 appeared quite diffusive in the absence of AMPH-1, indicating that AMPH-1 is important for endosomal recruitment of TBC-2 (Fig 3C). GFP-TBC-2 levels as assayed by western blot were not reduced in amph-1 mutants (Fig 3E).
We extended this analysis further, testing a form of TBC-2 impaired for AMPH-1 binding (P150A) for function in vivo, using the same hTfR-GFP localization assay described above. We found that while expression of full length wild-type TBC-2 efficiently rescued the localization of hTfR-GFP in a tbc-2 null mutant background (Fig 4A–4C and 4F), the expression of the interaction defective form of TBC-2 failed to rescue the localization of hTfR-GFP in a tbc-2 null mutant background (Fig 4A, 4B, 4D and 4F). Our results thus indicate that in addition to RAB-10, AMPH-1 also contributes to TBC-2 endosomal recruitment.
We also determined that AMPH-1 can interact with RAB-10 using a GST-pulldown approach with full length GST-AMPH-1 and HA-tagged RAB-10(Q68L) (S3C Fig). Addition of FLAG-TBC-2 to this assay showed that GST-AMPH-1 can pull down TBC-2 and RAB-10 at the same time, but the presence of TBC-2 in the reaction did not appear to increase the pulldown efficiency of RAB-10 (S3C Fig). Colocalization analysis indicated the presence of AMPH-1-GFP and tagRFP-RAB-10 on a significant fraction of the same endosomes, consistent with physiological significance for the AMPH-1/RAB-10 interaction (S3A and S3B Fig). However, loss of RAB-10 did not reduce association of AMPH-1-GFP with membranes (S4A–S4C Fig). Rather in rab-10 mutants we observed an increase in AMPH-1-GFP puncta and tubule intensity (S4A–S4C Fig). This may be an indirect effect of the increase in endosomal PI(4,5)P2 in rab-10 mutants that we previously showed occurs in part via another RAB-10 effector CNT-1, an ARF-6 GAP [11]. Alternatively RAB-10 may affect AMPH-1 recruitment or function more directly, perhaps affecting its conformation or interaction with other proteins.
If RAB-10 and AMPH-1 contribute to TBC-2 recruitment and function, then loss of RAB-10 or AMPH-1 would be expected to result in abnormally elevated levels of GTP-bound RAB-5. Furthermore, since the Rab protein nucleotide cycle is linked to Rab protein membrane association, an elevated "active" GTP-bound status for RAB-5 should result in an elevated level of membrane-bound RAB-5. This model predicts that in tbc-2, rab-10, and amph-1 mutants, where the RAB-5 GAP TBC-2 is either completely missing, or is mislocalized, RAB-5 association with membranes should be increased. Previous work showed that RAB-5 labeled endosomes are enlarged and/or more numerous in tbc-2 and rab-10 mutants, consistent with this model [10,25,27]. In our previous work we had assayed RAB-5 labeled early endosome number in amph-1 mutants and found no significant change [29]. However, in light of the interaction of AMPH-1 with TBC-2, we analyzed additional parameters, and found that RAB-5 puncta intensity is increased in amph-1 mutants, consistent with elevated RAB-5 membrane association (S5 Fig).
Endosome size and number can change for a number of reasons, so we extended this analysis to directly measure RAB-5 membrane association biochemically. We separated membranes from cytosol in C. elegans lysates using ultracentrifugation at 100,000g in the appropriate mutant backgrounds, comparing the amount of intestinally expressed GFP-RAB-5 present in each fraction by Western blot. Consistent with the predictions from this model, we observed an elevation in GFP-RAB-5 membrane-to-cytosol ratio in tbc-2, rab-10, and amph-1 mutants (Fig 5A–5C). Loss of RAB-10 or AMPH-1 increased the membrane association of RAB-5 to a lesser extent than that caused by loss of TBC-2, suggesting that some localized TBC-2 remains in rab-10 and amph-1 mutants, although endosome localized TBC-2 is difficult to visualize by microscopy in such mutant backgrounds (Fig 5A and Fig 5C). In summary, our data support a role for rab-10 and amph-1 in TBC-2 membrane recruitment that is required to complete the RAB-5 nucleotide cycle, removing RAB-5 from membranes. Since RAB-10 and AMPH-1 function in the recycling aspect of endocytic trafficking, these results suggest that removal of RAB-5 from endosomal membranes is an integral part of the recycling process, perhaps linked to cargo transition from early to recycling endosome transport.
Previous work showed that RAB-5 and RAB-10 display significant spatial overlap in the C. elegans intestine, consistent with functional data indicating that RAB-10 is important for exit of recycling cargo from RAB-5-positive endosomes [10]. To better understand the relationship between RAB-5 and RAB-10, we assayed for changes in their relative colocalization in tbc-2 and amph-1 mutants. Similar to previously published results, we found that under wild-type conditions tagRFP-RAB-5 and GFP-RAB-10 both label punctate endosomal structures that partially colocalize (Fig 6A–6A‴ and 6D). We detected dramatic morphological changes for both tagRFP-RAB-5 and GFP-RAB-10 labeled endosomes in a tbc-2 mutant background. Aside from some remaining punctate structures, in tbc-2 mutants tagRFP-RAB-5 and GFP-RAB-10 tended to label very large pleiomorphic structures that were never observed in wild-type animals (Fig 6B–6B‴). Quantification of RAB-5 colocalization with RAB-10 showed a significant increase in the correlation of tagRFP-RAB-5 and GFP-RAB-10 signals in tbc-2 mutants (Fig 6D), with colocalization mostly restricted to the grossly enlarged structures (Fig 6B–6B‴). amph-1 mutants also displayed a significant increase in the correlation of the tagRFP-RAB-5 and GFP-RAB-10 signals (Fig 6C–6C‴ and 6D), although the morphological size and shape changes were less severe than those in tbc-2 mutants (Fig 6C–6C‴). Taken together, these data suggest that TBC-2 and AMPH-1 cause recycling defects by altering the normal compartmentalization of RAB-5 and RAB-10 on endosomes.
Our previous work on RAB-10 function in the intestine showed that RAB-5 labeled endosomes in rab-10 mutants are grossly enlarged and accumulate an additional model recycling cargo, hTAC-GFP (human TAC, IL-2 receptor alpha chain)[10]. hTAC-GFP strongly labels the tubular aspects of the basolateral recycling pathway at steady state, and depends upon RAB-10, RME-1, and ARF-6 for its recycling [10–11,13]. To better understand the step in recycling transport affected by TBC-2 and AMPH-1 we assayed the relative localization of hTAC-GFP to tagRFP-RAB-5 and tagRFP-RAB-10 in tbc-2 and amph-1 mutants. Under wild-type conditions, hTAC-GFP displays little steady-state overlap with tagRFP-RAB-5 (Fig 7A–7A‴). In tbc-2 mutant animals, the tubular meshwork of hTAC-GFP appears disrupted, with hTAC-GFP mostly found in enlarged endosomes, many of which label for tagRFP-RAB-5 (Fig 7B–7B‴). We measured a striking increase in the degree of colocalization between hTAC-GFP and tagRFP-RAB-5 in tbc-2 mutants (Fig 7D). In animals lacking AMPH-1, we also detected a significantly larger degree of overlap between hTAC-GFP and tagRFP-RAB-5 in comparison to that of wild-type animals (Fig 7C–7C‴ and 7D). Consistent with our previous reports, we observed partial overlap of hTAC-GFP with tagRFP-RAB-10, mostly restricted to punctate rather than tubular aspects of the hTAC-GFP labeled endosomes (Fig 8A–8A‴). The degree of colocalization between hTAC-GFP and tagRFP-RAB-10 increased mildly in tbc-2 mutants and was basically unaltered in amph-1 mutants (Fig 8A–8A‴, 8B–8B‴, 8C–8C‴ and 8D). Taking into account the aforementioned increase in colocalization between RAB-5 and RAB-10 in these mutant backgrounds, these data suggest that most hTAC-GFP in tbc-2 mutant and in amph-1 mutant animals is trapped in the early endosome.
Given the continuous flow of proteins and membranes along the endocytic and exocytic pathways, cells face a formidable challenge in achieving accurate intracellular transport of membrane cargo. Such transport is likely to require tight regulation that enforces the directionality of sequential flow between membranous compartments [24]. Rab GTPases serve as master regulators of membrane trafficking by controlling the structural and functional characteristics of intracellular organelles [24]. The ability to switch between the "on" and "off" states through the Rab GTP/GDP cycle empowers Rab proteins to control the spatial and temporal regulation of cargo transport [30]. Rabs interact with a cohort of effector proteins that contribute to a variety of functions, ranging from vesicle tethering, to vesicle budding and movement, and regulating the activation state of other small GTPases [31]. An ordered relay of cargo between sequentially acting compartments is thought to entail coordination of Rab activation states, coordinating changes in organelle maturation and/or allowing distinct compartments to interact at the right time and the right place for cargo transfer [32]. A Rab cascade model has been proposed that likely defines a general principle in membrane transport. This model proposes that an upstream GTP-loaded Rab protein recruits the GEF for the next Rab-GTPase along a transport pathway, activating the downstream Rab. In turn a countercurrent activity is initiated by the downstream GTP-loaded Rab, which recruits the GAP for the upstream Rab to deactivate it [24]. Together these activities are proposed to help enforce unidirectional flow. Such Rab cascades have been proposed for maturation based transport steps, such as the early endosome to late endosome transition, as well as transport steps mediated by small vesicle transport between distinct compartments, such as ER to Golgi transport [33–35].
While the molecular details of how such Rab cascades work are beginning to come to light in a small number of cases, little is known of how such activities influence endocytic recycling. In this study, we focused on the transition from early endosomes, controlled by RAB-5, to recycling endosomes, controlled by RAB-10, acting in the basolateral recycling pathway of the C. elegans intestinal epithelia. Our study shows that the downstream Rab, RAB-10, in its GTP-bound form, binds to RAB-5 GAP TBC-2 and is required for its recruitment to endosomes. Consistent with a RAB-10 to RAB-5 negative regulatory loop via TBC-2, loss of TBC-2 or RAB-10 increases association of RAB-5 with membranes, indicating abnormally high RAB-5 activation. Lack of TBC-2 also causes a dramatic morphological change in the RAB-5 labeled early endosomes. We observed accumulation of abnormally large, RAB-5-positive, pleiomorphic endosome structures, many of which displayed increased overlap with RAB-10. Thus we propose that TBC-2 can serve as a bridge in the interaction between RAB-10 and RAB-5. This model suggests that without TBC-2, RAB-5 cannot be inactivated as part of the recycling pathway, and RAB-10 endosomes cannot properly separate from RAB-5 endosomes. Our cargo localization analysis shows that in tbc-2 mutants the recycling cargo hTAC is mostly trapped in RAB-5 positive endosomes, indicating a defect in the exit of recycling cargo from early endosomes that cannot inactivate RAB-5. Our work is consistent with, and extends, work in C. elegans neurons that independently identified a connection between RAB-10 and TBC-2 important for neuropeptide secretion [19]. Thus the biogenesis and/or cargo loading of dense-core granules appears to share mechanistic similarities with endocytic recycling. Our results are also reminiscent of a counter-current GAP cascade in Saccharomyces cerevisiae that is required to restrict the spatial overlap of early and late Golgi Rabs Ypt1p and Ypt32p [36].
Our study also showed that cargo transition from early endosomes to recycling endosomes requires the coordination of another regulator of the recycling pathway, BAR-domain protein AMPH-1. Like RAB-10, AMPH-1 contributes to endosomal recruitment of TBC-2. We also detected failure in proper separation of RAB-5 and RAB-10 and failure in the exit of recycling cargo from early endosomes in amph-1 mutants, although the endosomes did not appear as grossly enlarged as in tbc-2 mutants. The AMPH-1 BAR domain binds directly to PI(4,5)P2 enriched membranes, can potentially sense membrane curvature, and can promote tubule formation [29]. An interesting possibility is that AMPH-1 derived membrane tubules could be directly involved in cargo transfer. Our previous work also showed that AMPH-1 binds to RME-1, a later acting player in the basolateral recycling pathway, potentially acting to coordinate early and late aspects of recycling [29].
Our current study delineated distinct regions of TBC-2 bound by RAB-10 and AMPH-1. Combined with our previous work showing a connection of CED-10/Rac1 to TBC-2 and recycling [27], our observations indicate that TBC-2 is a key feedback regulator of RAB-5, acting as a molecular nexus that integrates signals from recycling endosome regulators RAB-10, AMPH-1, and CED-10. The correct localization of peripheral membrane proteins is often maintained by multiple weak physical interactions, perhaps to more precisely position such proteins at points where multiple binding partners converge, a concept sometimes called coincidence sensing. Precise recruitment of TBC-2 to endosomes during recycling is likely to be quite important in the complex process of endosomal transport, where RAB-5 activity is essential for early aspects of the pathway but needs to be deactivated for later events. Such localization mechanisms may also be easily reversible, an important characteristic in dynamic situations.
In wild-type animals we found that RAB-5-labeled endosomes and RAB-10-labeled endosomes appear as distinct puncta that show partial overlap, suggesting that only a subpopulation of RAB-5 and RAB-10 labeled endosomes is interacting at any given time. This could imply the existence of transient interactions between RAB-5 and RAB-10 labeled endosomes that function to transfer cargo, removing recycling cargo as the early endosome matures into the late endosome. Intermediates in this process could be trapped, or delayed in resolution, in tbc-2, rab-10, and amph-1 mutants. Such transient interactions between early and recycling endosomes have been proposed in other systems, although the detailed mechanisms remain obscure [37]. Interestingly that work also indicated a BAR domain protein (Nwk) was involved in early endosome to recycling endosome transport, perhaps indicating cargo transfer via membrane tubules. More work will be required to understand the dynamic interactions between early and recycling endosomes that mediate cargo transfer.
All C. elegans strains were derived originally from the wild-type Bristol strain N2. Worm cultures, genetic crosses, and other C. elegans husbandry were performed according to standard protocols [38]. Strains expressing transgenes were grown at 20°C. A complete list of strains used in this study can be found in S1 Table.
Secondary structures of TBC-2 protein were predicted using the Quick2D from the Bioinformatics Toolkit (Max-Planck Institute for Developmental Biology). (Web link: http://toolkit.tuebingen.mpg.de/quick2_d)
The yeast two-hybrid experiments were performed according to the procedure of the DupLEX-A yeast two-hybrid system (OriGene Technologies). All two-hybrid plasmids were generated as PCR products with Gateway attB1.1 and attB2.1 sequence extensions and were introduced into the Gateway entry vector pDONR221 by BP clonase II (Invitrogen) reaction. The bait vector pEG202-Gtwy and target vector pJG4-5-Gtwy have been described previously [39]. Origene plasmid pSH18-34 (URA3, 8 ops.-LacZ) was used as a reporter in all yeast two-hybrid experiments. Constructs were introduced into the yeast strain EGY48 (MATα trp1 his3 ura3 leu2::6 LexAop-LEU2) included in the system. Transformants were selected on plates lacking leucine, histidine, tryptophan, and uracil and containing 2% (wt/vol) galactose/1% (wt/vol) raffinose at 30°C for 3 days and were assayed for the expression of the LEU2 reporter. The constructs of mutated forms of TBC-2 with alanine substitution were constructed by Q5—Site Directed Mutagenesis Kit (New England Biolabs, Inc.) using the cDNA sequence of TBC-2 as template.
To construct GFP or RFP/mCherry fusion transgenes that express specifically in the worm intestine, we used a previously described vha-6 promoter-driven vector modified with a Gateway cassette inserted at the Asp718I site just upstream of the GFP or RFP coding region [10]. The PCR products of the genes of interest were first cloned into the Gateway entry vector pDONR221 by BP reaction (Invitrogen). Then the PDONR221 plasmids carrying the sequences of interest were transferred into the intestinal expression vectors by Gateway recombination cloning, in a LR clonase II (Invitrogen) reaction, to generate N-terminal/C-terminal fusions [10]. Low-copy integrated transgenic lines for all of these plasmids were obtained by the microparticle bombardment method [40]. Transgenic strains pwEx142-144 were generated as following. Full-length TBC-2, TBC-2(P150A), and TBC-2(288-292AAAAA) was first cloned into entry vector pDONR221. pSM47 pSNX-1::tagRFP, pDONR221 containing TBC-2, or TBC-2(P150A), TBC-2(288-292AAAAA), pCM1.36-TBB-2 3'-UTR was inserted into the pCFJ1001 vector via multi-site LR reaction (Gateway LR Clonase II Plus Enzyme by Life Technologies). Rescue plasmids pCFJ1001::pSNX-1::tagRFP::TBC-2 (full length, P150A, or 288-292AAAAA) (10 ng/ul), pCFJ601 (50 ng/ul) and pmyo-2::GFP (coinjection marker) (10 ng/ul) were microinjected and resulting extrachromosomal arrays were used in this study [41].
For yeast two-hybrid analysis pEG202-RAB-10(Q68L), pEG202-AMPH-1(SH3), and pJG4-5-TBC-2 were constructed by gateway cloning as described previously [10,29]. For GST pull-down experiments rab-10(Q68L) and tbc-2 cDNA clones were transferred to in-house modified pcDNA3.1 (+) (Invitrogen) vectors containing 2xHA or 3xFLAG epitope tags and a Gateway cassette (Invitrogen) as described previously [11].
Live worms were mounted on 2% agarose pads with 10mM levamisole as described previously [39]. Multiwavelength fluorescence colocalization images were obtained using an Axio Imager.Z1 microscope (Carl Zeiss Microimaging) equipped with a YOKOGAWA CSU-X1 spinning disk, Photometrics Evolve 512 EMCCD camera, captured using Metamorph software (Universal Imaging), and then deconvolved using AutoQuant X5 (AutoQuant Imaging). Images taken in the DAPI channel were used to identify broad-spectrum intestinal autofluorescence caused by lipofuscin-positive lysosome-like organelles [42,43]. Quantification of colocalization images was done using the open source Fiji (Image J) software [44]. GFP/RFP colocalization experiments were performed on L4 larvae expressing GFP and RFP markers as previously described. To obtain images of GFP fluorescence without interference from autofluorescence, we used argon 488-nm excitation and the spectral fingerprinting function of the Zeiss LSM710 Meta confocal microscope system (Carl Zeiss Microimaging). Quantification of images was performed with Metamorph Version 6.3r2 (Universal Imaging).
Worms expressing intestinal GFP-RAB-5 in wild-type, tbc-2(tm2241), rab-10(q373) and amph-1(tm1060) genetic backgrounds were synchronized and cultured on NGM. Mixed stage worms were washed off with M9 buffer, pelleted and resuspended in 500μl of lysis buffer (50 mM Tris-HCL PH 8.0, 20% Sucrose, 10% Glycerol, 2 mM DTT and protease inhibitors). The worms are then disrupted using a Mini-Beadbeater-16 (BioSpec Products). Carcasses and nuclei were removed by centrifugation at 1000g for 5 min at 4°C. 200 μl of the postnuclear lysate was centrifuged at 100,000g for 1h. Pellets were reconstituted in the same volume of lysis buffer as that recovered as supernatant.
Worms expressing intestinal GFP-TBC-2 in wild-type, rab-10(ok1494), and amph-1(tm1060) genetic backgrounds were synchronized and cultured on NGM plates. 50 young adult animals of each genotype were handpicked into 10 μl of lysis buffer (100 mM Tris pH 6.8, 8% SDS, 20 mM β-mercaptoethanol) and boiled at 100°C for 10 min. Extracted worm proteins were separated by 10% SDS-PAGE and blotted to nitrocellulose. After blocking, the blot was probed with HRP-conjugated anti-GFP antibody.
rab-10(Q68L) and tbc-2 cDNA clones were transferred to in-house modified pcDNA3.1 (+) (Invitrogen) vectors containing 2xHA or 3xFLAG epitope tags and a Gateway cassette (Invitrogen) for in vitro transcription/translation experiments using the TNT-coupled transcription-translation system (Promega). Full length GST and GST-AMPH-1 was expressed and purified as previously described [29]. Eluted proteins were separated by ExpressPlus PAGE (4–20%) (GenScript), blotted to nitrocellulose, and stained with Ponceau S to detect GST fusion proteins. After blocking, the blot was probed with anti-HA (16B12) antibody and anti-FLAG M2-Peroxidase antibody (Sigma-Aldrich).
|
10.1371/journal.ppat.1004680 | Coordinated Function of Cellular DEAD-Box Helicases in Suppression of Viral RNA Recombination and Maintenance of Viral Genome Integrity | The intricate interactions between viruses and hosts include an evolutionary arms race and adaptation that is facilitated by the ability of RNA viruses to evolve rapidly due to high frequency mutations and genetic RNA recombination. In this paper, we show evidence that the co-opted cellular DDX3-like Ded1 DEAD-box helicase suppresses tombusviral RNA recombination in yeast model host, and the orthologous RH20 helicase functions in a similar way in plants. In vitro replication and recombination assays confirm the direct role of the ATPase function of Ded1p in suppression of viral recombination. We also present data supporting a role for Ded1 in facilitating the switch from minus- to plus-strand synthesis. Interestingly, another co-opted cellular helicase, the eIF4AIII-like AtRH2, enhances TBSV recombination in the absence of Ded1/RH20, suggesting that the coordinated actions of these helicases control viral RNA recombination events. Altogether, these helicases are the first co-opted cellular factors in the viral replicase complex that directly affect viral RNA recombination. Ded1 helicase seems to be a key factor maintaining viral genome integrity by promoting the replication of viral RNAs with correct termini, but inhibiting the replication of defective RNAs lacking correct 5’ end sequences. Altogether, a co-opted cellular DEAD-box helicase facilitates the maintenance of full-length viral genome and suppresses viral recombination, thus limiting the appearance of defective viral RNAs during replication.
| A major force in virus evolution is the ability of viruses to recombine and change their genomes rapidly. Similar to viral replication that greatly depends on subverted cellular proteins, viral genetic recombination is also affected by host factors based on genome-wide screens with tomato bushy stunt virus (TBSV) in yeast model host. However, the roles of host factors in the viral genomic RNA recombination process remain elusive. In this paper, we show evidence, in yeast, plants and in vitro, that co-opted cellular helicases by TBSV affect viral recombination through suppressing template-switching and replication of the new recombinant viral RNAs. Based on the presented data, a new concept emerges on the roles of co-opted cellular helicases in maintaining viral genome integrity. Altogether, the hijacked cellular DEAD-box helicases are involved in maintenance of full-length viral RNA genome and suppression of viral RNA recombination, thus blocking the appearance of defective or recombinant viral RNAs during replication.
| RNA viruses replicate inside cells and they require many cellular factors to complete their infection cycle. The intricate interactions between viruses and hosts include evolutionary arms race and adaptation that is facilitated by the ability of RNA viruses to evolve rapidly due to high frequency mutations and genetic RNA recombination as well as reassortment of genomic components [1–3]. Interestingly, cellular and environmental factors affect viral RNA recombination, which is a process that joins two or more noncontiguous segments of the same RNA or two separate RNAs together [4,5]. Recombination could alter viral genomes by introducing insertions or duplications, combining new sequences, or leading to deletions or rearrangements. RNA recombination also functions to repair truncated or damaged viral RNA molecules [2,5–7]. Viral RNA recombination can affect virus population dynamics, contribute to virus variability, as well as function in genome repair that maintains the infectivity of RNA viruses [3,4].
Viral RNA recombination is intensively studied with Tomato bushy stunt virus (TBSV), a tombusvirus infecting plants, using yeast (Saccharomyces cerevisiae) model host. TBSV is an outstanding model for both replication and recombination studies [8–12]. Systematic genome-wide screens with TBSV have led to the identification of more than 30 host genes affecting viral RNA recombination in yeast [8,9,13–15]. Among the host factors identified is the cytosolic Xrn1p 5’-to-3’ exoribonuclease (Xrn4 in plants) that suppresses TBSV recombination [16–18]. Xrn1p was shown to rapidly degrade cellular endoribonuclease-cleaved TBSV RNAs, termed degRNAs (Fig. 1A) [16–19]. The combined effects of cellular exo- and endoribonucleases determine the accumulation of degRNAs, which are especially active in RNA recombination, and thus, these cellular factors affect the frequency of viral RNA recombination events [9,18]. An additional key cellular factor involved in TBSV recombination is Pmr1 Ca++/Mn++ pump that controls Mn++ level in the cytosol [15]. Studies revealed that the cytosolic Mn++ level could greatly affect the properties/activities of the viral replicase, including its ability to synthesize RNA and switch templates. For example, high Mn++ level (in the absence of Pmr1) leads to high frequency RNA recombination in yeast or plant cells as well as in a cell-free TBSV replication assay [15].
Tombusviruses code for two replication proteins, termed p33 and p92pol, which are translated directly from the genomic (g)RNA. p92pol RNA-dependent RNA polymerase [20,21] is produced through translational readthrough of the p33 stop codon [22–24]. The abundant p33 is an RNA chaperone that functions in RNA template selection/recruitment and in the assembly of the membrane-bound viral replicase complex (VRCs) [21,25–29].
A recent systematic screen with TBSV based on a temperature-sensitive (ts) library of yeast mutants (Prasanth and Nagy, unpublished), identified the yeast Ded1p ATP-dependent DEAD-box RNA helicase as a cellular factor affecting TBSV RNA recombination. Ded1p and ten other yeast DEAD-box proteins, which are the largest family of RNA helicases, were also identified as pro-viral factors in TBSV replication in yeast [13,30–35]. DEAD-box helicases are known to be involved in all aspects of cellular metabolism [36–38], in RNA virus replication [39–42], viral translation [43,44], and affect responses to abiotic stress and pathogen infections [45–47]. They function in RNA duplex unwinding, RNA folding, remodeling of RNA-protein complexes, and RNA clamping [48].
TBSV, which does not code for its own helicase, usurps the yeast DDX3-like Ded1p (similar to the Arabidopsis AtRH20 DEAD-box helicase), to promote (+)-strand synthesis [49]. Ded1p/AtRH20 bind to the 3’-end of the TBSV minus-strand RNA, and by locally unwinding the dsRNA replication intermediate structure [50], it renders the promoter sequence accessible to p92pol for initiation of (+)-strand RNA synthesis. Additional DEAD-box helicases, such as Dbp3p (human DDX5-like) and Fal1p (eukaryotic translation initiation factor eIF4AIII-like), which are involved in ribosome biogenesis in yeast [51–53], and the orthologous Arabidopsis RH2 and RH5 helicases bind to the 5’ proximal region in the TBSV (-)RNA [54]. This region harbors a critical replication enhancer element (REN) [55]. These co-opted cellular helicases can locally unwind the double-stranded (ds) structure within the REN of the replication intermediate and enhance (+)RNA synthesis [50,54]. Altogether, the co-opted cellular DEAD-box helicases work synergistically to enhance TBSV replication by interacting with the viral (-)RNA, dsRNA and the replication proteins within the VRCs [54].
In this work, we show evidence that Ded1p/AtRH20 helicases are strong suppressors of TBSV recombination in yeast and plants. In vitro assays show direct involvement of Ded1p in suppression of viral recombination, which requires its ATPase function. Moreover, the presented data support a new role for Ded1p in facilitating the switch from (-)-strand to (+)-strand synthesis. Interestingly, the eIF4AIII-like AtRH2 helicase enhances TBSV recombination in the absence of Ded1/AtRH20, suggesting that the coordinated action of cellular Ded1/AtRH20 and AtRH2 helicases control viral RNA recombination events. We propose a model on the role of Ded1/AtRH20 in facilitating the replication of full-length viral RNAs with intact 5’ ends while inhibiting the replication of 5’-truncated viral RNAs, thus playing a major role in maintaining the intact genome structure for TBSV.
To characterize the role of the DDX3-like Ded1p DEAD-box RNA helicase of yeast in TBSV RNA recombination, first we utilized genetic approaches in yeast. Depletion of Ded1p resulted in ~5-fold increase in TBSV recombinant (rec)RNA accumulation (Fig. 1B, lanes 13–16). Similarly, yeast expressing either Ded1–95ts or Ded1–199ts temperature-sensitive mutants as a single source for Ded1p, led up to 5-to-10-fold increase in TBSV recRNA levels at the semi-permissive temperature (Fig. 1C, lanes 13–16 and 21–24 versus 17–20). Ded1–199ts also supported ~35-fold higher recRNA accumulation at a lower (permissive) temperature (Fig. 1C, lanes 9–12), suggesting that this particular mutant is especially suitable for viral RNA recombination studies. Ded1–199ts (G368D mutation) is known to debilitate its function in protein translation and intron splicing [56], while Ded1–95ts (T408I mutation) does not affect splicing, but maybe involved in translation and RNA decay [57]. Altogether, the above yeast genetic approaches have conclusively demonstrated that the wt Ded1p helicase is a strong suppressor of TBSV RNA recombination in yeast cells.
The most frequent recombinants in the TBSV system are generated via template-switching mechanism by the viral replicase using viral RNA templates that are cleaved by cellular endo- and exoribonucleases (schematically shown in Fig. 1A) [5,8,9,15,17,18,58]. The partially degraded (5’-truncated) viral RNA products generated by the cellular nucleases are called degRNAs, which serve as templates for viral RNA recombination (Fig. 1A) [8,9]. Interestingly, the amounts of degRNAs also increased by ~3-to-10-fold, suggesting their efficient generation and replication in Ded1ts mutant yeasts (Fig. 1C) or in yeast with depleted Ded1p (Fig. 1B). Interestingly, the degRNAs are superior templates for high frequency recombination when expressed in yeast cells in the presence of the viral p33/p92pol replication proteins (Fig. 1D) [8,9]. Both Ded1–95ts and Ded1–199ts supported 3-to-8-fold higher recRNA accumulation from the DI-RIIΔ70 degRNA template than the wt Ded1p did in yeast (Fig. 1D). These data further supported the suppressor function of Ded1p in TBSV recRNA accumulation in yeast.
The surprisingly robust accumulation of the 5’-truncated degRNAs in both ded1–95ts and ded1–199ts yeasts expressing the full-length DI-AU-FP repRNA (Fig. 1C) was likely due to enhanced efficiency of their replication, because expression of the representative DI-RIIΔ70 degRNA accumulated to high level (up to ~5-fold increase) in ded1–95ts and ded1–199ts yeast strains in comparison with the wt yeast (Fig. 1D). These findings indicate an unexpected role of Ded1p in suppressing the replication of the 5’-truncated degRNAs. This is in contrast with the pro-viral role of Ded1p in enhancing the accumulation of TBSV DI-72 repRNA, which carries the authentic 5’ end sequence (see also below) [49,54].
Testing the accumulation of (+) versus the (-)RNA products revealed ~9-fold increased (-)recRNA production in case of ded1–199ts yeast at semi-permissive temperature in comparison with wt yeast (Fig. 2A). Interestingly, similar to the high level of (-)recRNAs, accumulation of truncated (-)degRNAs was also increased by ~4-fold, while the amount of full-length (-)repRNA changed only slightly (DI-AU-FP repRNA, Fig. 2A) in ded1–199ts yeast. Altogether, (-)recRNAs accumulated to ~3-fold higher level than the full-length DI-AU-FP (-)repRNA in ded1–199ts yeast (Fig. 2A). On the contrary, the DI-AU-FP repRNA was the most prevalent (+)RNA product, while the (+)recRNAs and (+)degRNA products accumulated to 3-to-7-fold lesser amounts than the (+)repRNA in ded1–199ts yeast (Fig. 2A).
Since ded1–95ts and ded1–199ts mutations are present within the RNA binding domain of the Ded1p helicase [56], we have tested if the mutants show altered viral RNA binding characteristic when compared with the wt Ded1p. The EMSA assay with DI-72(-) RNA template revealed that the purified ded1–95ts and ded1–199ts mutants bound to the viral (-)RNA with up to 25-fold reduced efficiency in vitro (Fig. 2B). The low efficiency in viral (-)RNA binding by these Ded1p mutants could be the reason for these mutants supporting the increased rate of viral recombination, high level of degRNA accumulation and reduction in viral (+)-strand synthesis (see Discussion).
In comparison with the results obtained via ded1–95ts and ded1–199ts mutants, we observed a similar trend with increased accumulation of (-)recRNA and (-)degRNA products obtained with the recombinogenic DI-AU-FP repRNA, when yeast expressed Ded1p at a reduced level (+doxycycline treatment, Fig. 3A-B). Altogether, these data revealed that Ded1p is important in regulation of (+) versus (-)RNA products and this regulation depends on the presence of the authentic 5’ end sequence from TBSV (+)repRNA.
To confirm the importance of co-opted Ded1p in viral RNA replication and recombination, we also tested the accumulation of various Δ+) and (-)RNA products with the efficient DI-72 repRNA, which replicates to the highest level among all TBSV RNAs in yeast and plants cells [59,60]. As expected based on previous publications [49,61], depletion of Ded1p by doxycycline in TET::DED1 yeast, reduced the accumulation of DI-72 (+)repRNAs by ~4-fold, while the accumulation of new (+)recRNAs and (+)degRNA products was below the detection limit (top image in Fig. 3C, lanes 3–4 and 7–8). Interestingly, however, (-)recRNA and (-)degRNA products, which were almost as abundant as the full-length DI-72 (-)repRNA, were detected in yeasts with depleted Ded1p level (bottom image in Fig. 3C, lanes 3–4 and 7–8). The corresponding (-)recRNA and (-)degRNA products were below detection limit in yeasts expressing Ded1p to high level (bottom image in Fig. 3C, lanes 1–2 and 5–6). Altogether, these results demonstrate that Ded1p plays a critical role in suppression of the formation and accumulation of recRNA and degRNA products during minus-strand synthesis.
To dissect the inhibitory function of Ded1p in recRNA formation and degRNA replication, first we used an in vitro assay with isolated yeast membranes [62]. The yeast membrane fraction contains the tombusvirus replicase in complex with the viral RNAs, thus facilitating studies on the viral RNAs functionally associated with the replicase. Denaturing PAGE analysis of the in vitro replicase products revealed that both recRNAs and degRNAs were actively replicated by the tombusvirus replicase derived from ded1–199ts yeast (~13-to-21-fold higher level than in wt replicase), while these RNAs were barely detectable in the replicase from wt yeast (Fig. 4A). In addition, ded1–199ts replicase supported ~6-to-7-fold higher level of (-)recRNAs and (-)degRNAs in comparison with slightly reduced DI-AU-FP (-)repRNA carrying the authentic 5’ end sequence in vitro (Fig. 4B). The (+)recRNAs and (+)degRNAs accumulated to 3-fold higher level in ded1–199ts yeast than the corresponding RNAs in wt yeast, but (+)recRNAs and (+)degRNAs were ~12-fold less abundant than the DI-AU-FP (+)repRNA in vitro (Fig. 4B-C). Thus, similar to the situation in yeast cells, wt Ded1p suppressed in vitro (-)-strand synthesis with the recRNAs and degRNAs, but not with DI-AU-FP repRNA carrying the authentic 5’ end sequence.
The second assay was based on a cell-free extract (CFE) from yeast with depleted Ded1p that was used to assemble the tombusvirus replicase in vitro using purified recombinant p33/p92pol and (+)repRNAs (Fig. 4D) [49]. The CFE supports a complete replication cycle resulting in both (-) and (+)-stranded repRNA products [29]. As expected, Ded1p facilitates the production of (+)repRNAs carrying the authentic 5’ end sequence [Fig. 4E, lanes 3–4, see reduced DI-72 repRNA accumulation in CFE with depleted Ded1p (+dox)] [49]. Similarly, addition of the purified recombinant wt Ded1p to the CFE programmed with DI-AU-FP (+)repRNA, which carries the authentic 5’ end sequence, led to a ~50% increase in repRNA accumulation (Fig. 4F, lane 2), while the ATPase-deficient D1 mutant of Ded1p [49,63] could not stimulate repRNA replication in vitro (Fig. 4F, lane 3). On the contrary, addition of the purified recombinant wt Ded1p or D11 mutant with increased ATPase activity [63] to the CFE with the 5’-truncated DI-RIIΔ70 degRNA, led to ~40–50% decrease in degRNA accumulation (Fig. 4G, lanes 2 and 4 versus lane 1), while D1 mutant did not affect the replication of DI-RIIΔ70 degRNA in the CFE (lane 3). Based on these data from CFE assays, we conclude that Ded1p inhibits the replication of recRNAs or degRNAs missing the authentic 5’ end sequence likely through blocking the (-)-strand synthesis on these RNA templates.
Based on known features of DEAD-box helicases in remodeling protein-RNA complexes [48,64], we reasoned that Ded1p might be involved in releasing the p92 RdRp protein from the (+)RNA template at the end of (-)-strand synthesis, thus decreasing the chance for template-switching events (see Discussion). To test this model, we developed an in vitro assay with a soluble form of p92, called p92-Δ167N, which can specifically use TBSV-derived (+)RNA template for RNA synthesis in vitro in the presence of biotynylated UTP and other ribonucleotides as shown schematically in Fig. 5A [21]. The biotynylated viral dsRNA form was then captured via streptavidin beads (Fig. 5B). We then added purified Ded1p to the beads to facilitate the putative release of the p92-Δ167N RdRp from the captured dsRNA product. The amount of dsRNA-bound versus released p92-Δ167N was measured by Western blotting (Fig. 5B-C). These experiments revealed that three-times more p92-Δ167N was released from the viral dsRNA product when purified wt Ded1p was included in the assay (Fig. 5C, lanes 4 versus 1).
In another assay, we used EMSA with MBP-p92-Δ167N and purified GST-Ded1p based on 32P-labeled RI(+) RNA as a probe. Both MBP-p92-Δ167N and GST-Ded1p bind to the probe when applied alone (Fig. 5D, lanes 3 and 14, respectively). However, when we added p92-Δ167N first to the probe, followed 15 min latter by addition of GST-Ded1p, then the release of the probe was detectable in the form of diffused label (“smear”) (Fig. 5D, lanes 4–5 versus 6–7 with purified GST as a control). Interestingly, the release of the probe was dependent on the presence of ATP, suggesting that Ded1p requires ATP for this function (Fig. 5D, lanes 8–9 versus 4–5). The diffused label was also observed when Ded1p was added first to the RNA, followed by p92-Δ167N (Fig. 5D, lanes 12–13), suggesting that Ded1p and p92-Δ167N likely form a complex that releases the viral RNA.
To establish the function of Ded1p during TBSV replication and RNA recombination, we examined if Ded1p affects these processes via controlling Xrn1p 5’-to-3’ exoribonuclease, which is a key enzyme in TBSV RNA stability and for suppression of TBSV RNA recombination in yeast [5,16–18,65]. For these studies, we expressed a 5’-truncated repRNA (DI-ΔRI, Fig. 6A), which goes through further 5’-truncations (up to ~70 nt, where RII(+)-SL hairpin structure stops the nuclease activity) in the presence of Xrn1p in wt yeast (Fig. 6A), while this truncation process is weak in xrn1Δ yeast (Fig. 6B) [65]. DI-ΔRI RNA did not accumulate in ded1–199ts yeast, similar to wt yeast (Fig. 6B, lanes 13–16 and 1–4), while DI-(RI accumulated to high level in xrn1Δ yeast (Fig. 6B, lanes 5–8). Also, the profile of recRNAs accumulating in ded1–199ts yeast was similar to that in wt yeast and different from that in xrn1Δ yeast (Fig. 6B). Thus, it seems that ded1p mutation does not affect TBSV RNA recombination and degRNA accumulation via inhibition of the Xrn1p activity. This conclusion was further supported by RNA stability experiments that showed comparable half-life for degRNA in ded1–95ts and ded1–199ts yeasts to the wt yeast (Fig. 6C).
Because TBSV replication is known to depend on two types of cellular DEAD-box helicases, namely the DDX3-like Ded1p/AtRH20 that bind to the 3’end of the (-)RNA and the eIF4AIII-like Fal1p/AtRH2 helicases that bind to a 5’ proximal enhancer element in the (-)RNA (Fig. 7A) [49,54], we also tested the effect of expression of AtRH2 on TBSV recombination in yeast. We observed up to ~12-fold enhanced level of TBSV RNA recombination in wt and 26-fold increase in ded1–199ts yeasts expressing AtRH2 (Fig. 7B, lanes 5–6, 17–18 and 11–12, 23–24). In contrast, expression of AtRH20 helicase (Fig. 7C), which is a Ded1p ortholog, suppressed recRNA accumulation in both wt and ded1–199ts yeasts (Fig. 7B). Thus, different co-opted cellular helicases have opposite effects on TBSV recombination in yeast.
To test if AtRH2 has direct function in TBSV recombination, we used the CFE-based TBSV replication assay prepared from yeast with depleted Ded1p (Fig. 7D). Interestingly, the addition of purified recombinant AtRH2 increased the replication of the 5’-truncated DI-RIIΔ70 degRNA by ~2-fold and RNA recombination also by ~2-fold (Fig. 7E, lanes 3–4 versus 1–2). However, the stimulatory effect of AtRH2 on RNA recombination is neutralized by the addition of purified recombinant Ded1p helicase (Fig. 7E, lanes 5–6), suggesting that AtRH2 only promotes formation of recRNAs and the replication of the 5’-truncated degRNAs when Ded1p helicase is depleted. In other words, Ded1p helicase seems to be the dominant factor with its recombination suppressor activity.
To confirm the roles of the above cellular helicases in TBSV RNA recombination in plants, we over-expressed AtRH2 and AtRH20 in Nicotiana benthamiana plants also expressing DI-AU-FP repRNA in the presence of Cucumber necrosis virus (CNV), a closely related tombusvirus that serves as a helper virus for the TBSV repRNA. The helper tombusvirus provides the p33 and p92 replication proteins in trans for the replication of repRNA and the de novo generated recRNAs in this system. We found that the Ded1p ortholog AtRH20 suppressed TBSV recRNA accumulation by ~2-fold, while AtRH2 increased recRNAs by ~2-fold (Fig. 8A). These data indicate that the different co-opted cellular helicases have opposite effects on TBSV recombination in plants. Over-expression of AtRH20 also suppressed the accumulation of the 5’-truncated DI-RIΔ degRNA and the further truncated degRNAs, ultimately resulting in ~3-fold less recRNA accumulation than in control plants (Fig. 8B, lanes 4–6 versus 1–3). Based on these results, we suggest that the roles of the two cellular helicases in plants are comparable to the functions of these helicases in vitro in the CFE assay and in yeast.
Viral RNA recombination and the generation of defective viral RNA molecules are thought to be chance events that take place during viral RNA replication. It is possible that RNA viruses regulate these unique processes to guarantee the efficient replication of the full-length viral RNA and to reduce the competition of various viral RNAs for viral- and host factors. Accordingly, the role of viral replicase proteins in viral RNA recombination and defective RNA generation has been documented before [4,66–71]. However, based on systematic genome-wide screens performed with TBSV in yeast surrogate host [8,9], a new concept on the key roles of cellular factors in viral RNA recombination and defective RNA generation is emerging [5,15–18,58,65]. Among such cellular factors are DEAD-box helicases as demonstrated in this paper.
This work based on genetic approaches with Ded1p ts mutants or depletion of Ded1p in yeast and in vitro approaches with cell-free replication of TBSV RNAs strongly supports a TBSV recombination suppressor activity of the co-opted Ded1p cellular helicase. Since the ATPase-deficient D1 mutant of Ded1p does not have recombination suppressor activity in vitro (Fig. 4), it seems that Ded1p helicase has a direct inhibitory function in TBSV RNA recombination. The AtRH20 helicase, a plant ortholog of Ded1p, also has similar recombination suppressor activity in yeast and in plants. Importantly, the recombination suppressor activity of Ded1p is independent of the recombination suppressor activity of the previously characterized Xrn1p 5’-to-3’ exoribonuclease, which acts by efficiently removing degRNAs and recRNAs generated during TBSV replication (Fig. 6) [16–18,65]. Altogether, Ded1p helicase is the first co-opted cellular factor in the viral replicase complex that has been shown to directly affect viral (+)RNA recombination.
A previously demonstrated function of co-opted Ded1p helicase is to locally unwind the double-stranded RNA replication product after the (-) RNA synthesis is completed on the (+)RNA template (Fig. 9A) [49,50,54]. Ded1p then facilitates the loading of the viral replicase onto the 3’ end of the (-)-stranded RNA portion of the dsRNA intermediate with the assistance of the co-opted cellular glyceraldehyde-3-phosphate dehydrogenase (GAPDH) [49,50,54]. Thus, ultimately, Ded1p promotes the asymmetrical (i.e., excess) production of new (+)-strand RNAs by allowing the selective use of the (-)RNA in the dsRNA intermediate template.
However, this work also reveals a new role of Ded1p in inhibition of (-)-strand synthesis, especially with those RNA templates that lack authentic 5’ sequences, such as degRNAs and recRNAs (Figs. 2–3). Interestingly, the amount of (-)recRNAs and (-)degRNAs far exceeds the DI-AU-FP (-)repRNA in Ded1p deficient yeast or in vitro when Ded1p mutant is present, while the (+)repRNA is more abundant than (+)recRNAs or (+)degRNAs (Figs. 2–3). In case of the highly efficient DI-72 repRNA, the (-)recRNAs and (-)degRNAs are only detected when Ded1p is depleted (Fig. 3). Thus, one major function of the co-opted wt Ded1p is to promote the efficient replication of only the full-length viral RNAs, while suppressing the replication of 5’-truncated viral RNAs, lacking critical cis-acting elements. This novel function of Ded1p in maintenance of genome integrity seems to be manifested during (-)-strand synthesis.
Ded1p-driven suppression of replication of degRNAs might be critical in cells loaded with cytosolic ribonucleases that likely generate many truncated viral RNAs. These defective RNAs could likely compete with the full-length viral RNAs for viral- and host factors, thus leading to reduced viral replication. However, the co-opted cellular Ded1p helicase facilitates proper replication of TBSV RNAs and protects TBSV from competition by defective viral genomes. Since Ded1p inhibits the replication of recRNAs or degRNAs lacking the authentic 5’ end sequence through blocking the (-)-strand synthesis on these RNA templates, lesser amount of defective viral dsRNAs could accumulate. The reduced amount of dsRNA is an advantage for the virus, because dsRNAs could efficiently trigger antiviral responses, such as RNAi (or RNA silencing in plants) [72–74].
Another surprising finding in this study is the stimulatory effect of a second group of co-opted cellular DEAD-box helicases on TBSV RNA recombination. Accordingly, over-expression of the eIF4AIII-like AtRH2 in yeast or plant cells led to increased level of recRNA accumulation (Figs. 7–8).
The AtRH2 helicase binds to the 5’ proximal region of the viral (-)RNA, which harbors the RIII(-) REN, resulting in localized unwinding of the dsRNA replication intermediate [50,54]. Although this unwinding process is important for the replication of the full-length TBSV RNA, it seems that it only works “properly” for TBSV replication when Ded1p/AtRH20 helicase is also present in the replicase complex. Based on these observations, the emerging concept is that the coordinated action of these two co-opted cellular helicases is required for efficient replication of the full-length viral RNA. If Ded1p is missing or the eIF4AIII-like AtRH2 is present in excess amount within the replicase complex, then the frequency of viral RNA recombination increases and replication of 5’-truncated viral degRNAs becomes more efficient. Therefore, these conditions favor the rapid evolution of TBSV, which could be advantageous under some circumstances, but disadvantageous when the wt TBSV is the best-adapted to the host/environment.
Previous works revealed roles for Ded1p/AtRH20 and AtRH2 DEAD-box helicases during TBSV (+)-strand synthesis [49,54], which was based on local unwinding of the dsRNA replication intermediate to facilitate initiation of (+)-strand synthesis by the viral replicase (Fig. 9A). However, this work unearthed a novel function for Ded1p helicase by showing an increased level of (-)RNA production from recRNAs and degRNAs in yeast expressing mutant Ded1p or with depleted level of Ded1p. To explain these findings, we propose that Ded1p helicase facilitates the displacement of the viral p92 RdRp protein from the dsRNA product at the end of (-)-strand synthesis, as shown schematically in Fig. 9A. In case of the full-length viral RNA, the localized unwinding of the “left side” of the dsRNA then promotes the association of the p92 RdRp with the 3’ cis-acting elements in the (-)RNA portion of dsRNA, followed by (+)-strand synthesis via strand-displacement mechanism as shown before [50]. Thus, basically, Ded1p/AtRH20 helicases could promote the switch from (-)- to (+)-strand synthesis.
In case of the 5’-truncated RNAs, Ded1p/AtRH20 helicases could likely displace the p92 RdRp from the 5’ end of the degRNAs (Fig. 9B). Displacement of p92 RdRp from the template would likely inhibit template-switching events during (-)-strand synthesis. Accordingly, in vitro assays support this model by providing evidence that Ded1p promotes dissociation of p92 RdRp from the viral RNA (Fig. 5). Moreover, Ded1p helicase might not be able to open the ds degRNA to facilitate initiation of (+)-strand synthesis due to the absence of RI(-) sequence (i.e. Ded1p binding sequence) in the (-)degRNA [49]. Indeed, all degRNAs identified lack the authentic 3’ end viral sequences in the (-)RNA [15,18,65]. Based on these, we propose that Ded1p helicase suppresses the use of 5’-truncated degRNAs in (+)-strand synthesis. Overall, the p92 displacement ability of Ded1p likely inhibits template-switching RNA recombination and the replication of recRNAs (Fig. 9B).
However, when Ded1p is depleted or mutant Ded1p is present, then p92 RdRp protein will not be efficiently displaced from the dsRNA [after finishing (-)RNA synthesis on the (+)RNA template], and this condition then facilitates template-switching-based RNA recombination (Fig. 9B). In addition, the replication of degRNAs and recRNAs is also increased in the absence of functional Ded1p, likely due to the presence of AtRH2 type helicase in the VRCs, which facilitates unwinding of the “right-side” of the dsRNA template, thus promoting re-initiation on the plus-strands of dsRNA templates to generate new minus-strands (Fig. 9B). AtRH2 cannot facilitate re-initiation on the (+)RNA when Ded1/AtRH20 is present due to the recruitment of p92 to the (-) 3’-end sequences by Ded1p, long-range RNA-RNA interactions and additional cellular factors, such as GAPDH, as described earlier [54]. Altogether, the above events could explain the increased level of (-)RNAs from degRNAs and recRNAs in yeast either expressing mutant Ded1p or with depleted Ded1p.
Overall, the novel function of the DDX3-like Ded1p/RH20 helicases is the down-regulation/inhibition of (-)RNA synthesis by promoting the efficient switch from (-)RNA to (+)RNA synthesis. Interestingly, this feature requires the authentic viral 3’ end sequences on the (-)RNA, suggesting similarities between telomeres and viral RNA synthesis in protection of the ends of linear nucleic acids [75,76]. Those viral RNAs lacking the authentic terminal sequences could replicate less efficiently in the presence of Ded1p/AtRH20 helicases, suggesting that TBSV recruits a cellular helicase to protect and promote the replication of the full-length viral RNAs, while suppressing the accumulation of recRNAs and degRNAs during viral infections. Therefore, based on this work, a new concept emerges on the roles of co-opted cellular helicases in maintaining viral genome integrity.
The yeast (Saccharomyces cerevisiae) strain BY4741 (MATa his3(1 leu2(0 met15Δ0 ura3(0), R1158 and TET::DED1 (yTHC library) were obtained from Open Biosystems. The temperature-sensitive (ts) yeast strains ded1–95ts and ded1–199ts were of generous gift from C. Boone (U. Toronto).
The yeast expression plasmids LpGAD-His92 (containing CNV p92pol gene behind the CUP1 promoter, LEU2 selection), and HpHisGBK-His33/DI-AU-FP (co-expressing p33 from the CUP1 promoter and DI-AU-FP repRNA from GAL1 promoter, HIS3 selection), HpHisGBK-HFHis33/DI-72 (co-expressing p33 from the CUP1 promoter and DI-72 repRNA from GAL1 promoter, HIS3 selection), UpGBK-His33/DI-AU-FP (co-expressing p33 from the ADH1 promoter and DI-AU-FP repRNA from GAL1 promoter, URA3 selection) have been previously described [8]. Plasmid HpGBK-His33/DI-RIIΔ70 was made by PCR-amplification of DI-72RIIΔ70 with primers #1546 (CCGCGAATTCACGGATTAGAAGCCGCCGAGCGGGT) and #1069 (CCGGTCGAGCTCTACCAGGTAATATACCACAACGTGTGT). The PCR product was restriction digested with EcoRI and SacI and ligated into the plasmid HpGBK-His33/DI-72 to replace the full-length DI-72 fragment.
The E. coli expression plasmids pMAL-Ded1, pMAL-D1, pMAL-D11 and pMAL-RH2 were prepared previously [49,54]. To prepare plasmids for expression of MBP-tagged Ded1p ts mutants, genomic DNA of ded1–95 or ded1–199 strains were used as templates in PCR with primers #3956 (CCAGCTGCAGTCACCACCAAGAAGAGTTG) and #3957 (CCAGGAATTCATGGCTGAACTGAGCGAACAAG). The PCR products were cloned into pMalc-2x vector at EcoRI and PstI sites. For GST-Ded1p expression, pMAL-Ded1 was used as a template in PCR with primers #4308 (TGGAACTTGGAATTGTTTACACCTTAGTCTGTTGACTTAA) and #4309 (CCAGCTCGAGTCACCACCAAGAAGAGTTG). The PCR products were cloned into pGEX-his-RE vector at SpeI and XhoI site.
The plant expression plasmids pGD-RH2, pGD-RH20, pGD-CNV and pGD-DI-AU-FP were described previously [54,65]. To obtain plasmids HpGBK-His-RH20 and HpGBK-His-RH2, the plasmids pMAL-RH20 or pMAL-RH2 [54,61], respectively, were used as templates in PCR with primers #4315 (CCAGGGATCCATGAGTGCATCATGGGCAG) and #4316 (CCAGCTGCAGCTAATCCCAAGCACTGGTC) for RH20; and #4816 (CCAGGGATCCATGGCGACAGCGAATCCTGG) and #5117 (CCAGCTGCAGTTAGATAAGATCAGCTACATTC) for RH2 open reading frames. The PCR products were cloned into HpGBK-His vector at BamHI and PstI sites.
Yeast strains were co-transformed with plasmids by using the lithium acetate/ssDNA/polyethylene glycol method, and transformants were selected by complementation of auxotrophic markers [77]. For TBSV recombination assay in BY4741, ded1–95ts, ded1–199ts, R1158 and TET::DED1, yeast strains were co-transformed with LpGAD-His92 and HpGBK-His33/DI-AU-FP, HpGBK-His33/DI-RIIΔ70 or HpHisGBK-HFHis33/DI-72. The transformed BY4741, ded1–95ts, and ded1–199ts strains were pre-grown at 23°C overnight in SC-LH- (synthetic complete media without histidine and leucine) media with 2% galactose. Then, 50 μM CuSO4 was added to the yeast cultures to launch virus replication and recombination. Yeast was grown at either 23°C or 29°C for 24 h before sample collection for analysis. The transformed R1158 and TET::DED1 strains were pre-grown at 29°C overnight in SC-ULH- (synthetic complete media without uracil, histidine and leucine) media with 2% galactose containing 10 μg/ml doxycycline. Then, 50 μM CuSO4 was added to the yeast cultures to launch virus replication and recombination at 23°C or 29°C for 24 h.
In the complementation study, BY4741 and ded1–199ts strains were co-transformed with LpGAD-His92, UpGBK-His33/DI-AU-FP and the indicated plasmids (HpGBK-HisRH20 or HpGBK-HisRH2) expressing one of the host helicases. The transformed yeast strains were pre-grown at 23°C overnight in SC-ULH- media with 2% galactose, followed by the addition of 50 μM CuSO4 and culturing at 23°C or 29°C for 24 h.
For viral RNA stability assay, BY4741 ded1–95ts, and ded1–199ts strains were transformed with UpYC-DI-RIIΔ70 [18]. The transformed yeast strains were grown at 23°C in SC-U- media with 2% galactose. After 24 h, the cultures were re-suspended in SC-U- media with 2% glucose and grown at 23°C or 29°C. The samples were collected at given time points mentioned in figure legend.
To observe the TBSV DI-(RI RNA recombination profile in BY4741, (Xrn1, (Met22, and ded1–199ts yeast strains, they were co-transformed with HpGBK-His33, LpGAD-His92 and pYC2-DI-(RI. The transformed cultures were inoculated on to ULH-/glucose media and grown at 23°C for 12 hrs. Yeast cultures were collected by centrifugation and dissolved in ULH-/galactose media supplemented with 50 μM CuSO4. Cultures were grown at 23°C for two days before sample collection for RNA analysis.
TBSV RNA recombination was analyzed using total RNA extracted from yeast and plants. Standard RNA extraction and Northern blot analysis was performed as described in previous publication [78]. To detect TBSV (+)RNA or (-)RNA, we prepared 32P-labeled DI-72RIII/IV probe with in vitro T7-based transcription using PCR-amplified DNA obtained on HpGBK-His33/Gal1-DI-72 template with primers #22 (GTAATACGACTCACTATAGGGCTGCATTTCTGCAATGTTCC)/ #1165 (AGCGAGTAAGACAGACTCTTCA) for (+)RNA detection; and primers #18 (GTAATACGACTCACTATAGGAGAAAGCGAGTAAGACAG) / #1190 (GGGCTGCATTTCTGCAATG) for (-)RNA detection.
Typhoon FLA 9500 system (GE) and ImageQuant TL software were used to detect and quantify the bands in the gels. The repRNA and degRNA bands were identified based on molecular markers, while the recRNAs were identified based on previously characterized recRNAs [15,16,58,65]. Only the bands representing the major recRNAs (which are pointed at in figures) were quantified. All these RNAs were normalized based on ribosomal RNA level in all samples.
Recombinant MBP-tagged helicase proteins and MBP-tagged TBSV p33 and p92 replication proteins or MBP-p92Δ167N were expressed in E coli and purified as published before [49]. Briefly, the expression plasmids were transformed into E. coli strain BL21 (DE3) CodonPlus. Protein expression was induced by isopropyl-β-D-thiogalactopyranoside (IPTG) at 16°C for 8 h. After collection of the cultures by centrifugation at 4000 xg for 5 min, the cells were re-suspended and broken in reduced-salt column buffer (25 mM NaCl, 30 mM HEPES-KOH pH 7.4, 1 mM EDTA, 10 mM β-mercaptoethanol). The lysate was centrifuged at 14,000 rpm for 10 min to remove cell debris. Then, the supernatant was incubated with amylose resin (NEB) at 4°C for 1 h. After washing the resin with 50 ml reduced-salt column buffer (without β-mercaptoethanol), the recombinant proteins were eluted in maltose buffer (column buffer containing 0.18% (W/V) maltose).
The membrane-enriched fraction (MEF) was obtained as published previously [62,78]. Briefly, yeast strains were transformed and grown as described above for TBSV recombination in yeast. Yeast cultures were collected and processed to obtain the MEFs containing the in vivo-assembled replicase complexes as previously described [78]. Each membrane fraction preparation was adjusted based on the relative amounts of His6-tagged p33 and comparable amounts of replicase (based on p33) from each preparation were used in the subsequent in vitro replicase assay. The replicase assay was performed as described [62,78]. Briefly, the in vitro assay (50 μl) contained 10 μl of the normalized MEF preparations, 10 mM DTT, 50 mM Tris-Cl pH 8.0, 10 mM MgCl2, 0.1 U RNase inhibitor, 1 mM ATP, 1 mM CTP, 1 mM GTP and 0.1 μl of α32P-UTP (3000 Ci/mmol). Reaction mixtures were incubated at 25°C for 3h, followed by phenol/chloroform extraction and isopropanol/ammonium acetate (10:1) precipitation. 32P-labeled RNA products were analyzed in 5% acrylamide/8 M urea gels. To detect the membrane associated RNA, membrane preparations that contained comparable amounts of replicase were used to extract the viral RNA by standard phenol/chloroform extraction and isopropanol/ammonium acetate (10:1) precipitation. Then, the RNAs were analyzed by Northern blotting with (+) or (-) RNA specific probes.
CFEs from BY4741 and TET::DED1 treated with 10 μg/ml doxycycline were prepared as described earlier [29,62] and adjusted to contain comparable amounts of total protein. The in vitro CFE-based assays were performed in 20 μl total volume containing 1 μl of adjusted CFE, 0.5 μg DI-72 (+), DI-RIIΔ70 (+) or DI-AU-FP (+)RNA transcripts (separately), 0.5 μg purified MBP-p33, 0.5 μg purified MBP-p92pol (both recombinant proteins were purified from E. coli) [79], 30 mM HEPES-KOH, pH 7.4, 150 mM potassium acetate, 5 mM magnesium acetate, 0.13 M sorbitol, 0.4 μl actinomycin D (5 mg/ml), 2 μl of 150 mM creatine phosphate, 0.2 μl of 10 mg/ml creatine kinase, 0.2 μl of RNase inhibitor, 0.2 μl of 1 M dithiothreitol (DTT), 2 μl of 10 mM ATP, CTP, and GTP and 0.25 mM UTP and 0.1 μl of 32P-UTP. 0.3 μg of MBP-Ded1, MBP-D1 or MBP-D11, respectively, was added to the assay to test their activities during viral RNA synthesis. Reaction mixtures were incubated for 3 h at 25°C, followed by phenol/chloroform extraction and isopropanol/ammonium acetate (10:1) precipitation. 32P-labeled RNA products were analyzed in 5% acrylamide/8 M urea gels [62].
The 32P-labeled full-length DI-72 (-)RNA and the RI(+) RNA were generated as described [79]. Ded1p and ts mutants were incubated with 5 ng of 32P-labeled DI-72(-) RNA probe in a binding buffer (50 mM Tris-HCl [pH 8.2], 10 mM MgCl2, 1 mM EDTA, 10% glycerol, 200 ng of yeast tRNA [Sigma], and 2 U of RNase inhibitor [Ambion]) at 25°C for 15 min. After the binding, the samples were analyzed by 5% nondenaturing PAGE performed at 200 V for 1 h in a cold room. To test the template release activity, briefly, 32P-labeled RI(+)RNA probe was incubated with p92-Δ167N at 25°C for 15 min, followed by adding affinity-purified GST-Ded1p to the reaction with or without 1 mM ATP, then the reaction was incubated at 25°C for 30 min. In addition, probe was also incubated with proteins in a different order mentioned in figure legend.
First, p92-Δ167N RdRp assay was used to produce the biotin-labeled partial dsRNA product [21]. Briefly, the in vitro RdRp reaction was performed in 20 μl total volume containing 1 μl of adjusted CFE (soluble fraction only), 0.5 μg DI-mini (+)RNA transcript [21], 0.5 μg affinity-purified MBP-p92-Δ167N, 30 mM HEPES-KOH, pH 7.4, 150 mM potassium acetate, 5 mM magnesium acetate, 0.13 M sorbitol, 0.2 μl actinomycin D (5 mg/ml), 2 μl of 150 mM creatine phosphate, 0.2 μl of 10 mg/ml creatine kinase, 0.2 μl of RNase inhibitor, 0.2 μl of 1 M dithiothreitol (DTT), 2 μl of 10 mM ATP, CTP, and GTP and 0.1 mM UTP and 0.1 μl of biotin-UTP. Reaction mixture was incubated at 25°C for 30 min. Note that we combined 10 separate in vitro reactions in the subsequent experiment. After incubation, the free, unincorporated biotin-UTP was removed by Sepharose G-25 column. Then, 200 μl of in vitro RdRp reaction mixture were incubated with Strepavidin-beads (MagneSphere Magnetic Separation Products, Promega) at 25°C for 10 min to capture the biotin-labeled RNA and the RNA-bound p92-Δ167N RdRp as well. Then, we washed the beads once with 0.1% SSC buffer, followed by incubation of the beads with GST-Ded1p or GST (as a control) in RdRp buffer (10 mM DTT, 50 mM Tris–Cl pH 8.0, 10 mM MgCl2) with 1 mM ATP at 25°C for 10 min to elute (release) the p92-Δ167N RdRp from the streptavidin-bound RNA. Then, we collected and precipitated the eluted proteins with 10% TCA. The precipitated proteins (eluate fraction in Fig. 5C) were dissolved in 30 μl SDS buffer. We also recovered the p92-Δ167N RdRp from the streptavidin-bound RNA by boiling the beads in 30 μl SDS buffer for 5 min (SDS fraction in Fig. 5C). All the protein samples were analyzed by Western blotting method with anti-MBP antibody to detect the amount of p92-Δ167N RdRp in the obtained samples.
Cultures of Agrobacterium tumefaciens C58C1 strain carrying pGD-RH2 or pGD-RH20 were used for transient expression of Arabidopsis thaliana RH2 and RH20 [54,61]. N. benthamiana plants were infiltrated with A. tumefaciens carrying pGD-RH2 (OD600 = 0.3) or pGD-RH20 (OD600 = 0.3) together with pGD-CNV (OD600 = 0.3) and pGD-DI-AU-FP (OD600 = 0.3) to launch tombusvirus replication and induce RNA recombination. Leaf infiltration with A. tumefaciens carrying “empty” pGD plasmid was used as a control. We also performed agroinfiltration with pGD-p33 + pGD-p92+ pGD-DI-ΔRI at a final concentration of 0.3 (OD600). Three and four days after agro-infiltration, samples from the agro-infiltrated leaves were collected from the agroinfiltrated leaves. Total RNA was extracted and Northern blot analysis was performed as previously described [16].
|
10.1371/journal.pntd.0002226 | Dual Targeting of Insulin and Venus Kinase Receptors of Schistosoma mansoni for Novel Anti-schistosome Therapy | Chemotherapy of schistosomiasis relies on a single drug, Praziquantel (PZQ) and mass-use of this compound has led to emergence of resistant strains of Schistosoma mansoni, therefore pointing out the necessity to find alternative drugs. Through their essential functions in development and metabolism, receptor tyrosine kinases (RTK) could represent valuable drug targets for novel anti-schistosome chemotherapies. Taking advantage of the similarity between the catalytic domains of S. mansoni insulin receptors (SmIR1 and SmIR2) and Venus Kinase Receptors (SmVKR1 and SmVKR2), we studied the possibility to fight schistosomes by targeting simultaneously the four receptors with a single drug.
Several commercial RTK inhibitors were tested for their potential to inhibit the kinase activities of SmIR1, SmIR2, SmVKR1 and SmVKR2 intracellular domains (ICD) expressed in Xenopus oocytes. We measured the inhibitory effect of chemicals on meiosis resumption induced by the active ICD of the schistosome kinases in oocytes. The IR inhibitor, tyrphostin AG1024, was the most potent inhibitory compound towards SmIR and SmVKR kinases. In vitro studies then allowed us to show that AG1024 affected the viability of both schistosomula and adult worms of S. mansoni. At micromolar doses, AG1024 induced apoptosis and caused schistosomula death in a dose-dependent manner. In adult worms, AG1024 provoked alterations of reproductive organs, as observed by confocal laser scanner microscopy. With 5 µM AG1024, parasites were no more feeding and laying eggs, and they died within 48 h with 10 µM.
IRs and VKRs are essential in S. mansoni for key biological processes including glucose uptake, metabolism and reproduction. Our results demonstrate that inhibiting the kinase potential and function of these receptors by a single chemical compound AG1024 at low concentrations, leads to death of schistosomula and adult worms. Thus, AG1024 represents a valuable hit compound for further design of anti-kinase drugs applicable to anti-schistosome chemotherapy.
| Schistosomiasis is a chronic, debilitating disease that affects over 200 million people in the world. The pathology of schistosomiasis is caused mainly by host immune responses to parasite eggs and due to the formation of granulomas in liver and other tissues. There is no vaccine for schistosomiasis and treatment relies essentially on a single drug, Praziquantel. However, reduced susceptibility of schistosome isolates to Praziquantel has been reported, raising serious concerns about the need to develop new drugs against schistosomes. Receptor tyrosine kinases (RTKs) control many cellular and developmental processes and they are important targets in cancer therapy. In this paper, we have investigated the possibility to fight schistosomes by targeting with a single drug, insulin receptors (IRs) involved in parasite growth and metabolism and Venus Kinase Receptors (VKRs) which are unusual IR-like RTKs expressed in the parasite reproductive organs of Schistosoma mansoni. Diverse RTK inhibitors have been tested on kinase activities of these RTKs. The well-known IR inhibitor, tyrphostin AG1024, was demonstrated to be a potent inhibitor of both S. mansoni VKRs and IRs, able to induce in vitro death of larvae and adult worms at micromolar doses. AG1024 could represent a good hit compound for the development of novel drugs against schistosomes.
| Schistosomiasis is the second important parasitic disease in the world. This water-borne disease occurs in over 70 tropical and subtropical countries, mainly in sub-Saharan Africa, with 200 million individuals infected, and a number of deaths estimated to be more than 200 thousands annually [1], [2]. Diverse programs aimed at a reduction of parasite transmission, including the control of vector snail populations or the improvement of sanitation conditions and water supplies, but mass treatment of human populations by chemotherapy remains the most efficient way to combat schistosomiasis [3]. Treatment relies essentially on the use of Praziquantel (PZQ), a safe and affordable drug effective against the three major human schistosome species and recommended by WHO to reduce morbidity and mortality caused by this disease. However, massive administration of PZQ in endemic areas, and the necessity to reiterate treatments because of the ineffectiveness of the drug towards immature parasites, have raised serious concerns regarding the development of parasite resistance to PZQ [4]. Therefore, intensive efforts have been made in recent years to identify novel schistosome molecular targets for chemotherapy [5] and protein Tyrosine Kinases (TKs) have been considered as good candidates because of their essential roles in development and metabolism [6]–[8]. Receptor tyrosine kinases (RTKs) regulate many cellular activities such as proliferation, migration or differentiation, and they are the major TK signalling protagonists, being able to integrate perception, response to extracellular signals and propagation by phosphorylation of intracellular targets [9]. Cancers are often associated with deregulation of RTK activity, and these receptors, such as Epidermal Growth Factor receptor HER-2, c-Kit or VEGF-R, constitute pertinent chemotherapeutical targets in diverse anti-cancer therapies [10]–[12]. Insulin-like Growth Factor 1 (IGF-1) receptor is also commonly overexpressed in cancer and its activation affects cell proliferation, adhesion, migration and cell death [13]. Blocking IGF-1R prevents tumor cell growth and increases apoptosis in malignant cells [14]. Moreover, insulin receptor (IR), closely related to IGF-1R, is also overexpressed in many cancers. Its activation has been shown to compensate IGF-1R inhibition in malignant cells, thus validating the interest of co-targeting IGF-1R and IR in cancer [15].
IR/IGFR molecules are conserved in a large variety of eumetazoan species, from sponges to mammals [16]. Two receptors of the IR family, SmIR1 and SmIR2, have been characterized in Schistosoma mansoni, and display differences in their tissue localization. SmIR-1 is expressed in muscles, intestinal epithelial cells and at the basal membrane of the tegument [17], colocalized with SGTP1 and SGTP4 schistosome glucose transporters [18]. SmIR2 is massively expressed in parenchymal cells of adult schistosomes, suggesting that the two receptors could have distinct functions [17]. Two IR members have been also found in Schistosoma japonicum (SjIR1, SjIR2) which are highly similar to SmIR1 and SmIR2 respectively [19]. In both schistosome species, these receptors might have conserved IR function in the regulation of glucose uptake, since treatment by IR specific inhibitors affect significantly glucose entry in parasites [19], [20].
Two additional RTKs (SmVKR1 and SmVKR2) with intracellular kinase domains similar to that of SmIRs were also characterised in S. mansoni. They were named VKR for Venus Kinase Receptor since they contain in their extracellular part an atypical Venus FlyTrap (VFT) motif usually found in G-protein-coupled receptors of class C. SmVKRs are members of a novel family of RTKs discovered few years ago [21], present only in invertebrates and activable by amino-acids [22], [23]. These receptors are highly expressed in larval stages of the parasite as well as in ovaries of female worms, suggesting functions in development and reproduction [23]. Considering the potential importance of SmIRs and SmVKRs in development, but also in metabolism and reproduction, the striking similarity observed between the catalytic domains of the four receptors led us to postulate that targeting simultaneously these four effectors by a single compound would be highly detrimental for the parasites and might possibly represent a novel multiple target strategy against schistosomes.
Here, we analyzed the potential of several IR and RTK inhibitors to inhibit kinase activities of both SmIR and SmVKR kinase domains recombinantly expressed in Xenopus oocytes. Among the different compounds tested, tyrphostin AG1024 emerged as the most potent inhibitor towards the four receptors. In vitro experiments then demonstrated that treatment with AG1024 led to dramatic effects on the viability of larval and adult schistosomes as well as on the fertility of adult worms.
All experiments involving hamsters within this study have been performed in accordance with the European Convention for the Protection of Vertebrate Animals used for Experimental and other Scientific Purposes (ETS No 123; revised Appendix A) and have been approved by the committee for ethics in animal experimentation of the region Nord Pas de Calais France (authorisation No. AF/2009) in the local animal house of the Pasteur Institute of Lille (Agreement No. A59-35009).
A Puerto-Rican strain of S. mansoni was maintained by passage through albino Biomphalaria glabrata snails and Mesocricetus auratus golden hamsters. Adult schistosomes were collected by portal perfusion from infected hamsters at 42–45 days p.i. Schistosomula were prepared as described previously [24].
Intracellular domains (ICD) of SmIR1, SmIR2, SmVKR1 and SmVKR2 were amplified by PCR from pcDNA3.1 plasmids encoding the receptor full-length sequences [17], [21], [23] using Fwd 5′-CCggatccAACGGAGAATTTCACGGAAACGTCTGCAG-3′ and Rev 5′-CCctgcagTCAAATATATAAGGAAGAAGATGTGAATG-3′ primers with BamH1 and Pst1 sites for SmIR1 ICD; Fwd 5′-CCgaattcCGTCGTTATTATTTAAAGGTTACAGCTTGG-3′ and Rev 5′-CCggatccTTATGCGATAACGTTTCTAGTTCTACTTAG-3′ primers with EcoRI and BamH1 sites for SmIR2 ICD; Fwd 5′-GGgaattcGTCAACCATATGAAAACCTTTG-3′ and Rev 5′-CCctgcagTCAAGGTAGAAACGCTAAACTGTTATC-3′ primers with EcoR1 and Pst1 sites for SmVKR1 ICD; Fwd 5′-AATggatccTAAACGGTCTTCCTACCGGAAAG-3′ and Rev 5′-CCccatggCGACGTAAACTGAAAGAAATTGAAAATCG-3′ primers with BamH1 and Nco1 sites for SmVKR2 ICD. PCR products were inserted into pCR2.1 TOPO vector (Invitrogen) before cloning in phase with the myc epitope tag in pGBKT7 expression vector (Clontech). pGBKT7 plasmids containing wild-type SmIR1WT, SmIR2WT, SmVKR1WT and SmVKR2WT ICDs were further submitted to site-directed mutagenesis in order to render the kinase protein domains constitutively active. For this, the second amino-acid next to the conserved YY autophosphorylation site contained in SmIR and SmVKR kinase domains was replaced by a glutamic residue according to the procedure already described [22]. Constitutively active mutants of SmIRs were obtained using 5′-CGTCTTGTAAATAATCAAGAATATTATAGAgAAATTGGACAAGC-3′ mutated sequence and its reverse complement to generate SmIR1YYRE and 5′-CAGATGTTTATGGACATAATTATTATCACgAAACAAGTCATGC-3′ and reverse complement sequences for the SmIR2YYHE mutant. SmVKR1YYRE and SmVKR2YYRE constructs were obtained as described in [23].
cRNA encoding wild-type or constitutively active mutants of receptor ICDs were produced using the T7 mMessage mMachine Kit (Ambion, USA). cRNAs were transcribed from T7 promoter-containing pGBKT7 plasmids (1 µg) previously linearised by HindIII restriction enzyme. cRNAs were precipitated by 2.5 M LiCl, washed in 70% ethanol, resuspended in 20 µl diethylpyrocarbonate (DEPC)-treated water, and quantified by spectrophotometry. cRNAs were analysed in a denaturating agarose gel. Gel staining with 10 µg ml−1 ethidium bromide allowed to confirm correct sizes and of absence of abortive transcripts. cRNA preparations (1 mg ml−1) were microinjected in Xenopus oocytes (stage VI) according to the protocol previously described [25]. Each oocyte was injected with 60 nl of cRNA in the equatorial region and incubated at 19°C in ND96 medium (96 mM NaCl, 2 mM KCl, 1 mM MgCl2, 1.8 mM CaCl2, 5 mM HEPES pH 7.4 supplemented with 50 µg/ml streptomycin/penicillin, 225 µg/ml sodium pyruvate, 30 µg/ml trypsin inhibitor) for 18 h. For inhibitor treatments, oocytes were incubated with tyrphostin AG538, AG1024, AG1478, HNMPA-(AM)3 (Santa Cruz Biotechnology), SU11274 or BIBF1120 (Selleckchem) at different concentrations. In all cases, germinal vesicle breakdown (GVBD) was detected by the appearance of a white spot at the apex of the cell, a witness of oocyte progression from G2 to M phase of the cell cycle.
Immunoprecipitation of myc-tagged ICD proteins expressed in oocytes was performed according to the procedure described previously [25]. Following 18 h of cRNA injection, oocytes were lysed in buffer (50 mM HEPES pH 7.4, 500 mM NaCl, 0.05% SDS, 0.5% Triton X100, 5 mM MgCl2, 1 mg/ml bovine serum albumin, 10 µg/ml leupeptin, 10 µg/ml aprotinin, 10 µg/ml soybean trypsin inhibitor, 10 µg/ml benzamidine, 1 mM sodium vanadate) and centrifuged at 4°C for 15 min at 10,000 g. Supernatants were incubated with anti-Myc (1/100; Invitrogen) antibodies for 4 h at 4°C. Protein A-Sepharose beads (5 mg, Amersham Biosciences) were added for 1 h at 4°C. Immune complexes were collected by centrifugation, rinsed three times, resuspended in Laemmli sample buffer, and subjected to a 10% SDS-PAGE. Immune complexes were analyzed by Western blotting using anti-myc (1/50,000) or pY-20 (1∶10,000; anti-phosphotyrosine, BD Biosciences) antibodies and the advanced ECL detection system (Amersham Biosciences).
500 schistosomula were incubated for 7 days in 24-well plates containing 2 ml of M199 medium (Invitrogen) supplemented with HEPES 10 mM, pH 7,4, antibiotic/antimycotic mixture (Sigma, 1.25%) and FCS (10% Gibco) (referred as M199 complete medium) with different concentrations (from 1 to 50 µM) of tyrphostin AG1024 (Santacruz Biotechnology) dissolved in DMSO. Culture medium was refreshed daily. Parasite mortality was assessed by eye each day using three criteria: absence of motility, tegument defects and granular appearance. A minimum of 300 larvae was observed for each condition, and the ratio dead larvae/total larvae calculated in three independent experiments.
Twenty adult paired couples of S. mansoni were incubated at 37°C in a 5% CO2 atmosphere in 10 ml M199 complete medium in the presence of tyrphostin AG1024 at different concentrations (from 1 to 10 µM) for 5 days. Culture medium was refreshed daily. The number of paired couples was estimated every day by stereomicroscopy. In each well, medium containing the eggs was harvested every day, and fractions were then pooled and centrifuged. Total number of eggs was determined from three independent countings.
Apoptosis was detected using the Terminal deoxynucleotidyl transferase dUTP nick end labeling (TUNEL) method and the In Situ Cell death detection kit (Roche). Briefly, 2,000 schistosomula were incubated for 48 h in 6-well plates containing 2 ml of M199 complete medium without or with 10 or 50 µM AG1024, then fixed in formaldehyde 2%. Labeling of schistosomula with DAPI and TMR red-dUTP was performed according to manufacturer's instructions and TUNEL-positive parasites were observed by fluorescence using an AxioImager Z1-Apotome microscope (Zeiss).
After 5 days of culture, worms were fixed for at least 24 h in AFA (ethanol 95%, formalin 3% and glacial acetic acid 2%), stained for 30 min with 2.5% hydrochloric carmine (Certistain, Merck), and destained in acidic 70% ethanol. Following dehydration in 70%, 90% and 100% ethanol, 1 min each, worms were preserved as whole-mounts in Canada balsam (Merck) on glass slides [26], [27]. The morphology of the reproductive organs of parasites was observed using a Confocal Laser Scanning Microscope (CLSM) Leica TCS SP2 microscope, with a 488 nm He/Ne laser and a 470 nm long-pass-filter under reflection mode.
Intracellular domains (ICD) of SmIRs and SmVKRs were amplified and cloned into the pGBKT7 vector which contains the T7 promoter sequence required for in vitro transcription. The expression of myc-tagged proteins of SmIR and SmVKR ICDs was obtained following injection of their respective cRNAs in oocytes. Proteins could be detected by western blot analysis of oocyte lysates with anti-myc antibodies. Myc-tagged proteins were detected at molecular weight of 41 kDa for SmIR1, 69 kDa for SmIR2, 68 kDa for SmVKR1 and 81 kDa for SmVKR2 constructs (Figure 1). Several studies have demonstrated that the Xenopus oocyte is a suitable model for expressing S. mansoni proteins and particularly for studying phosphorylating activity of protein kinases [23], [25], [28]–[32]. In oocytes, which are giant cells naturally blocked in prophase I of meiosis I, the kinase activity of any exogenous recombinant kinase is able to trigger resumption of meiosis and passage into metaphase II, following germinal vesicle breakdown (GVBD), a process easily detected by the appearance of a white spot at the animal pole of the oocyte. In order to analyze receptor kinase activities, we prepared constitutively active mutants by site-directed mutagenesis. In YYxE mutants, a glutamic (E) residue was introduced near the YY autophosphorylation site, in order to mimic its phosphorylation and to induce spontaneous kinase activation.
Results shown in Figure 1 indicated that SmIR1YYRE, SmIR2YYHE, SmVKR1YYRE and SmVKR1YYRE mutant proteins were effectively recognized by anti-phosphotyrosine antibodies, confirming their potential to autophosphorylate and thus demonstrating their constitutive kinase activity. As expected, only the oocytes expressing constitutively active kinases, but not the wild-type ones, underwent GVBD. The number of oocytes undergoing GVBD could be used in following tests as an indicator of the kinase activity of ICD proteins.
SmIR1YYRE, SmIR2YYHE, SmVKR1YYRE and SmVKR1YYRE ICDs were expressed in Xenopus oocytes and we tested the capacity of several TK inhibitors to inhibit their potential to induce GVBD in oocytes. As the kinase domains of SmIR and SmVKR proteins were previously shown to be highly similar to those of insulin receptors [21]–[23], we analysed the effect of three well-known IR and/or IGFR inhibitors (tyrphostins AG538 and AG1024, HNMPA-(AM)3) on schistosome receptor ICD kinase activity. Tyrphostin AG1478 (EGFR inhibitor), SU14278 (Met receptor inhibitor) and BIBF1120 (FGFR inhibitor) were tested in parallel at different concentrations (Figure 2A). First results showed that both SmIR and SmVKR kinases were sensitive to IR/IGFR inhibitors and that among these three compounds, AG1024 was the most effective, able to inhibit at 100% GVBD in oocytes expressing SmVKR1, SmVKR2, SmIR1 at a 0.1 µM dose, and SmIR2 at a 1 µM dose. Complete inhibition of GVBD induced by ICDs was obtained with the IGFR inhibitor AG538 at 1 µM, except in SmVKR1-expressing oocytes in which 0.1 µM AG538 was sufficient to totally inhibit the activity. The effectiveness of HNMPA-(AM)3 was similar to that of AG1024 on SmVKR1 and SmVKR2 (0.1 µM) but this drug was less effective on SmIR1 and SmIR2, that required respectively minimal doses of 1 and 10 µM to be inhibited. Surprisingly, AG1478, a potent inhibitor of EGFR, was effective on SmVKR1 and SmVKR2 at low doses (0.1 µM) whereas its action on SmIR1 and SmIR2 was relatively weak (10 µM needed to inhibit 100% GVBD). As expected, the Met kinase inhibitor SU14278 had no detectable activity on SmVKRs and SmIR1 (only 40% inhibition of GVBD in SmIR2-oocytes at 100 µM) and FGFR inhibitor BIBF1120 was also inactive on the four schistosome kinases. From these data, we concluded that AG1024 was the most potent drug to inhibit both SmIRs and SmVKRs, since it blocked completely the activity of SmIR1, SmVKR1 and SmVKR2 ICDs at a 100 nM concentration, and SmIR2 kinase activity at 1 µM. Western blot results (Figure 3) confirmed that inhibition of GVBD in the presence of the drug was associated with an absence of tyrosine phosphorylation and kinase activation for each ICD.
In order to investigate the effect of AG1024 on the viability of S. mansoni larvae, 24 h-old schistosomula were cultured in vitro for 5 days with different concentrations of AG1024, with daily renewal of drug-containing medium. Parasite death was assessed by eye, following three criteria: loss of motility, tegument alterations and granular aspect (Figure 4B). We observed that AG1024 treatment led to parasite death in a time and dose-dependent manner. Indeed, 50 µM of AG1024 induced 100% of parasite death within 48 h, whereas five days were required with 20 µM. Treatment of schistosomula with 1 and 10 µM AG1024 much lower affected parasite viability (with 15 and 30% mortality within 5 days respectively) (Figure 4A).
Since AG1024 is known to trigger apoptosis in human cell lines [33], we evaluated the occurrence of apoptosis-induced death in schistosomula using a TUNEL assay. In these experiments, schistosomula were treated with 10 or 50 µM AG1024 for 48 h, then fixed and stained with DAPI and TUNEL (Figure 5). Results indicated that AG1024 could induce apoptosis in schistosomula in a dose-dependent manner. Taken together, these results strongly suggest that AG1024 could lead to schistosomula death by inducing apoptotic signals, through inhibition of SmIR and SmVKR kinases.
The effect of AG1024 was also studied on adult worms in vitro. In these experiments, S. mansoni couples were cultured with different concentrations of AG1024, and we monitored pairing behaviour and egg production for each condition during 5 days. Results showed that drug treatment had drastic effects on parasite fitness and egg production. Indeed, 1 µM AG1024 was affecting the stability of worm pairing, showing only 30% of couples still paired after 5 days (Figure 6A) and 30% decrease of egg laying (Figure 6B). Striking effects of AG1024 on schistosomes were registered at 5 µM, a dose at which worms were no more paired and egg laying almost stopped at day 2. At this time point, worms also suffered of tetany and were drifting as a consequence of their inability to stick to well bottom walls. Gut peristalsis stopped after 5 days, suggesting that SmIRs and/or SmVKRs may also regulate functions in gastrodermis and/or smooth muscles. Finally, at higher concentrations (from 10 to 50 µM), AG1024 induced adult worm death within a 2 to 5 day period (not shown).
To complement these observations, we examined AG1024-treated adult worms by confocal laser scanning microscopy. Whereas no significant phenotype could be detected in gonads of adult worms treated with 1 µM of AG1024 (Figure 7 C, D), major changes occurred in worms treated with 5 µM (E, F). In females, we observed important size reduction and disorganization of the ovary, which in normal parasites (A) contains small immature oocytes in its anterior part and large mature oocytes in its posterior part. In AG1024-treated worms, immature cells were less abundant and mature cells seemed to invade the whole ovary. Focus on the ootype of female worms (Figure 8) further indicated that the drug significantly affected at 1 µM the composite structure of the egg formed in the ootype (B), inhibiting totally its formation when used at 5 µM (C and D). Moreover, we could note in treated parasites, a significant atrophy of Mehlis' glands, the cells that line the ootype [27].
In males, we could observe main changes inside of the sperm vesicle, that was full of undifferentiated cells in worms treated with 5 µM AG1024 (Figure 7 F), whereas a population of elongated spermatozoa was visible in control worms (B). These data indicated that AG1024 treatment could also affect spermatogenesis.
Current strategies of intervention against schistosomiasis are based on drug administration of PZQ. Since PZQ neither kills immature schistosomes nor prevents reinfection, PZQ-based control programs provide only a transitory effect on parasite transmission and a limited potential on eradication of the disease. Moreover, concerns have been raised about PZQ resistance that actually emphasize the need for new initiatives in search for alternative antischistosome compounds and in discovery of novel parasite drug targets [4], [5]. Recent studies have convincingly demonstrated the importance of protein kinases in schistosome biology and TKs have been considered as good candidates because of their essential roles in development and metabolism [6]–[8]. Targeting of Src (SmTK3) [34] and Syk (SmTK4) [31] kinases with their respective TK inhibitors herbimycin and piceatannol was shown to have a marked effect on reproduction processes of S. mansoni and the anti-cancer drug Imatinib (STI-571, Gleevec) that targets Abl kinases led to important alterations of parasite gastrodermis and caused the death of parasites in vitro [32], [35].
During the last few years we have demonstrated the peculiar nature of insulin-dependent or insulin-related signalling in schistosomes. First, two distinct IR homologs (SmIR1 and SmIR2) were found in S. mansoni whereas only one single IR is present in most of invertebrate species [17]. Second, two additional receptors (SmVKR1 and SmVKR2) were discovered that contain catalytic IR-like domains, and thus represent alternative candidates able to participate in IR-like pathways [21]–[23]. Such a diversification of the IR family, described for the first time in S. mansoni [17], was confirmed recently in S. japonicum and Clonorchis sinensis [19], [36] (Vanderstraete et al, submitted). This reflects a complex insulin-related network that we could consider as a real “Achille's heel” for these parasite trematodes in terms of targets for chemotherapy. Since schistosome IRs have been shown to participate in parasite metabolism by their regulatory function in glucose uptake in adult parasites [17], [19], [20] while SmVKR1 and SmVKR2 highly expressed in gonads [22], [23] play important functions in development and gametogenesis (M. Vanderstraete et al, to be published), we have investigated the possibility to fight parasites by dual targeting of metabolism and reproductive processes through the inhibition of their four IR-like receptors using a single drug.
In these studies, we have tested several commercial drugs known to inhibit human RTK activity and to be efficient on various cancer cells. Three inhibitors AG1024, AG538, HNMPA-(AM)3, which are specific for IR/IGFR [37]–[39] and whose detrimental effects especially on glucose uptake in adult parasites have been already described [19], [20] were analyzed along with three other compounds known to inhibit either EGFR (AG1478), or Met (SU11274) or FGF-R (BIBF1120). Inhibitory effect of these compounds was analyzed towards SmIR and SmVKR recombinant active kinases produced in Xenopus oocytes, a highly suitable cellular model in which we can directly relate the potential of proteins to induce meiosis resumption to their kinase activity [23], [25], [28]–[32]. Whereas tyrphostins AG1024 and AG538 were active at ≤1 µM on SmIR1 and SmIR2, surprisingly, HNMPA-(AM)3 was active only at ≥1 µM on SmIR1 and at ≥10 µM on SmIR2. The efficacy of the three IR inhibitors, was equal or even better towards SmVKRs than against SmIRs, and AG1024 emerged as the most potent drug, being able to inhibit the four receptors at a dose of ≤1 µM in the kinase assay developed in Xenopus oocyte. Concerning the EGFR inhibitor AG1478, its unexpected effect on SmVKRs at ≤0.1 µM was unexplained. Besides, the lower efficacy of AG1478 on SmIRs (10 µM) suggested structural differences between IR/IR-like catalytic domains of the two receptor classes. The Met receptor inhibitor SU11274 had almost no effect on SmIR and SmVKR kinases and BIBF1120 was not active at <10 µM on any of these kinases, confirming the conserved IR-like structure of SmIR and SmVKR catalytic domains.
From these data, we decided to analyze the effect of the selective inhibitor of IGF-1R, AG1024, on the viability of larval and adult stages of S. mansoni in vitro. We could demonstrate that AG1024 caused death of schistosomula in a dose and time-dependent manner, inducing apoptotic signals in the parasite, similarly to its effect caused on cancer cells [33]. Concerning the adult stage, results indicated that parasite couples, compared to schistosomula, were sensitive to lower amounts of the drug, and showed important loss of fitness and fertility at doses ≤5 µM. These concentrations are lower than those used on MCF7 human breast cancer cells (≥10 µM) to decrease proliferation and cause apoptosis [33], indicating the particular sensitivity of the parasites to the drug.
Since evidence has been given that gonads are important sites for the expression of SmVKR1 and SmVKR2 [22], [23], this tempts to assign in priority the decrease of egg formation and laying consecutive to AG1024 treatment, to the inhibition of SmVKR kinase activities. However, we propose that the concomitant targeting of SmIR receptors could be responsible also for alterations of reproductive processes in AG1024-treated worms. Indeed, You et al have shown that the insulin receptor SjIR2 was located in vitelline cells of S.japonicum females [19] and these authors demonstrated recently that vaccination of mice against the ligand-binding domain of SjIR2 resulted in a significant reduction of faecal eggs and liver granuloma density in infected animals, suggesting the importance of schistosome IR receptors both for nutrition (glucose consumption) and reproduction of parasites [40].
While AG1024 is a specific IGF-1R and IR inhibitor, it has been shown that AG1024 had an additional target in melanoma cells upstream of the Erk2 kinase [41]. We do not exclude that side-effects of the drug on other kinases and especially on other parasite RTKs, could contribute to the toxic effect of the drug on the parasite. Recent experiments in oocytes performed using the recombinant EGF receptor of S. mansoni, SER [25], demonstrated that its kinase activity was also sensitive to AG1024 at 1 µM, thus confirming the potential of this drug in the context of a multi-kinase targeting. Finally, it was shown that AG1024 was, among the tested IR inhibitors, the most toxic one for schistosomes. Tyrphostin AG538 and HNMPA-(AM) 3 had no visible effect on parasite viability in vitro when used at 10 µM during a 5 day culture (results not shown). Considering the large identities that exist between the kinase domains of VKR1 and VKR2 receptors from S.mansoni and Schistosoma haematobium (98% and 99% respectively) as well as between those of the IR1 and IR2 molecules of S. mansoni and S. japonicum (74% and 72% respectively [19]), it is likely that AG1024 would have a similar toxic effect on these three human schistosome species.
In conclusion, our results show that simultaneous inhibition of the functional activity of SmIRs and SmVKRs using a single chemical compound can lead in vitro to the death of both immature and adult stages, which is an attractive feature for an alternative drug to PZQ (whose action on immature worms is defective). Further work is needed to evaluate the potential of AG1024 to kill parasites in infected animals, but these data place this drug already as a good hit for the design of more specific anti-kinase drugs applicable to anti-schistosome chemotherapy.
|
10.1371/journal.ppat.1002671 | Macrophage Activation Associated with Chronic Murine Cytomegalovirus Infection Results in More Severe Experimental Choroidal Neovascularization | The neovascular (wet) form of age-related macular degeneration (AMD) leads to vision loss due to choroidal neovascularization (CNV). Since macrophages are important in CNV development, and cytomegalovirus (CMV)-specific IgG serum titers in patients with wet AMD are elevated, we hypothesized that chronic CMV infection contributes to wet AMD, possibly by pro-angiogenic macrophage activation. This hypothesis was tested using an established mouse model of experimental CNV. At 6 days, 6 weeks, or 12 weeks after infection with murine CMV (MCMV), laser-induced CNV was performed, and CNV severity was determined 4 weeks later by analysis of choroidal flatmounts. Although all MCMV-infected mice exhibited more severe CNV when compared with control mice, the most severe CNV developed in mice with chronic infection, a time when MCMV-specific gene sequences could not be detected within choroidal tissues. Splenic macrophages collected from mice with chronic MCMV infection, however, expressed significantly greater levels of TNF-α, COX-2, MMP-9, and, most significantly, VEGF transcripts by quantitative RT-PCR assay when compared to splenic macrophages from control mice. Direct MCMV infection of monolayers of IC-21 mouse macrophages confirmed significant stimulation of VEGF mRNA and VEGF protein as determined by quantitative RT-PCR assay, ELISA, and immunostaining. Stimulation of VEGF production in vivo and in vitro was sensitive to the antiviral ganciclovir. These studies suggest that chronic CMV infection may serve as a heretofore unrecognized risk factor in the pathogenesis of wet AMD. One mechanism by which chronic CMV infection might promote increased CNV severity is via stimulation of macrophages to make pro-angiogenic factors (VEGF), an outcome that requires active virus replication.
| Neovascular age-related macular degeneration (AMD) is the leading cause of vision loss in the elderly. Onset of AMD is due to local production of vascular endothelial growth factor (VEGF) that promotes formation of new blood vessels in the retina, thereby leading to retinal tissue destruction and blindness. Since a clinical study by us showed that AMD patients have high amounts of antibodies to human cytomegalovirus (HCMV), we postulated that infection with HCMV might be a risk factor for AMD. To investigate this possibility, mice were infected with murine cytomegalovirus (MCMV), and at various times after infection, subjected to laser treatment of the eye to induce choroidal neovascularization, an experimental model of AMD. Most severe CNV developed in mice with chronic MCMV infection, a time when MCMV gene sequences could not be detected within eye tissues. However, splenic macrophages collected from mice with chronic MCMV infection produced high levels of gene transcripts to several pro-angiogenic factors including VEGF. MCMV infection of mouse macrophages in culture also produced high amounts of VEGF. Stimulation of VEGF production in vivo and in vitro was sensitive to antiviral treatment. Chronic HCMV infection may therefore promote AMD by stimulation of VEGF production by activated macrophages.
| Angiogenesis, the formation of blood vessels, plays a critical role in embryonic development, wound healing, and normal physiologic processes associated with natural growth and development. On the other hand, new blood vessel growth (neovascularization) contributes to a number of pathologic conditions that include atherosclerosis and tumor formation [1], [2]. The eye is also particularly sensitive to neovascularization during which abnormal blood vessel growth within retinal or choroidal tissues leads to vision loss or blindness. Sight-threatening diseases of the eye associated with abnormal neovascularization include diabetic retinopathy [3], retinopathy of prematurity [4], and age-related macular degeneration (AMD) [5].
Of these, AMD is the leading cause of severe irreversible central vision loss and legal blindness in individuals 65 years of age or older in the United States and other developed countries [6]–[9]. Since the number of elderly persons will double by 2020, AMD is expected to become a major public health problem. Two forms of AMD are recognized [5]–[10]. The non-neovascular form (also known as “dry” or “nonexudative”) represents an early form of AMD usually associated with little visual acuity loss. It is characterized by atrophic abnormalities of the retinal pigment epithelium (RPE) and drusen, small lesions at the level of the RPE that contain granular and vesicular lipid-rich material. Over time, however, this form of AMD often progresses to the neovascular (also known as “wet” or “exudative”) form of AMD that results in significant vision loss due to the appearance of choroidal neovascularization (CNV). Although the precise events that contribute to the development of AMD remain uncertain, recent studies have implicated various immunological and inflammatory mechanisms. For example, complement deposition has been demonstrated within drusen and the choriocapillaris, and several publications have demonstrated that polymorphisms in complement factor-H are associated with an increased risk of AMD [11]–[14]. Several investigators have also identified macrophages in association with drusen as well as choroidal neovascular membranes [15]–[18] suggesting a role for macrophages in the pathophysiology of both forms of AMD. In support of this hypothesis, we [19], and others [20], have shown in a mouse model of experimental CNV that depletion of macrophages significantly decreases the size and severity of lesions.
Macrophages are immune cells of monocyte origin that are classically associated with innate immune responses, particularly inflammation [21], but they may also exhibit pro-angiogenic as well as anti-angiogenic activities [22]. Thus, macrophages may exist in different activation states [23], and individuals may therefore vary in activation states as defined by expression of cytokine transcripts as well as inducible cytokine production [24]. In fact, phenotypically polarized macrophages have been broadly classified into two main groups: classically activated (M1) macrophages and alternatively activated (M2) macrophages that are further subdivided into three subtypes [25]. Moreover, M1 macrophages exhibit an anti-angiogenic phenotype, whereas M2 macrophages exhibit a pro-angiogenic phenotype [22]–[27]. It is therefore possible that individuals with macrophages of one activation state will have a relative protective effect in AMD while individuals with macrophages of another activation state will be at risk for progressive complications. This idea is supported by our observation that the presence of highly activated macrophages is associated with a 5-fold increase in risk of having wet AMD [24].
The mechanism of macrophage activation is clearly multifactorial involving genetics, systemic health cofactors, and environmental cofactors including infection [15], [28]–[30]. Infectious pathogens have been implicated in several vascular diseases, especially atherosclerosis [29]–[32]. Chlamydia pneumoniae, human cytomegalovirus (HCMV), and Helicobacter pylori all have been implicated in promoting severity of atherosclerosis and inducing complications such as myocardial infarction [31], [33], [34]. These findings prompted us to perform a seroepidemiologic study to investigate a possible association for these infectious pathogens with neovascular AMD, a study that subsequently demonstrated a significant association with high HCMV IgG serum titers [35]. This finding differed from that of Kalayoglu and coworkers [36] whose study suggested an association between chlamydia and neovascular AMD.
Given our clinical findings [35] and our long-standing interests in the immunology and pathogenesis of cytomegalovirus retinal disease [37], we sought to test the hypothesis that chronic infection with HCMV, a common β-herpesvirus that targets myeloid lineage cells that give rise to activated macrophage cell populations in tissues [38], is a heretofore unrecognized risk factor for onset and progression of neovascular AMD. This hypothesis was tested herein using an established mouse model of laser-induced CNV [39] to evaluate the effect of systemic infection by murine cytomegalovirus, a mouse β-herpesvirus whose genomic structure and cellular/tissue tropisms parallels those of HCMV [38], on the severity of CNV lesions during acute and chronic virus infection. We observed that mice with chronic MCMV infection developed more severe CNV, and that macrophages collected from chronically infected animals were activated as determined by expression of high levels of transcripts for a number of pro-inflammatory and pro-angiogenic factors, especially the pro-angiogenic cytokine vascular endothelial growth factor (VEGF). In vitro studies confirmed that MCMV infection of a mouse macrophage cell line resulted in significant upregulation of VEGF mRNA and VEGF protein production. Subsequent in vivo and in vitro studies using the antiviral ganciclovir demonstrated that increased production of VEGF by splenic macrophages collected from chronically infected mice and by MCMV-infected mouse macrophages grown in culture was ganciclovir-sensitive, findings that suggest that active virus replication is indeed required for stimulation of VEGF production by macrophages.
We first explored the effect of systemic MCMV infection on experimental CNV. Three times relative to systemic virus inoculation were chosen for this study, one at time of acute systemic infection (6 days postinfection) and two at times of chronic systemic infection (6 weeks postinfection and 12 weeks postinfection) [38]. Groups of C57BL/6 mice were inoculated intraperitoneally with a sublethal dose of MCMV, and their eyes were subjected to laser-induced CNV [39] at 6 days, 6 weeks, or 12 weeks after systemic MCMV infection. In this study, control mice received UV-inactivated MCMV. Four weeks after laser treatment, propidium iodide-stained flatmounts of the posterior pole were prepared of all laser-treated eyes, and groups were compared for severity of CNV. Results are shown in Figure 1A–D and Figure 2A. As expected, mice inoculated with UV-inactivated MCMV exhibited small CNV lesions (1.8±0.1 disc areas). In comparison, mice inoculated with infectious MCMV exhibited CNV lesions of increased size. Lesion size also increased with time after MCMV infection. Whereas mice with MCMV infection of 6-days duration exhibited CNV lesions of moderate enlargement (2.7±0.2 disc areas) four weeks after laser treatment, progressively larger lesions were observed in mice with MCMV infection of 6-weeks (3.1±0.2 disc areas) and 12-weeks (4.4±0.6 disc areas) duration prior to laser treatment (Figure 2A). A statistical comparison of lesion sizes observed in animals with MCMV infection of 12-weeks duration versus control animals revealed significance (p = <0.0001). The frequency of large lesions also increased with progression of MCMV. Whereas only 10% of the total number of lesions in mice inoculated with UV-inactivated virus exceeded 2.2 disc areas (Figure 2B) (representing the 95% confidence interval for lesion size in control mice), 57.5, 92, and 100% of animals with systemic MCMV infection of 6-days, 6-weeks, and 12-weeks duration prior to laser treatment, respectively, developed large CNV lesions. Similar findings were observed when flatmounts were evaluated for degree of vascular size (Figure 2C) and vascularity (data not shown), although cellular density remained constant (Figure 2D). Taken together, these results suggest that systemic MCMV infection results in more severe CNV in mice, even during acute infection where a trend in increased severity is also observed. The most severe and statistically significant of CNV lesions, however, is found in mice with chronic MCMV infection of 12-weeks duration.
Histopathologic analysis of CNV lesions (Figure 3A) paralleled those of flatmount findings. When compared with mice that received UV-inactivated virus (Figure 3B), mice with laser-induced CNV at 12 weeks after infection demonstrated a near doubling of CNV surface area (64,977±7,267 pixels2 versus 119,149±8,578 pixels2; p = <0.0004). Importantly, neither active nor chronic systemic MCMV infection changed the typical morphological appearance of experimental CNV. There was an absolute absence of MCMV-induced cytopathology as well as retinal necrosis.
Since systemic MCMV infection was found to induce more severe CNV, we explored the possibility that direct virus infection of choroidal tissues might be responsible for this outcome. Choroidal tissues as well as several key tissues and cell populations known to be associated with MCMV pathogenesis [38], [40] were sampled for detection of MCMV DNA using primers for virus-specific immediate-early 1 (IE1) and glycoprotein H (gH) gene sequences in PCR assays. As expected, samples of spleen tissue, lung tissue (Figure 4), and salivary gland tissue as well as splenic macrophages collected from animals at time of acute MCMV infection (6 days postinfection) or chronic MCMV infection (12 weeks postinfection) provided positive signals for MCMV-specific DNA (Table 1) indicating extensive systemic MCMV infection. Purified CD34+ cells of bone marrow origin collected from mice with acute and chronic MCMV infection were also positive for MCMV DNA. In comparison, choroidal tissues from eyes of acutely infected mice were indeterminant for MCMV-specific DNA, and MCMV-specific DNA sequences could not be detected in choroidal tissues from eyes of chronically infected animals (Table 1). The apparent lack of MCMV infection of choroidal tissues taken from chronically infected animals was confirmed by our inability to recover infectious virus from whole eyes of parallel groups of chronically infected animals individually homogenized and individually inoculated onto MEF monolayers. Thus, no evidence was found for direct MCMV infection of choroidal tissues or subsequent active virus replication within the eye at time of chronic infection when CNV was found to be most severe.
Alternatively, systemic MCMV infection could contribute to increased severity of CNV indirectly via activation of macrophages to produce pro-angiogenic factors. It is well known for both HCMV and MCMV that peripheral blood monocytes are vehicles for systemic dissemination of virus during acute infection [38], [40], and these cells can harbor virus during chronic infection and with the potential to become activated macrophages within various tissues [23]. Since sufficient numbers of macrophages could not be collected from individual eyes of acutely and chronically infected mice for analysis, we subjected enriched populations of splenic F4/80+ macrophages collected from acutely and chronically infected mice to real time RT-PCR assay for detection and quantification of transcripts to several pro-inflammatory and pro-angiogenic cytokines and mediators associated with neovascular AMD. In this study, results were compared with baseline transcript levels established for splenic macrophages collected from mice inoculated with UV-inactivated virus. As shown in Table 2, significant differences were observed in the patterns of synthesis for a number of macrophage-associated transcripts examined during acute and chronic MCMV infection. Of importance was the finding of significant upregulation of VEGF (p = 0.04) and VEGFR1 (p = 0.05) transcripts that progressed from acute to chronic infection. This was associated with a concomitant significant upregulation of matrix metalloproteinase-9 (MMP-9) (p = 0.02), cyclooxygenase-2 (COX-2) (p = 0.05), and tumor necrosis factor alpha (TNF-α) (p = 0.03) transcripts, but only during chronic infection. Interestingly, macrophage-associated VEGFR2 transcript was significantly downregulated (p = 0.01) during acute and chronic infection. These results suggest that an increase in CNV size and severity during chronic MCMV infection may be due to virus-induced activation of macrophages that favor neovascularization.
Although splenic macrophages collected from mice with chronic systemic MCMV infection exhibited an approximate 20-fold increase in VEGF mRNA levels when compared with splenic macrophages collected from control mice, it is possible that increased VEGF mRNA production was not due to active MCMV replication. To explore directly the ability of mouse macrophages to produce increased amounts of VEGF during active virus replication, monolayers of IC-21 mouse macrophages, a macrophage cell line of C57BL/6 origin [41], were either mock-infected (control), treated with LPS (positive control), inoculated with UV-inactivated MCMV (negative control), or inoculated with infectious MCMV at a dose resulting in a low level of infection (2.5 PFU/cell). All monolayers were quantified at 24 hr and 48 hr later for levels of TNF-α mRNA and VEGF mRNA by quantitative RT-PCR assay. Results are shown in Figure 5A. When compared with mock-infected monolayers, monolayers of IC-21 mouse macrophages were activated by LPS treatment as demonstrated by large increases in VEGF mRNA and TNF-α mRNA levels, but parallel monolayers inoculated with UV-inactivated virus produced only low levels of VEGF mRNA and TNF-α mRNA suggesting little-to-no activation. In comparison, MCMV-infected monolayers of IC-21 mouse macrophages at 24 hr postinfection showed a 13-fold increase in VEGF mRNA levels, but interestingly failed to duplicate an increase in TNF-α mRNA production as seen in LPS-treated MCMV-infected monolayers. The same pattern of cytokine mRNA synthesis was observed in MCMV-infected IC-21 mouse macrophages at 48 hr postinfection. At this time after virus infection, VEGF mRNA levels were >50-fold greater than levels found in mock-infected monolayers (p = <0.04), but TNF-α mRNA levels were only ∼3-fold greater. This pattern of activation is consistent with a M2 phenotype of macrophage activation [25] since further analysis of MCMV-infected IC-21 macrophages when compared with mock-infected cells revealed increased levels of IL-10 and IL-1RA mRNA levels, equivalent levels of IL-23 mRNA production, and no detectable IL-21 mRNA production (data not shown). Confirmation that MCMV infection of IC-21 mouse macrophages resulted not only in a significant increase in VEGF mRNA levels, but also in a significant increase in VEGF protein, was provided by ELISA analysis of supernatants collected at 48 hr postinfection (p = 0.01) (Figure 5B). Taken together, these results provide proof-of-principal that the increase in VEGF mRNA levels observed in mice with chronic systemic infection could arise from direct MCMV infection, active virus replication, and subsequent macrophage activation associated with the M2 phenotype, a pro-angiogenic phenotype [25].
To further explore VEGF production by MCMV-infected mouse macrophages in culture, monolayers of IC-21 mouse macrophages were either MCMV-infected (2.5 PFU/cell) or mock-infected and subjected to immunostaining analysis for detection of VEGF production and for quantification of VEGF-positive cells at 24 hr and 48 hr postinfection. Results are shown in Figure 6. When compared with MCMV-infected and mock-infected cells reacted with control antibody, MCMV-infected cells reacted with anti-VEGF antibody at 24 hr and 48 hr postinfection exhibited positive cytoplasmic staining for VEGF. Whereas staining was generally stronger in MCMV-infected cells at 48 hr postinfection when compared with MCMV-infected cells at 24 hr postinfection, the strongest staining was observed in foci of MCMV-infected cells at 48 hr postinfection showing early stages of cytopathology during plaque formation. It is noteworthy that individual macrophages at 48 hr postinfection not involved in plaque formation were also VEGF positive. Quantification studies revealed that ∼55% and ∼93% of MCMV-infected IC-21 mouse macrophages exhibited positive staining for VEGF at 24 hr and 48 hr postinfection, respectively, whereas mock-infected controls showed background levels of VEGF production of ∼10%.
We found in studies described above that splenic macrophages collected from acutely infected mice and chronically infected mice produced significantly more VEGF mRNA when compared with splenic macrophages collected from control mice (Table 2). In addition, monolayers of MCMV-infected IC-21 mouse macrophages produced significantly more VEGF mRNA and VEGF protein when compared with monolayers of mock-infected cells (Figures 5–6). If stimulation of VEGF production in vivo and in vitro is induced directly by active virus replication, we hypothesized that stimulation of VEGF production should be sensitive to treatment with ganciclovir, an antiviral that inhibits HCMV and MCMV replication at the level of virus DNA synthesis [42], [43]. To test this hypothesis in vivo, a study was performed in which groups of C57BL/6 mice were either inoculated intraperitoneally with a sublethal dose of MCMV or mock-infected with maintenance medium (control). Unlike the study summarized in Table 2, it is noteworthy that this study did not use inoculation with UV-inactivated virus as a control. At 12 weeks postinfection, groups of MCMV-infected mice or mock-infected mice were treated intraperitoneally with ganciclovir (40 mg/kg/day) for 7 days [43]. Parallel groups of untreated control MCMV-infected mice or mock-infected mice were not treated with ganciclovir, but instead received daily intraperitoneal injections of phosphate-buffered saline for 7 days. Following the 7-day regimen of ganciclovir or phosphate-buffered saline treatment, splenic macrophages were collected from ganciclovir-treated and untreated chronically infected mice and compared by quantitative real time RT-PCR assay for levels of VEGF mRNA and TNF-α mRNA. In agreement with our previous study summarized in Table 2, splenic macrophages collected from untreated chronically infected mice showed dramatic stimulation of VEGF mRNA production as well as TNF-α mRNA production (Figure 7). In fact, the degree of stimulation for both VEGF mRNA and TNF-α mRNA production was greater than that observed in our previous study (Table 2), especially with respect to TNF-α mRNA production. This difference might be due to the different controls used in the two separate studies, UV-inactivated virus (Table 2) versus maintenance medium (Figure 7). When compared with untreated virus-infected animals, however, ganciclovir treatment resulted in a significant inhibition of VEGF mRNA production (p = ≤0.009), specifically an approximate 44-fold decrease in VEGF mRNA production. A similar degree of inhibition of TNF-α mRNA production was also observed in the presence of ganciclovir treatment (p = ≤0.009). Importantly, this significant inhibition of VEGF mRNA and TNF-α mRNA production in ganciclovir-treated animals could not be attributed to drug-related toxicity since splenic macrophages collected from these animals were found to be >95% viable at time of enrichment and just prior to RT-PCR assay when analyzed by the trypan blue exclusion and MTS assays (data not shown).
An in vitro study was performed to confirm our in vivo ganciclovir treatment findings. Monolayers of IC-21 mouse macrophages were inoculated with either a low dose of MCMV (2.5 PFU per cell) or mock-infected, and all monolayers were treated at 1 hour postinfection with either 0, 15, 30, or 60 uM of ganciclovir. At 24 hr postinfection, all monolayers were harvested and subjected to quantitative RT-PCR assay for comparison of VEGF mRNA levels. In agreement with in vivo ganciclovir treatment findings, increasing amounts of the antiviral reduced in a relatively dose-dependent manner the amounts of VEGF mRNA produced when compared with untreated MCMV-infected mouse macrophages (Figure 8A). As expected, untreated MCMV-infected mouse macrophages produced VEGF mRNA at increased levels, and at levels equivalent to that observed for MCMV-infected macrophages at 24 hr postinfection as shown in Figure 5A. With increasing doses of ganciclovir, however, amounts of VEGF mRNA were dampened, ultimately being reduced by ∼5-fold at the highest doses of ganciclovir, 30 and 60 uM. This reduction in VEGF mRNA production could not be attributed to drug-induced toxicity since mock-infected ganciclovir-treated IC-21 mouse macrophages remained >95% viable at all doses tested when subjected to the trypan blue exclusion and MTS assays (data not shown).
Since ganciclovir treatment appeared to reduce, but not eliminate, VEGF mRNA production by MCMV-infected IC-21 mouse macrophages in a relatively dose-dependent manner, we sought to determine if VEGF could be detected within ganciclovir-treated MCMV-infected IC-21 mouse macrophages, albeit at reduced levels, with increasing doses of drug. We therefore performed an immunostaining study to visualize VEGF production within monolayers of MCMV-infected IC-21 mouse macrophages at 24 hr postinfection following treatment with 0, 15, 30, or 60 uM of ganciclovir at 1 hr after virus inoculation. Results are shown in Figure 8B. In agreement with previous findings (Figure 6), MCMV-infected IC-21 mouse macrophages not treated with drug exhibited prominent cytoplasmic staining for VEGF. In comparison, increasing doses of ganciclovir treatment appeared to dampen VEGF protein production within the MCMV-infected cells. Nonetheless, positive staining for VEGF could still be detected within MCMV-infected IC-21 mouse macrophages treated with the highest dose of ganciclovir, 60 uM. Only faint background staining or no detectable staining was observed in parallel control monolayers of mouse macrophages that were either not infected with virus or virus-infected and reacted with control antibody (data not shown). Taken together, these in vitro findings suggest that VEGF mRNA and VEGF protein production during MCMV infection of IC-21 mouse macrophages are indeed ganciclovir-sensitive, although VEGF production is not completely eliminated in the presence of the antiviral. Moreover, the significant reduction of VEGF mRNA and VEGF protein during ganciclovir treatment of MCMV-infected IC-21 mouse macrophages in culture is in agreement with in vivo findings, thereby supporting the hypothesis that upregulation of VEGF mRNA within splenic macrophages collected from MCMV-infected mice with chronic infection is due to active virus replication.
The number of investigations of angiogenesis in the eye has increased significantly in recent years due to findings that neovascularization of the retina and choroid plays a central role in the development of a number of major blinding diseases. These include AMD as well as diabetic retinopathy, polypoidal choroidal vasculopathy, myopic choroidal neovascularization, neovascular glaucoma, retinopathy of prematurity, and ocular tumorigenesis (all reviewed in [44]). Since a seroepidemiologic clinical study by us demonstrated an apparent association between HCMV infection and neovascular AMD [35], we used an experimental C57BL/6 mouse model of CNV to test the hypothesis that systemic MCMV infection will contribute to the severity of CNV. It has not escaped our attention that mouse strain-dependent factors might play a factor in CNV development during systemic MCMV infection since macrophages from C57BL/6 mice (a prototypical Th1 mouse strain) and macrophages from BALB/c mice (a prototypical Th2 mouse strain) exhibit distinct M1- or M2-dominant responses [45]. Nonetheless, our results collectively showed that systemic MCMV infection of C57BL/6 mice did indeed result in more severe CNV, and, more importantly, chronically infected mice showed the greatest severity of CNV. Although MCMV DNA sequences could not be detected within choroidal tissues of chronically infected animals, splenic macrophages collected from chronically infected animals produced increased amounts of transcripts to several pro-inflammatory and pro-angiogenic cytokines including VEGF. That MCMV infection of mouse macrophages will modulate a pro-angiogenic M2 phenotype that included significant stimulation of VEGF production was shown directly by in vitro studies using a mouse macrophage cell line of C57BL/6 origin. Further evidence that virus infection induced stimulation of VEGF production both in vivo and in vitro was provided by ganciclovir treatment studies that demonstrated sensitivity of VEGF production to the antiviral both in vivo and in vitro. Thus, our findings are novel with respect to chronic eye disease since they provide for the first time new data that suggests that chronic cytomegalovirus infection can contribute to the pathogenesis of wet AMD, possibly via activation of macrophages towards a pro-angiogenic phenotype and stimulation of VEGF production. While we have not yet demonstrated in our model direct visualization of MCMV-infected, VEGF-producing macrophages associated with areas of CNV, several observations would argue that this is a likely occurrence. Firstly, we [19] and others [20] have shown previously that macrophages are essential for development of CNV. Secondly, we have shown previously in the context of MCMV retinitis that IC-21 macrophages infected with a β-galactsidase-expressing LacZ recombinant MCMV will travel to ocular tissues of C57BL/6 mice following tail vein injection [46]. Finally, we show herein that MCMV infection of IC-21 macrophages stimulates VEGF production, a stimulation that is also observed in splenic macrophages collected from chronically infected mice with severe CNV. Future in vivo immunostaining studies will directly address this important issue.
The concept that infectious agents might contribute to the pathogenesis of vascular diseases has become an intense and controversial area of investigation. Two major hypotheses have emerged. One hypothesis proposes that vascular disease is caused by direct infection of the target tissue [36], while the second hypothesis proposes a bystander effect caused by infection at a distant tissue [47]–[49]. In atherosclerosis, direct infection of the atheromatous plaque by Chlamydia pneumoniae has been suggested as a stimulus for recruitment of inflammatory cells. Arguing against this hypothesis, however, are antibiotic treatment trials designed to suppress Chlamydia pneumoniae infection and failing to demonstrate a measurable clinical effect on preventing myocardial infarction or other sequelae [34]. On the other hand, patients with chronic periodontal infection and inflammation have provided evidence suggesting that chronic infection at a distant site may play a role in vascular disease. In this patient population, infection by a variety of different organisms appeared to lead to more severe vascular disease [34], [50], [51]. Since in our study, MCMV-specific DNA sequences could not be detected in choroidal tissues of eyes with the most severe choroidal neovascularization, we propose a similar bystander hypothesis for the role of HCMV infection in chroroidal neovascularization of the eye.
HCMV is a common β-herpesvirus that persists for the life of its host following primary infection. While chronic HCMV infection of healthy, immunologically normal persons was initially thought to have no significant disease consequence, chronic HCMV infection has now been associated with a growing number of long-term diseases that include the vascular disease atherosclerosis, restenosis following angioplasty, transplant vascular sclerosis associated with chronic allograft rejection of solid organ grafts (reviewed in [52]), and possibly tumor formation (reviewed in [53]). Evidence for a link between HCMV and vascular disease was first provided by Melnick, DeBakey, and coworkers [54] when virus antigen was detected within arterial tissues from carotid artery plaques obtained from patients with atherosclerosis. Since this fundamental observation of ∼20 years ago, however, it has been difficult to determine the precise mechanisms by which HCMV might participate in the pathophysiology of vascular disease because the etiologies of chronic diseases are complex and multifactorial. Nonetheless, seropositive HCMV persons are two to three-times more likely to develop coronary artery disease when compared with HCMV seronegative patients [55]. In support of this association are recent findings that 76% of patients with ischemic heart disease have detectable HCMV DNA within their vascular tissues [56], and up to 53% of carotid artery atherosclerotic lesions are positive for HCMV DNA [57]. A number of animal studies have also provided compelling evidence that cytomegalovirus plays an important role in the pathophysiology of atherosclerosis, including several studies that have demonstrated more severe atherosclerosis in apoE −/− mice following systemic MCMV infection [58]–[61].
While an association has been recognized between cytomegalovirus infection and atherosclerosis, the strongest association of cytomegalovirus in vascular disease is with the development of restenosis and transplant vascular sclerosis. Several clinical studies have shown that HCMV infection is involved in accelerating both acute and chronic graft failure in all types of solid organ transplants by promoting vascular disease associated with rejection [52], probably by virus originating from the vasculature of transplanted organs from HCMV seropositive donors [62]. For example, HCMV infection was shown to double the 5-year rate of graft failure in cardiac allograft recipients due to accelerated transplant vascular sclerosis [63]. Similarly, kidney transplant allograft survival was decreased in asymptomatic HCMV-infected recipients during the first 100 days after transplantation when compared with recipient patients who had no evidence for HCMV infection, an outcome suggesting that HCMV infection, even when asymptomatic, has a negative impact on graft survival [64]. These clinical findings have been supported by a number of rat models of heart, kidney, lung, and small bowel transplantation in which infection with rat cytomegalovirus (RCMV) significantly decreased the mean time to graft failure while concomitantly increasing the degree of vasculopathy within the allograft tissue [65], [66].
Neovascularization is a complex, multi-step process of angiogenesis that rapidly takes place in response to inflammation and tissue injury, and involves many cell types, cytokines, chemokines, and proteases that work in concert to form new blood vessels from existing blood vessels. In brief (reviewed in [52]), angiogenesis is initiated by release of pro-angiogenic factors from activated endothelial cells and tissue-resident macrophages, followed by removal of pericytes that surround the existing blood vessels. This results in the breakdown of the basement membrane of the existing blood vessel wall through activation of several proteases including matrix metalloproteinases (MMPs). The release of extracellular remodeling proteins during continued degradation of the blood vessel wall leads to the release of growth factors that promote endothelial cell migration toward the angiogenic stimulus and ultimately mediates endothelial cell proliferation that drives the formation of neotubules. These neotubules in turn release additional growth factors such as platelet-derived growth factor (PDGF) that recruit vascular smooth muscle cells and pericytes that stabilize the newly formed blood vessel. Importantly, pro-angiogenic M2 macrophages have been shown recently to act as bridging cells that promote the fusion of neotubules into one continuous blood vessel [67]. Cytomegalovirus infection could therefore enhance neovascularization at various stages of angiogenesis through a number of direct and indirect mechanisms.
Monocytes are the primary target in vivo for HCMV (and MCMV and RCMV) infection [68], [69]. They serve as a site for virus latency and persistence [70], and help to disseminate virus throughout the host including the vasculature. When virus-infected monocytes enter the vasculature, they mature, and during the maturation process to become macrophages, they initiate an activation program that also serves to stimulate virus replication [71]. In this manner, infected macrophages may disseminate virus to other cells of the vasculature that are involved in angiogenesis and vascular disease. These include endothelial cells, smooth muscle cells, pericytes, and fibroblasts [52]. Given this complexity, the precise temporal relationship between virus infection of individual cell types and disease pathogenesis remains obscure and difficult to determine. Nonetheless, it is known that HCMV infection of endothelial cells induces the expression of adhesion molecules ICAM-1 and VCAM-1 [72] that serve to magnify transendothelial cell migration of inflammatory cells including monocytes. These monocytes become resident macrophages that promote angiogenesis by secretion of VEGF and other pro-angiogenic factors such as IL-6 [52].
During virus replication, the HCMV-encoded chemokine receptor US28 also plays a prominent yet multifaceted role in angiogenesis. Firstly, US28 has been shown to stimulate VEGF production directly by induction of COX-2 via activation of the NF-κB pathway [73]. Secondly, this HCMV-encoded chemokine receptor promotes the migration of macrophages in response to the CX3CL1 chemokine Fractalkine [52], a function that may help to attract additional HCMV-infected macrophages to areas of inflammation and thereby amplify angiogenesis. Thirdly, US28 also promotes the migration of vascular smooth muscle cells [74], but does so by binding to CC-chemokines and not Fractalkine [75]. Thus, US28 appears to stimulate the migration of both macrophages and vascular smooth muscle cells, but in a ligand-dependent manner. Whereas US28-induced migration of macrophages takes place after ligation with Fractalkine, but not CC-chemokines, US28-induced migration of vascular smooth muscle cells is mediated by binding to CC-chemokines, but not Fractalkine. Since HCMV-encoded US28 apparently plays multiple roles in promoting angiogenesis, we postulate the same is true for M33, the MCMV homologue of US28 [76]. Ongoing studies are therefore oriented toward testing the hypothesis that MCMV-encoded M33 plays significant roles in the pathophysiology and increased severity of CNV during chronic MCMV infection.
Additional direct and indirect mechanisms by which cytomegalovirus might contribute to angiogenesis and vascular disease are suggested by other studies. Examples include studies that have shown that HCMV infection induces a reduction of endothelial nitric oxide synthase activity commonly observed during cardiovascular disease [77]; RCMV induces the stimulation of a number of proteases including MMPs that are involved in degradation of the basement membrane required during the angiogenesis process [78]; HCMV induces an upregulation of a number of cellular chemokines including macrophage inflammatory protein 1 alpha (MIP1-α), MIP1-β, RANTES, and IL-2 that play critical roles in angiogenesis and development of vascular disease [74], [79]; and HCMV infection of coronary artery smooth muscle cells stimulates VEGF expression [80].
Since angiogenesis in health and disease is a process of great complexity that offers a number of mechanisms by which cytomegalovirus infection of multiple cell types might serve as a stimulatory cofactor in the development of more severe choroidal neovascularization, we elected to focus our study on a possible role for macrophages during chronic systemic MCMV infection. Macrophages can be either pro-angiogenic or anti-angiogenic depending on their polarization phenotype [25] that is regulated by the cytokine patterns encountered by macrophages within the resident tissue milieu [23], [26]. Classically activated macrophages, or M1 macrophages, exhibit an anti-angiogenic phenotype and produce high amounts of IL-12, IL-23, IL-6, and TNF-α, but low amounts of IL-10 [81]. In comparison, alternatively activated macrophages, or M2 macrophages, exhibit a pro-angiogenic phenotype and produce high amounts of IL-10, but low amounts of pro-inflammatory cytokines such as IL-6 and TNF-α [81]. Moreover, M1 macrophages inhibit angiogenesis by inducing a cell-death program in endothelial cells, whereas M2 macrophages promote angiogenesis by stimulating production and release of pro-angiogenic factors such as VEGF that encourage endothelial tip cell formation [52]. In this regard, Fantin and coworkers [67] have recently made the extraordinary observation that M2 macrophages may also play a critical role during formation of new blood vessels by serving as bridge cells to properly position and fuse neotubules into one continuous blood vessel, possibly via activation of the DII4-a ligand and expression of Notch receptors [82]. Thus, cytomegalovirus infection of monocytes and macrophages may influence angiogenesis-related activities by several possible mechanisms. For example, HCMV infection of monocytes appears to influence the polarization phenotype of the activated macrophage by modulating in a selective manner many M1/M2-associated factors [52], [83], thereby inducing angiogenesis through stimulation of VEGF production and other angiogenic factors. Importantly, MCMV-infected IC-21 mouse macrophages exhibited a pro-angiogenic M2 phenotype in our studies. Alternatively, HCMV infection could conceivably have a detrimental on the normal angiogenic process by promoting inflammation. HCMV infection of endothelial cells may also enhance the stability of newly formed blood vessels through stimulation and release of several cytokines and growth factors including the Notch 2 receptor [83]. We therefore postulate that chronic MCMV infection results in more severe choroidal neovascularization in our study by driving monocytes toward a M2 macrophage phenotype that favors angiogenesis through stimulation and release of pro-angiogenic factors that includes VEGF. It has not escaped our attention, however, that chronic MCMV infection might also cause more severe choroidal neovascularization by direct or indirect mechanisms associated with endothelial cell infection, a focus of future studies.
Splenic macrophages collected from chronically infected mice with the most severe choroidal neovascularization in our study showed significant increases in the amounts of transcripts to MMP-9 and COX-2, two proteins known to be involved in angiogenesis [52]. The most dramatic increase in transcript level, however, was observed for that of VEGF, a critical pro-angiogenic factor. This observation was confirmed in a second independent animal study by us that demonstrated an even greater increase in VEGF transcript production in splenic macrophages collected from chronically infected animals. One interpretation of these reproducible findings is that when chronically infected monocytes are recruited to choroidal sites of laser-induced damage, their activation programs are initiated and oriented toward the pro-angiogenic M2 phenotype. Since they are also chronically infected with MCMV, this activation program stimulates virus replication, an event that leads to enhanced production and secretion of several pro-angiogenic factors including VEGF. Inoculation of cultures of human foreskin fibroblasts or cultures of coronary artery smooth muscle cells with HCMV has been shown to result in stimulation of functionally active VEGF production [78]. It is therefore not surprising in the present study that MCMV infection of cultures of IC-21 mouse macrophages significantly stimulated production of VEGF mRNA and VEGF protein. Additional observations made during immunostaining studies also demonstrated that VEGF is indeed produced in high amounts by MCMV-infected IC-21 mouse macrophages, especially those in the early stages of cytopathology during plaque formation.
Of particular interest, however, was the additional observation that monolayer cells too early to be infected with MCMV (given the low multiplicity of infection used) were also VEGF-positive, an observation suggesting the attractive hypothesis that uninfected bystander macrophages might also be stimulated by adjacent MCMV-infected macrophages to produce enhanced amounts of VEGF during virus infection. Thus, MCMV infection of resident macrophages of tissues of the lung and spleen, and even bone marrow cells, could conceivably contribute to macrophage activation during chronic infection. MCMV-infected bone marrow cells, especially stromal cells, could favor a pro-angiogenic microenvironment that induces bystander activation during development within the marrow since stromal cells serve as a substrate upon which monocytes are induced to differentiate [84]. In addition, due to their high vascularity, both lung and spleen experience high monocyte traffic, and chronic MCMV infection of these tissues could induce bystander activation. We therefore postulate that chronic MCMV infection of monocytes and macrophages distant from the eye serves as an important mechanism for macrophage activation of the M2 phenotype that would contribute to the pro-angiogenic microenvironment of the choroidal tissues of the eye.
Treatment with ganciclovir, a potent inhibitor of active cytomegalovirus replication and HCMV disease in the clinical setting [42], [43], has been shown to delay the time to development of allograft rejection in heart transplant recipients [85], [86], a finding that underscores the importance for active HCMV replication in acceleration of vascular disease. Additional studies using experimental rat transplant models have provided similar data showing that ganciclovir therapy also reduced or prevented RCMV-associated acceleration of tissue rejection when compared with RCMV-infected animals not treated with the antiviral [87], [88]. Since we hypothesize that MCMV infection of macrophages plays a central role in amplifying the severity of experimental choroidal neovascularization in mice by stimulation of pro-angiogenic factors including VEGF, we used a similar antiviral approach to demonstrate that VEGF-specific transcript production by splenic macrophages collected from chronically infected mice was indeed ganciclovir-sensitive. This outcome strongly supports the need for active MCMV virus replication in stimulation of production of pro-angiogenic factors such as VEGF that is required for increased severity of choroidal neovascularization. These in vivo findings were duplicated and extended in culture using ganciclovir-treated, MCMV-infected monolayers of IC-21 mouse macrophages, and in a relatively dose-dependent manner.
In summary, the findings reported herein using an experimental mouse model of CNV serve to clarify our previous seroepidemiologic clinical study in which a significant association was identified between high titers of anti-HCMV IgG and development of neovascular AMD [35]. The presence of high anti-HCMV titers may indicate a subset of patients who harbor a greater total body burden of chronic HCMV infection, or who have experienced a recent, significant reactivation event. In either case, we hypothesize that the blood load of circulating HCMV-infected monocytes would be exceptionally high in this subset of patients. Upon recruitment to sites of drusen formation in patients who manifest the dry form of AMD, HCMV-infected monocytes would mature into tissue-resident macrophages with active virus replication, become polarized toward the pro-angiogenic M2 phenotype, and become a major source for production of a number of pro-angiogenic factors including VEGF that would amplify choroidal neovascularization associated with the wet form of AMD. We therefore believe that HCMV infection should be considered as a heretofore unrecognized risk factor for development of neovascular AMD. If true, subsets of patients who harbor a low virus load of HCMV would be predicted to experience decreased onset and progression of choroidal neovascularization, an occurrence that would impact their clinical outcome in terms of time of onset of visual loss and degree of visual loss. It is therefore possible that antiviral treatment might be effective in suppressing choroidal neovascularization associated with wet AMD in a fashion similar to that for suppression of allograft rejection in heart transplant recipients. Future studies will be oriented toward this investigation.
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 with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. Animal research protocols were approved by the Institutional Animal Care and Use Committees of the University of Miami Miller School of Medicine (A3324-01) and Georgia State University (A3914-01). All laser treatments were performed under anesthesia (intramuscular administration of ketamine hydrochloride, xylazine, and adepromazine), and all efforts were made to minimize suffering.
Adult female C57BL/6 mice were purchased from the National Institute of Aging (Bethesda, Maryland), and used throughout this investigation. Mice were allowed unrestricted access to food and water and maintained in alternating 12-hour light-dark cycles.
Stocks of MCMV were prepared in mouse salivary glands as described previously [89]. Briefly, BALB/c mice (Taconic Farms, Germantown, New York) were infected intraperitoneally with 1×102 to 1×103 plaque-forming units (PFU) of the Smith strain of MCMV (American Type Culture Collection, Manassas, VA) contained within a 0.2-ml volume. Approximately 14 days later, the salivary glands were removed aseptically, homogenized (10% wt/vol) in Dulbecco's modified Eagle's tissue culture medium containing 10% fetal bovine serum (DMEM), clarified by centrifugation, and 0.25 ml aliquots of the supernatant stored in liquid N2. Virus stocks were titered on monolayers of mouse embryo fibroblasts (MEF) grown in DMEM. Fresh aliquots of MCMV stock were thawed and used for single experiments. UV-inactivated virus was prepared by exposure of aliquots of MCMV stock to ultraviolet radiation for 30 min to inactivate virus infectivity as determined by no detectable plaque formation on MEF monolayers 7 days after undiluted inoculation.
Plan 1: To evaluate CNV severity after acute or chronic infection with MCMV, four groups of mice (n = 10 mice per group) were injected intraperitoneally with 40 ul of a non-lethal dose of infectious MCMV (1.5×106 plaque-forming units) or with an equivalent dose of UV-inactivated MCMV (controls). At 6 days (acute infection) and at 6 weeks and 12 weeks (chronic infection) after inoculation, the eyes of age-matched mice were subjected to bilateral laser treatment to induce CNV as described below. Mice were matched in age (10 months) at time of laser treatment. Four weeks later, the right eyes were collected for flat-mount analysis, and the left eyes were collected for histopathologic analysis.
Plan 2: To evaluate macrophages for their patterns of production of various pro-angiogenic factors during CNV at time of acute versus chronic MCMV infection, the eyes of groups of mice (n = 10 mice per group) were subjected to bilateral laser treatment at 6 days or at 12 weeks after intraperitoneal injection with infectious MCMV. The control group for this study consisted of groups of mice injected intraperitoneally with UV-inactivated virus. Mice were matched in age (10 months) at time of laser treatment for these animal groups. Four weeks after CNV induction, splenic macrophages were collected from all animals for quantitative RT-PCR assay analysis of several gene transcripts relevant to inflammation and/or neovascularization. Whole eyes, choroidal tissues, tissues from various organs (salivary glands, lung, spleen), and bone-marrow cells (CD34+ cells) were also collected from mice of the same animal groups and analyzed by standard plaque assay for detection of infectious virus or analyzed by PCR assay for detection of MCMV-specific DNA sequences.
Plan 3: To confirm mouse macrophages as a source for VEGF production following MCMV infection, monolayers of the IC-21 mouse macrophage cell line (American Type Culture Collection, Manassas, VA, USA) [41] were inoculated either with MCMV (moi = 2.5), UV-inactivated MCMV, maintenance medium only, or maintenance medium containing lipopolysaccharide (LPS) (100 ng/ml). All cells were harvested at 24 or 48 hrs postinfection and subjected to quantitative real time RT-PCR assay for quantification of VEGF mRNA and TNF-α mRNA, standard ELISA for quantification of VEGF protein production, or immunostaining for detection and pattern of VEGF production.
Plan 4: To determine the effect of antiviral treatment on production of VEGF mRNA and TNF-α mRNA by splenic macrophages at time of chronic MCMV infection, groups of mice (n = 10 mice per group) were injected intraperitoneally with 40 ul of a non-lethal dose of infectious MCMV (1.5×106 plaque-forming units) or maintenance medium (mock infected). At 12 weeks after inoculation, MCMV-infected or mock-infected mice were treated intraperitoneally with ganciclovir for 7 days at a dose of 40 mg/kg/day, a dose that reflects the relative decreased sensitivity of MCMV to ganciclovir when compared with the sensitivity of HCMV to ganciclovir [43]. Untreated control MCMV-infected or mock-infected mice received daily intraperitoneal injections of phosphate-buffered saline for 7 days. Following the 7-day regimen of ganciclovir or phosphate-buffered saline treatment, splenic macrophages were collected from ganciclovir-treated and untreated chronically infected mice and compared by quantitiative real time RT-PCR assay for levels of VEGF mRNA production. To determine the effect of antiviral treatment on production of VEGF mRNA and TNF-α mRNA by mouse macrophages during acute MCMV infection, monolayers of IC-21 mouse macrophages were inoculated either with MCMV (moi = 2.5) or mock-infected with maintenance medium. At 1-hr postinfection, MCMV-infected and mock-infected monolayers were treated either with 15, 30, or 60 uM of ganciclovir or treated with phosphate-buffered saline (control). At 24 hr postinfection, all monolayers were harvested and subjected to quantitative RT-PCR assay for quantification and comparison of levels of VEGF mRNA production.
At 6 days, 6 weeks, or 12 weeks after injection with infectious or UV-inactivated MCMV, diode red laser was used to create choroidal thermal burns bilaterally and induce experimental CNV as described previously [39]. Four weeks after laser application, mice were euthanized, and subjected to histopathologic analysis as well as flat-mount analysis of surface area, vascularity, and cell density of CNV. All images were digitally acquired (Axiovision, Zeiss) and recompiled (Photoshop version 6.0; Adobe, San Jose, California). Surface area of CNV lesions was determined by using either fluorescein-isothiocyanate (FITC)-dextran (Sigma, St. Louis, Missouri) fluorescence or propidium iodide (PI, Sigma) fluorescence, and outlining the margins of the lesion with a computer analysis software (Photoshop 6.0). The area in pixels was normalized by dividing the average of the optic disc measured in 10 independent eyes. Five eyes were examined 4 weeks after laser treatment to determine the average spot size (0.48 disc areas). A CNV was determined to be present if the surface area of an individual lesion was greater than 0.50 disc areas.
Four weeks after bilateral laser treatment of groups of mice infected systemically with MCMV for 6 days, 6 weeks, or 12 weeks, left eyes were carefully removed from all animals following euthanasia, fixed in 10% buffered formalin, paraffin embedded, sectioned with hematoxylin and eosin, and examined by light microscopy for detection and quantification of areas of CNV.
Following removal of spleens under sterile conditions from euthanized mice, a Spectra/Mesh macroporus 210 µm filter (Spectrum Laboratories, Inc., Los Angeles, California) was used to obtain splenic macrophages after maceration of individual spleens in a Hanks balance salt solution (HBSS) medium containing 1 M HEPES, 1 M NaAZ, and fetal bovine serum. ACK buffer was added to the spleen suspension to lyse red blood cells. The remaining cells were centrifuged and resuspended in HBSS medium containing rat anti-mouse F4/80 conjugated with PE (Caltag, Burlingame, California). Splenic macrophages were then purified by magnetic column separation using MACS Anti-PE Microbeads (Miltenyi, Auburn, California) as specified by manufacturer's instructions.
At the time of euthanasia and under sterile conditions, tibias and femurs were dissected and bone marrow was extracted by slowly flushing the dyaphyseal channel with HBSS medium using a 27-gauge needle. Bone marrow was homogenized, filtered, centrifuged, and resuspended in HBBS medium. Red blood cells were lysed with ACK buffer, and the remaining cells were incubated with rat anti-mouse CD34 conjugated with PE (BD Biosciences, Pharmingen, San Diego, California). CD34+ vascular precursor cells were then purified by magnetic column separation using MACS Anti-PE Microbeads (Miltenyi) as specified by manufacturer's instructions.
Whole eyes collected from mice at 30 days after laser-induced CNV were frozen individually at −80°C. At time of quantitative plaque assay, eyes were thawed, homogenized individually in 1.0 ml of cold Dulbecco's modified Eagle's medium (DMEM) containing 10% fetal bovine serum, and clarified by centrifugation. Ten-fold dilutions of the resulting supernatants were titered in duplicate onto monolayers of MEF contained within 6-well plates, allowed to absorb for 1 hour at 37°C, overlaid with methylcellulose containing DMEM, and incubated for 5 or 6 days at 37°C in a humidified CO2 atmosphere. Monolayers were screened daily for 7 days using an inverted light microscope for detection of plaques of MCMV-induced cytopathology.
DNA was extracted from whole eyes, salivary glands, lungs, bone marrow, spleens, and isolated macrophages collected from euthanized mice using the QIAamp Tissue kit (QIAGEN GmbH, Valencia, California) according to manufacturer's instructions and subjected to PCR assay to detect MCMV-specific DNA using primers for immediate early 1 (IE1) and glycoprotein H (gH) genes. The primers used were kindly provided by Dr. Daniel D. Sedmak, Ohio State University College of Medicine, Columbus, Ohio. The primer pair for MCMV IE1 gene was 5′-TAGCCAATG ATATCTTCGAGCG-3′ and 3′-ATCTGGTGCTCCTCAGATCAGCTAA-5′, and the primer pair for MCMV gH gene was 5′-TTCAGTTCAACTCGAA-3′ and 3′-GGGAAGAAGTACTCGACCGG-5′. PCR amplification of β-actin was performed as an internal control. Actin primers consisted of 5′-ATTGTGATGGACTCCGGTGA-3′ and 3′-AGCTCATAGCTCTTCTCCAG-5′. DNA extracted from tissue homogenates was eluted in 100 µl of distilled water, and stored at −20 C until analysis. DNA was amplified in a total volume of 25 µl with 200 nM of each primer and 1.0 U of Taq DNA polymerase (Gibco BRL) added in 2.5 µl of a PCR buffer (50 mM KCL, 20 mM Tris-HCl [pH 8.4], and 1.5 mM MgCl2). PCR assays were performed on a Perkin Elmer 9600 thermocycler (PE Applied Biosystems). PCR assay conditions consisted of an initial denaturation step of 4 min at 94 C, followed by 35 cycles, with 1 cycle consisting of 30 sec at 94 C, 30 sec at 53 C, and 30 sec 72 C. Amplification products were separated by electrophoresis through 1% agarose gels, and stained with ethidium bromide for visualization.
Total RNA was extracted from whole bone marrow cells (CD34+ cells), splenic macrophages, or MCMV-infected IC-21 mouse macrophage monolayers using Tri-Reagent and prepared for quantitative RT-PCR reactions as described previously [90]. Real time RT-PCR assay was used to quantify several cellular transcripts of interest that included mouse tumor necrosis factor-alpha (TNF-α), matrix metalloproteinase-9 (MMP-9), vascular endothelial growth factor (VEGF), VEGF receptor 1 (VEGFR1), VEGF receptor 2 (VEGFR2), platelet-derived growth factor-beta (PDGF-β), cyclooxygenase (COX-2), and inducible nitric oxide synthase (iNOS). Real time RT-PCR assays were performed for TNF-α, VEGFR1, VEGFR2, and COX-2 mouse transcripts using commercially available kits (Perkin Elmer Applied Biosciences). The primer pair for real time RT-PCR assay of mouse PDGF-β mRNA was 5′-AAGCACACGCATGACAAG-3′ and 3′-GGGGCAATACAGCAAATAC-5′; for VEGF mRNA was 5′- CGAAACCATGAACTTTCTGC-3′ and 3′-CCTCAGTGGGCACACACTCC-5′; for MMP-9 mRNA was 5′-CAGGATAAACTGTATGGCTTCTGC-3′ and 3′- GCCGAGTTGCCCCCA-5′; and for iNOS mRNA was 5′-TGACGCCAAACATGACTTCAG-3′ and 3′-GCCATCGGGCATCTGGTA. Transcripts of these molecules were normalized to 18S ribosomal RNA transcripts via standard curves generated using serially diluted samples of mRNA (0.001–100 ng). Real time RT-PCR assays were performed in duplicate with quantitative values determined for each molecule as the ratio of the mean values for a specific mRNA versus 18S mRNA. Median values for each molecule were calculated and normalized to samples obtained from sham-inoculated control animals (100%).
MCMV-infected and mock-infected monolayers of IC-21 mouse macrophages grown on 6-well chamber slides were harvested at 24 and 48 hr postinfection, fixed in cold ethanol, dried, and reacted with 5% normal goat serum containing 0.2% Triton X-100. Following three washings in phosphate-buffered saline, slides were incubated for 1 hr with either rabbit anti-mouse VEGF IgG (1∶100 dilution) (Santa Cruz Biotechnology, Santa Cruz, CA) or normal rabbit IgG (1∶100 dilution (Santa Cruz Biotechnology, Santa Cruz, CA), washed three times with phosphate-buffered saline, and reacted with biotinylated anti-rabbit IgG secondary antibody using the Rabbit ABC Staining system (Santa Cruz Biotechnology, Santa Cruz, CA). Chamber slides were mounted on standard microscope slides and cell nuclei were counterstained using Vectashield Mounting Medium containing DAPI (Vector Laboratories, Burlingame, CA). All slides were examined and photographed using a Nikon Eclipse 50i microscope equipped with an X-Cite Series 120 Epi-fl illuminator.
Morphometric data for individual lesions in each eye were averaged to provide one value per eye. Mean and standard deviation values for each group was calculated and p values were determined using student t-test and one-way analysis of variance+Dunnett's multiple comparison post-hoc test (GraphPad Prism 4.0, San Diego, CA). Values of p≤0.05 were considered statistically significant for all forms of statistical analysis used.
|
10.1371/journal.pcbi.1004687 | A Boolean Function for Neural Induction Reveals a Critical Role of Direct Intercellular Interactions in Patterning the Ectoderm of the Ascidian Embryo | A complex system of multiple signaling molecules often produce differential gene expression patterns in animal embryos. In the ascidian embryo, four signaling ligands, Ephrin-A.d (Efna.d), Fgf9/16/20, Admp, and Gdf1/3-r, coordinately induce Otx expression in the neural lineage at the 32-cell stage. However, it has not been determined whether differential inputs of all of these signaling pathways are really necessary. It is possible that differential activation of one of these signaling pathways is sufficient and the remaining signaling pathways are activated in all cells at similar levels. To address this question, we developed a parameter-free method for determining a Boolean function for Otx expression in the present study. We treated activities of signaling pathways as Boolean values, and we also took all possible patterns of signaling gradients into consideration. We successfully determined a Boolean function that explains Otx expression in the animal hemisphere of wild-type and morphant embryos at the 32-cell stage. This Boolean function was not inconsistent with three sensing patterns, which represented whether or not individual cells received sufficient amounts of the signaling molecules. These sensing patterns all indicated that differential expression of Otx in the neural lineage is primarily determined by Efna.d, but not by differential inputs of Fgf9/16/20, Admp, and Gdf1/3-r signaling. To confirm this hypothesis experimentally, we simultaneously knocked-down Admp, Gdf1/3-r, and Fgf9/16/20, and treated this triple morphant with recombinant bFGF and BMP4 proteins, which mimic Fgf9/16/20 and Admp/Gdf1/3-r activity, respectively. Although no differential inputs of Admp, Gdf1/3-r and Fgf9/16/20 signaling were expected under this experimental condition, Otx was expressed specifically in the neural lineage. Thus, direct cell–cell interactions through Efna.d play a critical role in patterning the ectoderm of the early ascidian embryo.
| It is often difficult to understand a complex system of multiple signaling molecules in animal embryos only with experimental procedures. Although theoretical analysis might solve this problem, it is often difficult to precisely determine parameters for signaling gradients and kinetics of signaling molecules. In the present study, we developed a parameter-free method for determining a Boolean function for understanding a complex signaling system using gene expression patterns of signaling molecules and geometrical configurations of individual cells within the embryo. In the ascidian embryo, four signaling ligands, Ephrin-A.d (Efna.d), Fgf9/16/20, Admp, and Gdf1/3-r, coordinately induce Otx expression in the neural lineage at the 32-cell stage. In addition to determining a Boolean function, our method determined sensing patterns, which represented whether or not individual cells received sufficient amounts of the signaling molecules. The sensing patterns predicted that differential expression of Otx in the neural lineage is primarily determined by Efna.d, but not by differential inputs of Fgf9/16/20, Admp, and Gdf1/3-r. We confirmed this prediction by an experiment. As a result, we found that only Efna.d signaling pathway is differentially activated between ectodermal cells and the remaining signaling pathways are activated in all ectodermal cells at similar levels.
| In animal embryos, cell-cell interactions directed by secreted and membrane-bound signaling ligands play an important role in establishing specific gene expression patterns. There are 16 ectodermal cells in the animal hemisphere of the 32-cell embryo of the ascidian, Ciona intestinalis, and all 16 have the potential to express Otx upon induction (Fig 1A). Four signaling molecules, Fgf9/16/20, Admp (anti-dorsalizing morphogenetic protein; a signaling molecule belonging to the BMP subfamily in the TGFβ superfamily), Gdf1/3-r [formerly called Gdf1/3-like and renamed according to the nomenclature guideline recently published [1]], and Efna.d (formerly EphrinA-d), cooperatively regulate Otx expression in four cells, which give rise to neural cells [2–4]. Fgf9/16/20 activates Otx expression through the ERK pathway, which is antagonized by Efna.d [5, 6]. Admp and Gdf1/3-r negatively regulate Otx expression by inducing the binding of the effector transcription factor Smad to an Otx enhancer (Fig 1B and 1C). The observation that Otx expression expands throughout the ectoderm upon knockdown of Efna.d or double-knockdown of Admp and Gdf1/3-r [4] indicates that these three genes are essential for differential expression of Otx within the ectodermal cells and patterning the ectoderm. On the other hand, another study indicated that a differential input of Fgf9/16/20 signaling could direct differential Otx expression in the ectoderm [7]. Thus, it has not yet been established which of these factors is critical for patterning of the ectoderm of normal embryos. In other words, it has not been determined whether differential inputs of all of these signaling pathways are really necessary. For instance, it is possible that differential activation of one of these signaling pathways is sufficient and the remaining signaling pathways are activated in all cells at similar levels. Because our previous experiments [4] did not necessarily give an answer to this question, we took an advantage of theoretical analysis in the present study.
Although quantitative models have successfully simulated molecular gradients for embryonic patterning in other model systems [8–10], it is difficult to precisely determine parameters for signaling gradients and kinetics of signaling molecules in the ascidian embryo. Boolean functions provide an alternative, because inputs and outputs are treated as binary values, and parameters that are difficult to determine are not used. In previous studies, Boolean functions have successfully explained how combinations of different transcription factors determine specific gene expression patterns [11–15]. Here we report determination of a Boolean function for Otx expression in the 32-cell embryo of Ciona intestinalis. This function reveals how individual cells sense signaling inputs and which signaling is the limiting factor for patterning the ectoderm.
Here we introduce a method for determining a Boolean function of gene expression directed by extracellular signals within a population of equivalent cells. Before formalizing neural induction of the ascidian embryo, we first considered a Boolean function describing a simple hypothetical biological system illustrated in Fig 2A. This system consists of two cells, I and II, and two signaling molecules, a and b. Cells I and II initially express the same set of transcription factors, and are therefore equivalent. After a sufficient period of time, gene o is activated only in cell I but not in cell II under control of signaling molecules a and/or b. Hence, the Boolean function for the expression of gene o is represented by [Xo = F(Xa,Xb)], where Xo represents expression of gene o, and Xa and Xb represent the signaling states of a and b. Inputs and outputs are considered in binary space. If a signal sufficiently activates its intracellular pathway, it is represented as ‘1’, and otherwise as ‘0’ hereafter. Because there are 4 (= 22) possible combinations of input signaling states in each cell in this hypothetical system, there are 16 (= 4×4) possible states in the whole system (Fig 2B). These states, which are represented as (Xa,Xb), are called "sensing patterns" hereafter, because they represent how individual cells sense individual signaling inputs. In this hypothetical system, signaling molecule a comes from the upper side of Fig 2A, and signaling molecule b comes from the lower side. Obviously, eleven sensing patterns are incompatible with the following two simple principles, which we call Rule 1 and Rule 2.
We next applied this logic to the signaling system that induces the neural marker gene Otx in the neural lineage (a6.5 and b6.5) of the 32-cell Ciona embryos (Fig 1A). Previous studies revealed that four signaling molecules, Admp, Efna.d, Fgf9/16/20 and Gdf1/3-r, are directly involved in inducing Otx expression in two pairs of cells (a6.5 and b6.5) within 16 equivalent ectodermal cells (eight pairs of cells) in the animal hemisphere (Fig 1A–1C) [2–4]. The activities of these signaling pathways are denoted by the binary variables, Xadmp, Xefn, Xfgf, and Xgdf, and the expression of Otx (Xotx) is represented by a Boolean function, Xotx = F(Xadmp, Xefn, Xfgf, Xgdf).
In this biological system, there are four binary input variables and eight pairs of equivalent cells, and the number of possible sensing patterns is 4,294,967,296 [= (24)8]. As in the case of the hypothetical biological systems, we first screened individual sensing patterns with Rule 1 and Rule 2.
During the 32-cell stage, the Ciona embryo dynamically changes its shape. Therefore we tried to determine sensing patterns and Boolean functions for three early 32-cell embryos, for which geometric data were obtained in a previous study [7] (S3 Table; S4A–S4C Fig). In these analyses, only geometric data were different from the first analysis for the mid-to-late 32-cell embryo. We obtained the same four sensing patterns from each of these three virtual embryos (S4D Fig). Three of them were the same as the ones obtained from the mid-to-late 32-cell embryo, and were compatible with the same Boolean function as that obtained from the mid-to-late 32-cell embryo. The remaining one, sensing pattern 4, was slightly different, and was compatible with eight Boolean functions, one of which was the Boolean function compatible with the other three sensing patterns (S4D and S4E Fig).
Even under sensing pattern 4, the Boolean functions compatible with this sensing pattern indicated that Efna.d is a critical factor for patterning the ectoderm. Sensing pattern 4 showed that Efna.d signaling and Gdf1/3-r signaling are not sufficiently active in a6.5 and b6.5, implying that these two factors are candidates for a factor for patterning the ectoderm. However, all eight Boolean functions compatible with sensing pattern 4 indicated that Gdf1/3-r signaling cannot pattern the ectoderm under this sensing pattern, because F(Xadmp = 1, Xefn = 0, Xfgf = 1, Xgdf = 0) and F(Xadmp = 1, Xefn = 0, Xfgf = 1, Xgdf = 1) gave the same result, Xo = 1. Therefore, even if Otx expression in early 32-cell embryos is directed by a Boolean function different from the one in the mid-to-late 32-cell embryo, our analysis indicated that Efna.d is the limiting factor for patterning the ectoderm of the 32-cell embryo.
The above prediction that Efna.d is the limiting factor for patterning the ectoderm of the 32-cell embryo was consistent with our observation in a previous study that Otx expression is expanded throughout epidermal cells of Efna.d morphants [4]. However, it has not been determined whether differential inputs of Fgf9/16/20, Admp, and Gdf1/3-r are really unnecessary for patterning the ectoderm. To test this, we used triple morphants of Fgf9/16/20, Admp, and Gdf1/3-r. First we confirmed our previously result that Fgf9/16/20/Admp/Gdf1/3-r morphants do not express Otx in the animal hemisphere [4] (Fig 4A). Next, we incubated Fgf9/16/20/Admp/Gdf1/3-r morphants in sea water containing recombinant bFGF and BMP4 proteins, which mimic Fgf9/16/20 and Admp/Gdf1/3-r activity [4]. The bFGF concentration was determined empirically on the basis of our previous study [4]; Otx is expressed on average in two cells of Fgf9/16/20 morphants incubated with 1 ng/mL of bFGF [4]. In this experimental condition, in which no gradients of bFGF and BMP4 within embryos were expected, Otx was expressed predominantly in the neural lineage, as in control embryos (Fig 4B; Table 1). Although relatively weak Otx expression in the epidermal lineage was observed only in a very small number of embryos, cells with neural fate almost always expressed Otx in these embryos, and ectopic expression was also observed in one unperturbed embryo (Table 1). In addition, as expected in our previous study [4], Otx expression was observed in the neural and epidermal lineages of quadruple morphants of Fgf9/16/20, Admp, Gdf1/3-r, and Efna.d incubated in sea water containing recombinant bFGF and BMP4 proteins (Fig 4C; Table 1), while injection of the same amount of a control morpholino oligonucleotide did not affect Otx expression (Fig 4D; Table 1). Thus, as predicted by the theoretical model, a difference in strength of Efna.d signaling, which is known to attenuate ERK activation [5, 6], can evoke specific Otx expression without differential inputs of Fgf9/16/20, Admp and Gdf1/3-r, even if differential inputs of these factors might contribute to specific Otx expression in normal embryos.
We determined a Boolean function for Otx expression in the animal hemisphere of the mid-to-late 32-cell ascidian embryo, based on a theoretical analysis using data obtained in previous studies [2–4, 7, 18] and in this study. We found that three sensing patterns of signals are compatible with this Boolean function. It is possible that the 32-cell-embryo normally takes only one of these sensing patterns. However, because the Boolean function indicates that Otx is specifically expressed in the neural lineage under either of these sensing patterns, the choice of sensing patterns by 32-cell embryos might not be strictly determined. In other words, these sensing patterns might represent fluctuations of signaling and robustness of this system.
The cis-regulatory module of Otx for the expression in the neural lineage contains multiple Ets-binding sites and Smad-binding elements (SBEs) [3, 4] (Fig 1B). Ets is positively regulated by Fgf9/16/20 signaling and negatively regulated by Efna.d signaling. SBEs are responsive to signaling of Admp and Gdf1/3-r, and negatively regulate the expression of Otx in the neural lineage. The Boolean function in Fig 3D indicates that Fgf9/16/20 and Efna.d work positively and negatively. It also indicates that Admp and Gdf1/3-r have a redundant function, because these two molecules are interchangeable. Thus, although no particular cis-regulatory mechanism was assumed in the present study, the cis-regulatory module is not inconsistent with the Boolean function that we revealed in the present study.
Our theoretical method does not use quantitative parameters that cannot be easily measured, such as the kinetics of individual signaling molecules. Instead, we only use expression patterns of signaling molecules and geometrical configurations of individual cells within the embryo. The former was determined by in situ hybridization [3, 18], and the latter was determined by computation of a series of confocal images [7]. Although activities of signaling pathways were treated as binary values, gradients or differential inputs of signaling molecules were taken into consideration. For this purpose, we assumed that the area of the contact of a cell with its surrounding cells that express a ligand is correlated with the strength of signaling. This is the case at least for Fgf9/16/20 [7], and it will be hard to imagine cases in which this assumption is inappropriate in early ascidian embryos with the following two reasons. First, our assumption also takes into consideration a case in which diffusion is very fast and no gradient is formed. Second, if an antagonist altered the activity of a signaling molecule within the embryo, this molecule could be considered as an additional signaling molecule. However, no genes for known antagonists for Fgf9/16/20, Admp, and Gdf1/3-r are expressed from the zygotic genome at or before the 32-cell stage [18].
The sensing patterns of individual cells in normal embryos showed that cells that do not sense Efna.d signaling above a threshold level give rise to neural cells, whereas cells that sense sufficient levels of Efna.d signaling give rise to epidermal cells. A previous study indicated that a differential input of Fgf signaling can differentiate ectodermal cells to neural cells under some experimental conditions and Fgf signaling is thought to be transmitted stronger in neural cells [7]. The present study does not necessarily rule out a possibility that a differential input of Fgf signaling contributes to patterning of the ectoderm in a normal embryo. A differential input of Fgf signaling will indeed contribute to patterning of the ectoderm in a normal embryo with the following three reasons: (1) Fgf signaling might be stronger in neural cells than in epidermal cells [7] (Fig 3A); (2) Efna.d signaling attenuates the ERK pathway activated by Fgf9/16/20 [5, 6] (Fig 1B); (3) a differential level of activation of the ERK pathway controls the expression of Otx [2–4]. However, our results indicate that a differential input of Efna.d is essential for the initial patterning of the ectoderm at the 32-cell stage in a normal embryo.
Secreted molecules often form continuous gradients, which are used for patterning of animal embryos [19]. Our result indicates that concentration gradients of Fgf9/16/20, Admp and Gdf1/3-r, or differential inputs of them, are not required, although these molecules are required for establishing the proper expression pattern of Otx. Efna.d is a membrane-bound protein, and therefore cannot form a continuous gradient as secreted molecules do. Cells located near the animal pole are surrounded by ectodermal cells, and are therefore expected to receive a stronger Efna.d signal. On the other hand, cells located in the periphery of the animal hemisphere are not completely surrounded by ectodermal cells, and are therefore expected to receive a weaker Efna.d signal. This is reminiscent of the differentiation of inner cell mass and trophectoderm of mammalian embryos. The fate choice between them mainly depends on Hippo signaling, which is thought to be activated through direct cell–cell interaction as Efna.d signaling [20, 21]. In the early animal embryo, cell–cell interaction through direct contacts may provide a more robust system for creating sharp boundaries of gene expression.
The contact areas of individual animal blastomeres of the 32-cell embryo with cells expressing Admp, Efna.d, Fgf9/16/20 and Gdf1/3-r were calculated using four different 3D-virtual embryos, which were reconstructed virtually from several series of confocal images [7] and the expression patterns of these genes [3, 18]. Given the delay between gene expression and protein translation, we assumed that cells descended from cells expressing a ligand gene at the 16-cell stage would express the encoded protein at the 32-cell stage [4]. The contact surfaces of individual animal blastomeres of the 32-cell embryo with anterior vegetal cells expressing Fgf9/16/20 were previously calculated [7]. We recalculated the contact surfaces of individual animal blastomeres of the 32-cell embryo with all cells expressing Fgf9/16/20 using geometrical data [7], in our previous study [4] and the present study. The contact surfaces of individual animal blastomeres with cells expressing Efna.d for one early 32-cell embryo were also calculated previously [4]. In the present study, we calculated the contact areas of individual cells with cells expressing Admp and Gdf1/3-r (S1 and S3 Tables) for three early 32-cell embryos and one mid-to-late 32-cell embryo. The files we used were downloaded from the Aniseed database [22] (http://www.aniseed.cnrs.fr), and the file names are shown in S1 and S3 Tables. Note that we ruled out autocrine effects of Efna.d, because it is a GPI-anchored membrane protein.
C. intestinalis (type A) adults were obtained from the National Bio-Resource Project for Ciona. The morpholino oligonucleotides for Fgf9/16/20, Admp, Gdf1/3-r, and Efna.d used in this study were those used in our previous study [4]. These morpholino oligonucleotides were designed to block translation. We also used a standard control MO (5’-CCTCTTACCTCAGTTACAATTTATA-3’) purchased from Gene Tools, LLC. DIG-RNA probes for whole-mount in situ hybridization were synthesized by in vitro transcription with T7 RNA polymerase as described previously [18]. Human recombinant bFGF (Sigma) and BMP4 (HumanZyme) were used at concentrations of 1 ng/mL and 100 ng/mL, respectively.
Identifiers for genes examined in the present study are as follows: CG.KH2012.C2.125 for Fgf9/16/20, CG.KH2012.C3.716 for Efna.d, CG.KH2012.C2.421 for Admp, CG.KH2012.C4.547 for Gdf1/3-r, and CG.KH2012.C4.84 for Otx.
|
10.1371/journal.pbio.0060285 | Specialization Does Not Predict Individual Efficiency in an Ant | The ecological success of social insects is often attributed to an increase in efficiency achieved through division of labor between workers in a colony. Much research has therefore focused on the mechanism by which a division of labor is implemented, i.e., on how tasks are allocated to workers. However, the important assumption that specialists are indeed more efficient at their work than generalist individuals—the “Jack-of-all-trades is master of none” hypothesis—has rarely been tested. Here, I quantify worker efficiency, measured as work completed per time, in four different tasks in the ant Temnothorax albipennis: honey and protein foraging, collection of nest-building material, and brood transports in a colony emigration. I show that individual efficiency is not predicted by how specialized workers were on the respective task. Worker efficiency is also not consistently predicted by that worker's overall activity or delay to begin the task. Even when only the worker's rank relative to nestmates in the same colony was used, specialization did not predict efficiency in three out of the four tasks, and more specialized workers actually performed worse than others in the fourth task (collection of sand grains). I also show that the above relationships, as well as median individual efficiency, do not change with colony size. My results demonstrate that in an ant species without morphologically differentiated worker castes, workers may nevertheless differ in their ability to perform different tasks. Surprisingly, this variation is not utilized by the colony—worker allocation to tasks is unrelated to their ability to perform them. What, then, are the adaptive benefits of behavioral specialization, and why do workers choose tasks without regard for whether they can perform them well? We are still far from an understanding of the adaptive benefits of division of labor in social insects.
| Social insects, including ants, bees, and termites, may make up 75% of the world's insect biomass. This success is often attributed to their complex colony organization. Each individual is thought to specialize in a particular task and thus become an “expert” for this task. Researchers have long assumed that the ecological success of social insects derives from division of labor, just as the increase in productivity achieved in human societies; however, this assumption has not been thoroughly tested. Here, I have measured task performance of specialized and unspecialized ants. In the ant species studied here, it turns out that specialists are no better at their jobs than generalists, and sometimes even perform worse. In addition, most of the work in the colony is not performed by the most efficient workers. So the old adage “The Jack of all trades is a master of none” does not seem to apply to these ants, suggesting that we may have to revise our understanding of the benefits of colony organization.
| Social insects are enormously successful ecologically. Ants, social bees, social wasps, and termites may make up 75% of the world's insect biomass, they play a major role in soil turnover and nutrient cycling, and they often surpass vertebrates in their biomass in a habitat [1]. Division of labor is often cited as the primary reason for the ecological success of social insects, particularly ants [1–5]. Division of labor implies that individuals within a colony specialize on particular tasks, such as brood care, foraging, nest building, or defense, and conversely that each task is performed by a particular subset of the workers [6–12]. If division of labor caused ecological success in social insects, it must have conferred benefits to colonies. What exactly are these adaptive benefits of specialization? According to the famous economist Adam Smith [13], specialization in human industry had three benefits: (1) increased individual efficiency through learning, (2) reduction of switching costs, and (3) the invention of machines. The first of these may be called the “Jack-of-all-trades is master of none” hypothesis: specialists are individually more efficient at performing their task than generalists. Although this hypothesis underlies many discussions of division of labor [2,6,14–18], it has rarely been tested, as pointed out by many authors [8,9,11,17,19–27]. Here, I address this issue by measuring individual efficiency of more than 1,100 workers of the ant species Temnothorax albipennis in several tasks. This allows me to test whether more specialized individuals are also more efficient.
Most previous research on division of labor in insects concentrates on the mechanisms of task allocation (e.g., [6,11,27–32]) instead of its consequences for individual or colony-level performance. Remarkably few studies have investigated the efficiency of individuals and how it relates to which tasks they perform [9,15,22,23,25,33–36]. In principle, there are two ways in which individuals who are specialists may achieve higher efficiency in performing “their” task: they may learn to perform a task better with frequent experience; or colonies may produce different specialists that are evolutionarily adapted to particular tasks. Worker polymorphism may be such an evolutionary adaptation: in ants with morphological castes, we know that “majors” (morphologically specialized ants) tend to be better at some tasks than the generalist “minors,” for example, they may be good at cutting leaves or walking fast to transport them [35,37]; they may also be good at other transport, defense, or food storage: [25,38–41], but are bad at performing brood care [22]. Polymorphism among workers, however, is rare, only occurring in less than 15% of ant genera [21,42]. Worker polymorphism also does not occur in bees or wasps. In bumble bees, workers exhibit size polymorphism, albeit not the variation in shape (allometry) characteristic of polymorphic ants. Workers may also differ genetically, leading to variation in behavior [43]. Although this may be unlikely to produce evolutionarily stable colony-level benefits [44], it is clear that worker differences in task preferences [30,31,45], as well as variation in quality of task performance [31,46,47], may be linked to genetic variation, and such variation may therefore play a role in specialization.
Individual efficiency can also increase through learning, without morphological adaptations. For example, bumble bees and honey bees need to invest time in learning to handle particular flower types efficiently [48,49]. Since learning incurs various types of cost (production and maintenance of neural tissue, energy costs of actually using it, and costs in errors made and time invested [50–55]), it may be beneficial to minimize the number of skills that an individual has to learn. This could lead to increased individual efficiency in specialists. Indeed, many bee foragers specialize on particular flower types, possibly to minimize costs of learning handling procedures [48,56,57] (although see [52]).
So, is the specialist worker in a colony the “master” of one task, while the generalist is a “master of none”? The Jack-of-all-trades is a master of none hypothesis would predict that more specialized workers perform a task with higher efficiency than generalists, whether this is a result of learning or adaptation. This hypothesis is what I test here for the ant Temnothorax albipennis. A set of tasks that are relevant in different contexts (foraging, emigrations, and nest building) will be used. Specifically, I will test which of the following specific hypotheses best predicts individual efficiency: (1) More specialized workers perform a task more efficiently. (2) Workers that are more active overall (in different tasks) perform tasks more efficiently. (3) Workers that engage in tasks with a short delay (who may have low response thresholds) perform tasks more efficiently than those who delay longer.
To understand the role that individual efficiency plays in colony division of labor, I will also test the following two hypotheses: (1) At the colony level, most labor is contributed by highly efficient workers. (2) At the colony level, most labor is contributed by specialized workers.
A total of 1,142 ants from 11 colonies of Temnothorax albipennis were marked individually and filmed performing four tasks: carrying brood items in a colony emigration, foraging for honey solution, foraging for protein (dead Drosophila flies), and collecting sand grains (hereafter, “stones”—they are about a third the size of a worker ant) as nest-building material. The four tasks studied have in common that it is possible to measure the amount of work performed per time by individual workers, without manipulation of colony composition, which may upset normal task allocation patterns. Of the colonies, four were large relative to the average colony size [58] in this species (147–233 workers) and seven were small (27–100 workers). These colonies were also used in other studies, in which colony size effects on individual workload in emigrations and the difference between “elite workers” and “specialized workers” were investigated ([12] and A. Dornhaus, J.-A. Holley, and N. R. Franks, unpublished data). In the study presented here, I focus on the quality of performance of individuals. For each individual ant, it was recorded how often it performed each task, how long it delayed before starting to perform the task (“delay,” see Materials and Methods), and how long it took the ant to perform two task units. A task unit (hereafter: “trip”) is defined as leaving the nest and returning to it while delivering either a brood item to a new nest site, a load of honey solution to a nestmate, a piece of fly (i.e., proteinaceous food) to the nest, or a stone to the wall being built. Thus, the duration per trip reflects how much work a worker accomplished per time, and therefore can be used as a performance measure. I calculated the average duration per trip for the first two trips of each ant that performed a task (hereafter called “performance”). For each ant and each task, a measure of specialization was also calculated: an ant was considered more specialized the more it concentrated its work effort in a single task. I used the proportion of total task performances (trips) that were in task X as a measure of specialization on task X. Thus, if all performances were in a single task, the worker's specialization for that task was 100%; if a task was never performed, specialization was 0%; and if all four tasks were performed equally frequently, specialization was 25%. Using these measures and ranking workers within each colony according to their performance, I find no correlation between specialization and performance for three tasks: brood transporting, honey foraging, and fly foraging (Figure 1). This means that in each colony, the workers that were the most specialized in a task were not necessarily the best performers. In the fourth task, collection of stones, I did find a significant impact of specialization on colony-level rank of performance (Figure 1; regression for large colonies: p = 0.03, R2 = 0.13, df = 27; small colonies: p = 0.01, R2 = 0.15, df = 36). However, this relationship was not in the direction predicted: workers with a high rank on specialization (mostly performing stone collection) had a high rank on duration of two trips, which means they were less efficient (took longer to perform the same amount of work). Note also that these results remain unchanged if workers who performed a task only once are excluded (large colonies: brood transports, p = 0.062; small colonies: brood transports, p = 0.38; honey foraging, p = 0.64; stone collection, p = 0.041; in no case were more than two workers excluded; if p-value is not given here, all workers performed the task at least twice or not at all).
Instead of specialization, it may be that the overall number of trips across all tasks predicts a worker's performance. To test this, I performed a stepwise regression of performance on three factors: overall activity level (total number of trips in all tasks performed by that worker), task-specific delay (time from start of experiment to first task performance), and specialization. None of these factors consistently predicted performance (Table 1). Only in one case (performance of brood transports in small colonies) was specialization a significant factor, although even here, activity, i.e., amount of work performed overall rather than in a specific task, was more predictive of performance, and only a small amount of the variation in performance was explained by either of these two factors.
Another puzzling result is that I do not find that the contribution a worker makes to overall colony workload is predicted by worker efficiency (Table 2). This means that in a given colony, most of the work is not necessarily performed by those who are best at it; it is, however, often performed by those most specialized in a task (Table 2). This agrees with a previous result that work in Temnothorax albipennis is generally performed by specialists, not generalists (A. Dornhaus, J.-A. Holley, and N. R. Franks, unpublished data).
Colonies did not differ significantly in how well their workers performed, except in the task of brood transports (Kruskal-Wallis Test, large colonies, transports: p = 0.024, df = 3, n = 115 workers; honey foraging: p = 0.47, n = 56, fly foraging: p = 0.38, n = 25; stone collection: p = 0.20, n = 28; small colonies, transports: p = 0.36, df = 6, n = 64 workers; honey foraging: p = 0.016, n = 32, stone collection: p = 0.26, n = 37). Median individual performance for colonies does not seem to depend on colony size (Figure 2A; brood transports: p = 0.15, R2 = 0.13; honey foraging: p = 0.89, R2 < 0.001; fly foraging: p = 0.58, R2 < 0.001; stone collection: p = 0.20, R2 = 0.08); variation, measured as the interquartile interval, among individuals also seems mostly constant even in different-sized colonies, although it was significantly higher in larger colonies for brood transports (Figure 2B; brood transports: p = 0.048, R2 = 0.30; honey foraging: p = 0.06, R2 = 0.33; fly foraging: p = 0.44, R2 < 0.001; stone collection: p = 0.69, R2 < 0.001).
Some individuals were only seen performing one task (35% of all workers in large, 32% in small colonies), and many were never seen performing any of the four tasks investigated here (52% of all workers in large, 57% in small colonies; see also [12] for inclusion of more tasks). Thus, the experimental conditions did not simply induce all workers to work at maximum level, which would have potentially obscured a normal relationship between specialization and individual performance. For workers whose performance was measured in at least two different tasks, their colony-specific rank in performance in one task did not correlate with that in the other task (Figure 3; Regression p = 0.44, R2 < 0.001, n = 62). As stated above, for each worker in each task, two task performances (trips) were measured: duration of the first trip correlates significantly with duration of the second trip for transports, but not for the other tasks (regression on colony-specific ranks, transports: p < 0.001, R2 = 0.19, n = 101; honey foraging: p = 0.07, R2 = 0.11, n = 21; fly foraging: p = 0.83, R2 < 0.001, n = 8; stone collection: p = 0.18, R2 = 0.05, n = 19). In honey foraging and stone collection, there was a trend in the same direction (higher rank in trip 1 ∼ higher rank in trip 2), but the sample sizes were lower than for transports; it thus cannot be said with certainty whether individuals were consistent in their performance over time.
Among the results presented here, two are particularly surprising: that colonies are not adapted to allocate the most efficient workers to each task, and that efficiency seems unrelated to the level of behavioral specialization of individuals. Many previous studies have simply assumed that if there is specialization, it will correlate with improved performance at a task (although see [22,23,59–61]). My results indicate that at least in this species, a task is not primarily performed by individuals that are especially adapted to it (by whatever mechanism). This result implies that if social insects are collectively successful, this is not obviously for the reason that they employ specialized workers who perform better individually. It also seems that individual performance (at least in the four tasks investigated) is not predicted by overall activity or that ant's delay to engaging in a task. Delay may or may not correspond to a task response threshold, an individual- and task-specific factor that defined the probability of engaging in a task. This factor has been used in many previous models of task allocation. It will be interesting for future studies to investigate whether response thresholds do or do not predict individual efficiency in performing tasks.
The performance measure here was the average duration of two individual trips, which corresponds to the number of items brought to the nest per time. Although this is a performance measure that is often used and can be objectively quantified, it is possible that the performance of specialist ants was superior to that of generalists in some other way. Perhaps specialists carried larger loads (although this seems unlikely in the nest-building case, as the sand grains were sieved to a uniform size), or perhaps specialists were able to collect more information or watch for predators while performing tasks. However, the time used per load, as measured here, varied by more than a factor of 40 (for example, fastest brood transport was 100 s, slowest 4,363 s). Although it cannot be excluded, it seems unlikely that these differences were compensated by load size or minimization of predation risk.
It is tempting to say that the ant species studied here, Temnothorax albipennis, is unusual in its colony organization. Maybe it employs less strict division of labor than other ant species (although other measurements indicate that this is not the case: A. Dornhaus, J.-A. Holley, and N. R. Franks, unpublished data), or maybe because of their long lifespan (workers can live several years in the lab), each individual already has had the opportunity to perfect its performance in each task. Also note that, as stated above, T. albipennis does not have worker polymorphism, so any differences in specialization among workers are the result of behavioral specialization only. However, the level of specialization in most social insect species is not known, and it can be argued that Temnothorax is representative of the majority of ant species: it has the same small colony sizes that are typical for most ants [1,62]; it forages by preying on and scavenging other arthropods in the leaf litter, as many other ants do [1,21]; it is monomorphic (no allometry among workers) as most other ants are [21,42]; and the genus Temnothorax is cosmopolitan and does not consist of ecological specialists adapted to particularly restricted habitats. To test whether the present results are widely applicable throughout the social insects, it would be desirable if future research employed a wide variety of study systems. That would enable an assessment of how widespread, across species, individual behavioral specialization is, and how it relates to efficiency.
In addition, only four tasks that ants perform on a regular basis were studied here. There are a number of other relevant tasks, most notably brood care and colony defense against predators and parasites. Efficiency assays for these tasks should be developed and used to study the benefits of specialization. Although studying other tasks is important, the tasks studied here were previously thought to be the ones that are most likely to be influenced by learning and thus suitable for specialization [30,48,63]. Tasks that involve leaving the colony require skills of orientation and the specific learning of landmarks [64,65]; even in small laboratory settings, such learning can significantly affect performance [66]. Similarly, tasks that involve collecting prey or building material involve identification and handling skills that cannot be easily genetically preprogrammed, as the precise location and type of prey and building material is likely to vary with microhabitat, even within a population (e.g., [34,63,67–69]).
Indeed the results presented here do not show that learning is absent in this species. Learning, in the context of task performance, may theoretically occur at three time scales. Short-term learning may increase performance from one trip to the next on the same day, without leading to long-term individual differences. Second, performance at foraging and emigration tasks may differ primarily between completely naive individuals who have never left the nest and individuals who have left at least once, i.e., participated in at least one emigration or foraging bout (or perhaps performed the equivalent of the well-studied “orientation flights” in honey bees, e.g., [51,64]). Third, amount of experience may directly correlate with performance, such that the more experience an individual gains at a specific task over its lifetime, the better it is able to perform a task. The results in this study show that there is no correlation between quality of performance and specialization—suggesting that differential improvement through learning, as in the third type of learning listed above, does not occur. However, it is quite likely that the first two types of learning do occur. Previous studies have demonstrated learning in colony emigrations and foraging in this species [70–75]. However, in many cases the main improvements were achieved after the first performance of the task (although see [69] for honey bees), suggesting an effect of learning similar to the second type discussed above.
In summary, this study finds that there is a large amount of variation in individual quality of task performance, not explained by any of the factors studied. The mechanism creating this variation is unknown, and may be genetic, developmental, or an effect of experience. Thus, learning may well affect task performance, but either it affects all individuals equally, or workers do not preferentially perform the tasks in which they are experienced (although the latter would contradict previous studies: [71,76,77]). Either way, learning does not seem to lead to superior performance by specialists. It will be important for future research to quantify at what time scales learning occurs, and whether it increases or decreases variance among individuals in the long term.
In this study, I quantify quality of task performance for individual workers in several tasks. To show that specialization exists, it is necessary to show that an individual performs more of one task, and less of another, compared to nestmates [10,11]. It is not sufficient to measure how much an individual performs a single task: this may merely identify high-activity workers from low-activity workers. To measure the benefits of specialization (at least in terms of individual efficiency or quality of task performance), it is equally necessary to compare performance in multiple tasks; otherwise one may simply identify high-quality workers from low-quality ones, without necessarily demonstrating that specialists are better at their task and worse (or at least no better) at another. If this is not the case, then one has demonstrated merely that there is variation in quality of task performance among individuals, and possibly that high-quality individuals tend to be allocated to particular tasks, but not that there are benefits to specialization.
What, then, are the benefits of division of labor in species without polymorphic workers? As mentioned above, there are at least three potential benefits of division of labor. Individually increased efficiency was only one of them. Others were a decrease in the costs associated with switching between tasks. For example, division of labor may lead to increased spatial efficiency, as hypothesized for ants [28], or reduction of other, possibly cognitive, switching costs [57]. It is also possible that specialization simplifies the process of task allocation (i.e., minimizes neural or other costs associated with the task selection process itself), or optimizes material flow in multistep tasks, [15]. Any of these processes may create colony-level fitness benefits from division of labor, even without improvement in individual efficiency as measured here. Future studies should attempt to quantify switching costs, spatial constraints, the role of learning, and the time scales at which individuals specialize in social insect colonies. My study also highlights that findings from commonly used model species, such as honey bees or leaf-cutting ants, which have very unusual and specific ecology and morphology, cannot necessarily be extended to other species [20,21,78]. We have much yet to learn about the benefits and evolution of division of labor.
As mentioned above, all workers in 11 colonies of the ant species Temnothorax albipennis, collected in Dorset, England, were individually marked with paints (a total of 1,142 ants were marked; details on this method, as well as colony collection and housing, can be found in [12]). The colonies were housed in artificial nests in the laboratory, made of a cardboard perimeter sandwiched between two glass slides. All colonies were filmed in three different contexts spaced at least a week apart (emigrations, wall building, and foraging). Each colony was filmed for at least 180 min in each context, starting at the time of manipulation as described below. This resulted in 166 h of digital video tape. Each context was initiated as follows: emigration—removal of the top glass slide, exposing the ants (a new, identical nest was offered in 10 cm distance); foraging—colonies were starved for 2 wk (no food, but water ad lib was offered), and then a small dish with honey solution (1:10 honey:water) and a pile of ten frozen Drosophila flies were placed 10 cm from the nest entrance; building—colonies were housed in a nest that had no front wall, creating a 33-mm-wide gap; on the day after the ants had moved into this nest, and a pile of colored and sieved sand grains was offered 10 cm from the nest. Under the latter conditions, ants use the sand grains to build a wall to narrow the nest entrance to approximately 1–3 mm. No food was offered in the emigration and building contexts, and no building material was offered in the emigration and foraging contexts.
Each of the three different contexts thus introduced the need to perform a particular set of tasks, creating an opportunity for each ant to take part in these tasks. By using separate contexts, ants' task choices were thus less affected by competing stimuli for different tasks, but solely by the individual's preferences for performing the task at hand. For example, if some individuals had both a high tendency to transport brood and to participate in nest building, their activity level in either of these tasks was not constrained by that in the other task. If all tasks had been offered at once, such individuals may have spent all their time transporting brood simply because it is the more urgent task. Only a separation of tasks as employed here allows the identification of specialists from highly active generalist individuals.
From the video tapes, the time that each ant picked up a brood item in the old nest, a sand grain from the pile, or left the nest in a foraging trip was extracted (start of trip). The time from the start of the experiment (e.g., removal of the nest cover in emigration experiments) to the start of the first trip for each ant was designated its task-specific delay. Then, the time that the same ant returned to the old nest, dropped the sand grain at the nest, returned and performed trophallaxis (to unload honey solution to a nestmate), or returned with a piece of dead fly to the nest was recorded (end of trip). The time difference between start and end of trip give the trip duration. The sum of all trips made in one context by one colony was called the colony-level workload (e.g., total number of brood transports made in an emigration). An individual's colony-level work contribution was measured as that individual's number of trips divided by the total colony-level workload. All measurements were double-checked by a second person to ensure accurate records of ant identity and timing of task performances.
|
10.1371/journal.pgen.1000557 | Pch2 Links Chromosome Axis Remodeling at Future Crossover Sites and Crossover Distribution during Yeast Meiosis | Segregation of homologous chromosomes during meiosis I depends on appropriately positioned crossovers/chiasmata. Crossover assurance ensures at least one crossover per homolog pair, while interference reduces double crossovers. Here, we have investigated the interplay between chromosome axis morphogenesis and non-random crossover placement. We demonstrate that chromosome axes are structurally modified at future crossover sites as indicated by correspondence between crossover designation marker Zip3 and domains enriched for axis ensemble Hop1/Red1. This association is first detected at the zygotene stage, persists until double Holliday junction resolution, and is controlled by the conserved AAA+ ATPase Pch2. Pch2 further mediates crossover interference, although it is dispensable for crossover formation at normal levels. Thus, interference appears to be superimposed on underlying mechanisms of crossover formation. When recombination-initiating DSBs are reduced, Pch2 is also required for viable spore formation, consistent with further functions in chiasma formation. pch2Δ mutant defects in crossover interference and spore viability at reduced DSB levels are oppositely modulated by temperature, suggesting contributions of two separable pathways to crossover control. Roles of Pch2 in controlling both chromosome axis morphogenesis and crossover placement suggest linkage between these processes. Pch2 is proposed to reorganize chromosome axes into a tiling array of long-range crossover control modules, resulting in chiasma formation at minimum levels and with maximum spacing.
| In the germ line of sexually reproducing organisms, haploid gametes are generated from diploid precursor cells by a specialized cell division called meiosis. Reduction by half of chromosome numbers during the first meiotic division depends on genetic exchange, resulting in the formation of crossovers. Without crossovers, pairs of homologous chromosomes frequently fail to separate, resulting in unbalanced gametes with a surplus or deficit of individual chromosomes. Along a given chromosome, crossovers form in different locations in different cells, but distribution of crossovers within each cell is controlled in two ways: first, at least one crossover is formed along each homolog pair, irrespective of size; second, a crossover in a given interval reduces the frequency of crossovers in adjacent chromosome regions. Here, we identify functions of the evolutionarily conserved protein Pch2 in suppressing additional crossovers in adjacent regions and ensuring homolog segregation under certain conditions. Pch2 further controls the assembly of chromosome axis protein Hop1 at future crossover sites. Our findings reveal that chromosome axes undergo structural changes at the same positions where crossovers occur. Thus, axis remodeling and crossover placement are linked via Pch2.
| During meiosis, a single round of DNA replication is followed by two rounds of chromosome separation, with homologous chromosomes (homologs) segregating during meiosis I and sister chromatids during meiosis II. Homolog segregation critically depends on formation of crossovers (COs) between homologs. COs, cytologically detectable as chiasmata, in combination with sister chromatid connections, mediate the correct positioning of homolog pairs in the meiosis I spindle. Without COs, homologs frequently fail to segregate, resulting in formation of aneuploid gametes, i.e. gametes with a chromosome surplus or deficit. Aneuploid gametes are one of the major causes for stillbirths and birth defects in humans [1].
CO formation occurs via a carefully orchestrated program during prophase of meiosis I entails close homolog juxtaposition, followed by reciprocal exchange of chromosome arms through homologous recombination [2]. On the DNA level, meiotic recombination is initiated by formation of programmed double strand breaks (DSBs) at multiple genome positions [3]–[5]. A non-random subset of DSBs undergoes stable interaction with a homologous chromatid, giving rise to COs, while the remainder of DSBs progress to alternative fates, including non-crossovers (NCOs), i.e. recombination events without exchange of flanking chromosome arms, as well as repair events with the sister chromatid [6]–[8].
Studies in fungi including S. cerevisiae have provided an understanding of meiotic recombination at the molecular level. Processing of meiotically induced DSBs depends on numerous proteins with related roles in mitotic DSB repair, but there are also prominent differences between these processes: First, during meiosis, homologs rather than sister chromatids serve as partners for homologous recombination [6]. Second, CO formation is enhanced over that of NCOs [7]. Following 5′ resection, DSBs undergo strand invasion of intact non-sister homologous chromatids. Pathways leading to COs and NCOs appear to bifurcate no later than the stage of strand invasion: Single end invasions (SEIs) emerge as the first CO-specific intermediate, subsequently giving rise to double Holliday junctions which are specifically resolved as COs [9]–[11]. NCOs likely arise via an alternative pathway characterized by a more transient strand invasion [7]. Notably, only COs provide interhomolog connections as required for homolog segregation.
Recombination is temporally and spatially coordinated with dramatic changes in global chromosome structure culminating in the assembly of the synaptonemal complex (SC). The SC, a widely conserved proteinaceous structure, stably juxtaposes homologs along their entire lengths during the pachytene stage [12]. SC formation is initiated during the leptotene stage when axial elements first form between and along sister chromatids. During the zygotene stage, axial elements of homologous chromosomes become closely juxtaposed via the SC central element which starts polymerizing from discrete sites; achieving full length homolog synapsis during the pachytene stage. Recombination is initiated via induction of DSBs during the leptotene stage, followed by onset of strand invasion at the transition from the leptotene to the zygotene stage [9]. During the pachytene stage, in the context of fully formed SC, double Holliday junctions are formed and resolved into COs, with NCOs emerging somewhat earlier [9]–[11].
Morphogenesis of the SC and recombination are highly interdependent, as indicated by (i) requirements for recombination proteins for SC assembly, and (ii) functions of SC components in recombination. In S. cerevisiae, DSBs are introduced by the widely conserved topoisomerase homolog Spo11 [5]. Spo11-dependent DSB formation is also required for SC assembly. Prominent components of yeast axial elements include Hop1 and its binding partner Red1, as well as meiosis-specific cohesin Rec8 and cohesin-associated proteins, e.g. Spo76/Pds5 [13]–[16]. Hop1 and Red1 further mediate normal DSB formation and preferential interaction of DSBs with homologs rather than sister chromatids [6], [17]–[19]. Zip1 is a prominent component of the SC central element. Zip1 starts polymerizing from both centromeres and from positions of designated CO sites [20]–[24]. Prior to assembly into full length SC, Zip1 mediates timely and efficient CO-specific strand invasion during recombination [11]. Two ZMM proteins, Zip2 and Zip3, are required for formation of most COs and also mediate normal SC assembly. Zip3 is present along fully formed SC with the number and distribution expected for CO designated sites, in S. cerevisiae and C.elegans [21],[22],[25],[26]. Finally, regions surrounding emerging COs are structurally modified as suggested by localized separation of sister chromatids at sites of ongoing recombination [16]. Later, when chiasmata emerge, they are characterized by extended regions of sister axis separation flanking the position of an established CO (see example in ref. [27]).
CO placement along homolog pairs is non-random at several levels: First, CO assurance guarantees formation of at least one CO per bivalent (e.g. ref. [28]). Second, CO homeostasis enhances CO formation at the expense of NCOs when initiating DSBs are artificially reduced [29]. Third, CO interference reduces the frequency of COs in regions adjacent to established COs resulting in maximally spaced COs [30].
The three levels of CO control indicate communication along chromosomes between sites of ongoing recombination. CO interference reduces CO frequencies over large physical distances, >100 kb in yeast and >100 Mb in higher eukaryotes [8],[31]. CO assurance and CO homeostasis suggest mechanism(s) that sense overall CO and/or DSB levels, affecting the outcome of ongoing recombination events. Timing, mechanism and the functional relationship between CO control and meiotic recombination pathway(s) are poorly understood. CO control is thought to operate on randomly distributed recombination interactions, a non-random subset of which become designated as future COs with the remainder progressing to NCOs. CO designation likely occurs no later than zygotene, as suggested by occurrence of cytological markers of CO-designation at this stage, and by concurrent appearance of CO specific recombination intermediates [11]. Linkage between CO assurance and CO interference was inferred from coordinate loss or retention of both features in certain mutant situations [29],[32]. Conversely, CO interference is retained in two zmm mutants (zip4Δ, spo16Δ) despite apparent loss of CO assurance, indicating that separable pathways contribute to CO control [28]. Structural chromosome components responsible for CO control also remain unknown. Normal interference distribution of CO-designation marker Zip2 in zip1Δ suggests that the SC central element is not required for crossover interference [22]. Notably, in zip1Δ, CO designation sites/Zip2 foci exhibit interference distribution, while CO interference is defective, indicating uncoupling between chromosome morphogenesis and events on the DNA level [22].
The widely conserved AAA+ ATPase Pch2 performs important functions in cell cycle control, recombination and chromosome morphogenesis during mutant and WT meiosis. Identified as a yeast mutant that bypasses meiotic arrest in zip1Δ, Pch2 also mediates mutant delay/arrest in C. elegans and Drosophila [33]–[38]. During yeast WT meiosis, Pch2 mediates timely resolution of double Holliday junctions and formation of COs and NCOs [35],[37]. Processing of a subset of recombination intermediates also depends on Pch2 in mouse [36]. In yeast, Pch2 further mediates assembly of structurally normal SC, controlling installation of axis component Hop1 and SC central element protein Zip1 along meiotic chromosomes in a pattern of alternating hyperabundance [37]. This pattern likely arises due to uniform loading of Hop1 and Zip1 at base levels along the length of the SC, corresponding to the uniform appearance of the SC detected by electron microscopy, in combination with additional domainal loading of either protein. Absence of Pch2 results in uniform localization patterns of Hop1 and Zip1 along the length of meiotic chromosomes [37].
Here, we have investigated the interplay between meiotic chromosome morphogenesis and CO control in yeast. We demonstrate intimate coordination between controlled CO distribution and axial element morphogenesis, as suggested by frequent association between Zip3-marked CO-designation sites and domains of preferential Hop1/Red1 loading. Association between Zip3 and Hop1/Red1 becomes detectable prior to substantial SC polymerization, consistent with axis differentiation at future CO sites early during meiosis. Furthermore, Hop1-Zip3 association is detected in ndt80Δ-arrested cells indicating its establishment independent of and prior to double Holliday junction resolution. Pch2 controls chromosome axis status by (i) specifying amount and pattern of chromosomal Hop1, (ii) limiting Zip3 positions along pachytene chromosomes and (iii) mediating global axis shortening. While competent for CO formation at normal levels, pch2Δ is defective in controlling the distribution of COs along chromosome arms. In pch2Δ, (i) CO interference is defective, and (ii) spore viability is drastically reduced upon global reduction of initiating DSBs. The pch2Δ phenotype is dramatically modulated by incubation conditions, including temperature, suggesting the existence of Pch2-independent back-up systems for crossover interference and for maintenance of normal spore viability despite reduced DSB levels. We propose a model where Pch2 mediates establishment of multiple CO control modules along each chromosome, with potential effects on CO interference and chiasma function.
Pch2 mediates domainal hyperabundance of axis protein Hop1 along pachytene chromosomes [37]. Loss of domain structure in pch2Δ during zygotene suggests Pch2 functions at or before this stage. To examine Pch2 localization throughout meiosis I prophase, an isogenic SK1 strain homozygous for N-terminally 3×HA-tagged Pch2 ( = HA-Pch2) was induced to undergo synchronized meiosis at 33°C. The 3×HA-tagged Pch2 construct used here complements Pch2 function as suggested by its ability to confer arrest in zip1Δ. It is identical to a construct previously examined in a different strain background (data not shown, A. Hochwagen, personal communication, see Materials and Methods for details; [33]).
Pch2 localization was examined at all stages of meiosis I prophase. At specified time points, cells were surface spread and immunodecorated with antibodies against the HA-epitope, and SC central element component Zip1 [20]. Cells progressed through meiosis with appropriate timing [11]: Late leptotene nuclei, containing <10 Zip1 staining foci, first appear at 2 hrs (Figure S1). Zygotene nuclei carrying multiple Zip1 foci (“early zygotene”) and/or Zip1 in partial lines (“late zygotene”) are prominent at 4 to 5 hrs (Figure 1A, 1AE, and 1I). Pachytene cells exhibiting mostly continuous lines of Zip1 along most of the 16 homolog pairs reach peak levels between 4 to 7 hours and disappear shortly before the onset of nuclear divisions (Figure 1M and Figure S1).
In pachytene nuclei, multiple Pch2 foci of comparable intensities are detected on most of the chromatin mass (Figure 1M–1P). The majority of Pch2 foci is associated with Zip1, but cells frequently also contain ∼five grouped foci in a crescent-shaped, Zip1-free chromatin region, likely corresponding to the nucleolus (Figure 1K and 1O; see ref. [33]). In pachytene nuclei, 21 (±7 S.D.) Pch2 foci residing outside the nucleolus (referred to as chromosomal Pch2 hereafter) are detected (n = 149 nuclei; Figure 1O). Pch2 is also present at chromosomal and presumed nucleolar positions in early and late zygotene nuclei where 16 (±5 S.D.) Pch2 foci are detected (n = 25 nuclei), some of which localize to several Zip1-free regions, suggesting localization to unsynapsed chromosomes. Together, these data indicate that Pch2 starts localizing abundantly to chromosome arms during the early zygotene stage, reaching maximum levels during the pachytene stage. Nucleolar and chromosomal Pch2 staining exhibits comparable intensities here, yet appears more prominent in the nucleolus in an earlier report [33]. Such differences could be due to effects of different spreading protocols and/or imaging systems.
Pch2 promotes timely formation of recombination products [37], and plays roles in CO control (see below). Zip3 is a cytological marker for CO-designated sites, forming interference-distributed foci along pachytene chromosomes with numbers corresponding to COs [22]. To examine localization of chromosomal Pch2 with respect to ongoing recombination interactions, meiosis was induced in strains homozygous for HA-Pch2 and C-terminally GFP-tagged Zip3. (ZIP3-GFP complements ZIP3 function as suggested by spore viabilities >85%, normal CO levels by physical analysis and WT-like progression through meiosis (G.V.B. and O. Nanassy, unpublished)).
Cells from a synchronous time course carried out at 33°C were spread and stained with appropriate antibodies. Anti-Zip1 antibody was used to determine stages of cells. Number and localization patterns of Zip3 in zygotene and pachytene nuclei correspond well with earlier reports [21]–[23]. Pachytene nuclei contain 61 (±6 S.D.) Zip3 and 31 (±12 S.D.) Pch2 foci (n = 42 nuclei; see Figure 2E–2H). Importantly, a substantial number of Pch2 foci colocalizes with Zip3: In pachytene nuclei, 54% (±17% S.D.) of Pch2 foci are associated with Zip3 foci, compared to 18% (±10% S.D.) fortuitous colocalization (n = 17 nuclei; see Materials and Methods for details on analysis of fortuitous colocalization). Colocalization of Pch2 with Zip3 is also observed in zygotene nuclei where 44 (±16 S.D.) Zip3 and 25 (±12 S.D.) Pch2 foci are detected (n = 28 nuclei; Figure 2A–2D). Of the Pch2 foci detected, 58% (±18% S.D.) colocalize with Zip3, compared to 13% (±10% S.D.) fortuitous colocalization (n = 13 nuclei). Similar localization patterns are observed in the same strain incubated at 30°C (data not shown). Together, these data demonstrate that chromosomal Pch2 partially and/or transiently associates with Zip3-marked CO-designated sites. This association could be related to Pch2's function in CO placement and/or CO-associated domain organization (see below).
To gain insights into positional identities of Hop1-enriched axis domains, Zip3 and Hop1 localization were examined in a synchronous WT time course at a time when pachytene cells are abundant [11]. In WT, at t = 7 hrs, >50% of undivided nuclei are at the pachytene stage, as indicated by Zip1 staining patterns (data not shown). In the same cell population, Hop1 and Zip3 localization are remarkably similar in number and position: Hop1 localizes to 55 (±13 S.D.) foci while Zip3-GFP localizes to 56 (±13 S.D.) foci per nucleus (Figure 3A–3D; n = 68 nuclei). When Hop1 and Zip3 localization patterns in the same nuclei are compared, a striking correspondence in position emerges: 72% (±10% S.D.) of Zip3 foci colocalize with Hop1, and 73% (±15% S.D.) of Hop1 foci colocalize with Zip3. Fortuitous colocalization in the same nuclei is 17% (±7% S.D.) and 17% (±8% S.D.), respectively. These results suggest that CO-designated recombination interactions frequently localize to chromosome domains enriched for Hop1.
To examine Zip3 localization relative to another axis protein and to determine the stage of meiosis in the same cells, spread nuclei were triple-stained for Red1, Zip3, and Zip1 in a strain homozygous for C-terminally HA-tagged Red1 (Red1-HA) [17] and Zip3-GFP. High levels of colocalization between Zip3 and Red1 were observed at both the zygotene and pachytene stages (Figure 3E–3H). In zygotene nuclei, 57 (±16 S.D.) Red1 foci and 50 (±19 S.D.) Zip3 foci are detected: 65% (±10% S.D.) of Zip3 foci colocalize with Red1, and 55% (±14% S.D.) of Red1 foci colocalize with Zip3 (n = 72 nuclei; Figure 3I–3M). In pachytene nuclei, Red1 localizes to 53 (±11 S.D.) foci, and Zip3 to 55 (±15 S.D.) foci; 59% (±10% S.D.) Zip3 foci colocalize with Red1, and 60% (±14% S.D.) Red1 foci colocalize with Zip3 (n = 57 nuclei; Figure 3N–3R). Thus, Red1 is also frequently associated with chromosome regions designated to undergo CO formation.
Together, these results have two key implications: Association of Zip3 and Hop1/Red1 (i) at the same sites along pachytene chromosomes suggests spatial linkage between Hop1-enriched domains and CO placement, and (ii) temporal coincidence with CO/NCO differentiation during the zygotene stage [10],[11].
We note that not all Zip3 foci are associated with Hop1/Red1 in every cell. This association may be transient and/or only a subset of Zip3 associates with Red1/Hop1. Furthermore, Zip3 occupies presumed CO designation sites only during the pachytene stage, while it localizes to centromeres in pre-zygotene cells [22],[24]. In pre-zygotene cells, Zip3 is detected at small numbers and rarely colocalizes with abundantly staining Hop1 or Red1 (data not shown).
We next investigated Hop1 and Zip3 localization in dmc1Δ and ndt80Δ, two meiotic mutants exhibiting distinct recombination blocks: In the absence of Rad51-paralog Dmc1, hyperresected DSBs accumulate, and COs and NCOs are eliminated, consistent with a role of Dmc1 in strand invasion (G.V.B., unpublished data; ref. [39]). Without transcription factor Ndt80, NCOs appear normally, but double Holliday junctions accumulate and COs are reduced accordingly [10]. Cells further arrest in ndt80Δ at a bona fide normal pachytene stage, as suggested by formation of viable spores upon induction of Ndt80 [40].
In dmc1Δ at 33°C, Zip3 localizes to nuclei abundantly, although at reduced numbers. At a time when most cells have completed DSB formation (t = 5 hrs; G.V.B., unpublished data), 33 (±7 S.D.) Zip3 foci are detected (n = 58 nuclei), compared to ∼50 Zip3 foci in WT cells (Figure 4A–4C and 4P). In dmc1Δ, maximum Zip3 localization is reached at t = 5 hrs, as indicated by comparable numbers of foci at t = 4 and t = 6 hrs (data not shown). Thus, Dmc1 is required for association of Zip3 with meiotic chromosomes at normal levels. Hop1 also localizes at high levels to meiotic chromosomes in dmc1Δ, but poor spreading in these cells interferes with quantitation of Hop1 foci. Of the Zip3 foci detected in dmc1Δ, 77% (±11% S.D.) colocalize with Hop1, compared to 26% (±8% S.D.) fortuitous colocalization. About 10% of dmc1Δ nuclei further exhibit WT-like patterns of Zip3 staining, with several Zip3 foci located in a linear array, consistent with staining along condensed chromosome axes. In these nuclei, Zip3 again colocalizes with-Hop1 at high levels (Figure 4A–4C). In summary, Zip3 foci form with reduced numbers in dmc1Δ, but tend to be associated with Hop1.
In ndt80Δ at 33°C, at a time when most cells have undergone pachytene arrest (t = 8 hrs), Zip3 and Hop1 localize to meiotic chromosomes with patterns and numbers similar to wild-type pachytene nuclei (compare Figure 4J and 4K with Figure 4D and 4E). Both Zip3 and Hop1 are detected as foci, with 50 (±9 S.D.) Zip3 foci and 63 (±8 S.D.) Hop1 foci detected (n = 51 nuclei). Colocalization between Hop1 and Zip3 is also high in ndt80Δ, with 76% (±11% S.D.) of Zip3 foci colocalizing with Hop1, similar to the WT pachytene stage (n = 51 nuclei; Figure 4F and 4L).
We conclude that Dmc1 is required for normal levels of both Zip3 localization and Hop1-Zip3 co-staining domains. Importantly, association of Zip3 and Hop1 is independent of NDT80, indicating that it is established prior to and independent of double Holliday junction resolution into COs.
Pch2's role in chromosome morphogenesis was examined in more detail by analyzing patterns and levels of Hop1 localization in ndt80Δ-arrested cells. Detection of 63 (±8 S.D.) Hop1 foci in ndt80Δ confirms that Hop1 localizes as foci rather than in lines along pachytene-arrested chromosomes (Figure 4K). Conversely, in both NDT80 and ndt80Δ nuclei, with maximized visualization of near-background signals, Hop1 foci frequently coalesce into lines, consistent with continuous localization of Hop1 at base levels along pachytene chromosomes (data not shown).
Absence of Pch2 affects Hop1 patterns similarly in ndt80Δ and NDT80 (compare Figure 4H and 4N; ref. [37]): Hop1 localizes as continuous, mostly uniform lines along the 16 homolog pairs. Quantitative analysis further identifies roles of Pch2 in controlling both Hop1 loading levels and patterns: Hop1 signal intensities are ∼three-fold increased in pch2Δ (p<0.0001; see Figure S2A for details). The Hop1 staining observed in pch2Δ could be due to a uniform increase exclusively or concurrent Hop1 redistribution. To examine this question, number and contour length of high intensity Hop1 signals were determined in 20 WT and pch2Δ nuclei (see Materials and Methods for details). In pch2Δ, strong Hop1 signals exhibit ∼two-fold increased average contour lengths and are present at reduced numbers (see Figure S2B, S2C; p<0.0001; two-sided Wilcoxon rank sum test). If extra loading had occurred universally, patterns of strong Hop1 signals would be similar in WT and pch2Δ. We conclude that the more uniform Hop1 signal in pch2Δ is due to an overall increase in Hop1 signal intensities concurrent with changes in Hop1 loading patterns.
Together, these findings have four important implications. (i) Hop1 is a prominent component of pachytene SC. (ii) In WT, Hop1 is present along chromosome axes in a mostly continuous pattern at base levels, with hyperabundance at distinct chromosome domains [37]. (iii) Pch2 controls both overall levels and patterns of Hop1 localization along meiotic chromosomes. (iv) Changes in Hop1 localization are caused by absence of Pch2, rather than being an indirect result of meiotic arrest.
Next, the role of Pch2 in controlling Zip3 association with chromosomes was investigated in WT and pch2Δ at a time point exhibiting maximum levels of pachytene cells (>50%; t = 7 hrs) as well as in ndt80Δ-arrested cells (T = 33°C): In WT, Zip3 and Hop1 predominantly localize as distinct foci (Figure 4D–4F; see above). In pch2Δ (t = 7 hrs), by contrast, Hop1 localizes in continuous lines and Zip3 foci are occasionally not well separated (Figure 4G; see above; ref. [37]). Further, in pch2Δ, Hop1 (but not Zip3) is detected in the nucleolus (Figure 4H) [33].
In WT nuclei, 56 (±1.7 S.E.) Zip3 foci are detected along meiotic chromosomes (n = 68, see above), while in pch2Δ, the average number of Zip3 foci per nucleus is 62 (±1.5 S.E.) (n = 64 nuclei; Figure 4D, 4G, and 4P). Accordingly, the number of Zip3 foci in pch2Δ is significantly increased (P = 0.0026, two-sided Wilcoxon rank sum test). To exclude possible effects of differences in meiotic progression, the number of Zip3 foci in WT and pch2Δ was also examined in the ndt80Δ background. In PCH2ndt80Δ, 50 (±1.3 S.E.) Zip3 foci are detected, compared to 67 (±1.3 S.E.) Zip3 foci pch2Δndt80Δ, reflecting an increase by 34% (Figure 4P). Again, this increase is statistically significant (p<0.0001, two-sided Wilcoxon rank sum test). Thus, Pch2 controls the number of Zip3 foci along pachytene chromosomes. Notably, increased numbers of Zip3 foci are not caused by accumulation of cells at the pachytene stage in pch2Δ: In ndt80Δ arrested cells, the number of Zip3 foci is substantially increased in pch2Δ compared to the corresponding PCH2 strain, indicating that Pch2 controls the number of Zip3 association sites.
To examine effects of pch2Δ on chromosome axis length, SC contour length was measured by staining for Hop1 and Zip1. In WT pachytene nuclei (identified based on Zip1 staining), Hop1 and Zip1 preferentially localize to alternating domains, whereas largely overlapping localization patterns are observed in pch2Δ pachytene cells (Figure 4S and 4V) [37]. The combined Hop1/Zip1 SC contour length of an entire chromosome complement is 34 µm (±0.9 µm S.E.) in WT, in accordance with published results (see ref. [41]). In pch2Δ, the axis length is increased by 18% to 40 µm (±1.2 µm S.E.), representing a significant increase (p = 0.00023, two-sided Wilcoxon rank sum test; see Figure 4W). Thus, homolog axes fail to shorten appropriately in the absence of Pch2.
Roles of Pch2 in controlling the number of Zip3-marked presumed CO-designated sites, Hop1's localization to the same regions, and meiotic chromosome axis length as well as Pch2's role in CO interference have important implications for the mechanism of CO control (Discussion).
We examined the roles of Pch2 in recombination in an interference tester strain carrying twelve pairs of heterozygous markers, defining nine genetic intervals along three homologs (designated as intervals 1 to 9 in Figure 5A), [42]. Chromosomes III, VII, and VIII represent small, large and intermediately sized yeast chromosomes, respectively. Marked regions span physical distances of 132 kb, 229 kb and 106 kb, corresponding to WT map distances of 43 cM, 66 cM and 47 cM, respectively (Figure 5A; below).
Map distances were determined using tetrads with four viable spores and Mendelian (i.e. 2∶2) segregation at a given pair of markers: Three types of tetrads can be distinguished (Figure 5B, boxed region): (i) All four spores exhibit parental marker combinations, giving rise to a parental ditype (PD); (ii) two spores are parental and two recombinant, constituting a tetratype (TT); (iii) all four spores carry nonparental marker configurations, constituting a nonparental ditype (NPD). The majority of PDs are derived from tetrads where no CO has occurred. TTs preferentially arise from tetrads that have undergone a single CO, while NPDs are derived from double COs involving all four chromatids within an interval (Figure 5B).
Double COs involving two or three chromatids give rise to PDs or TTs, respectively, and are indistinguishable from tetrads involving a single or no CO (Figure 5B, lower part). Thus, total frequencies of double COs are extrapolated from NPD frequencies [43]. Note that this formula assumes absence of chromatid interference which has been validated for WT and pch2Δ (data not shown).
Following meiosis at 33°C, tetrads from WT and pch2Δ strains were dissected and markers were scored. Dissection of WT and pch2Δ asci gave rise to >1200 four spore-viable tetrads for each strain. In WT, genetic distances are similar to those previously reported (Figure 5C; Table 1), [42]. Map distances are remarkably similar between WT and pch2Δ, with two of nine intervals in pch2Δ exhibiting a significant increase (intervals 1 and 3). We note an apparent increase in NPD frequencies in pch2Δ. Accordingly, in pch2Δ double COs when calculated separately, contribute disproportionally to total map distances in seven intervals (see Figure 5C, no differences in intervals 4 and 9).
Thus, Pch2 is not required for formation of COs at normal levels in most genome regions, consistent with prior physical analysis at a particular recombination hotspot [35],[37]. However, Pch2 appears to limit the occurrence of closely spaced double COs.
Increased levels of double COs in pch2Δ raise the question of Pch2's role in CO control. Modified coincidence analysis and analysis of NPD frequencies were used to determine effects of pch2Δ on CO interference using the tetrad set generated at 33°C. In modified coincidence analysis, map distances for each test interval are determined for two distinct tetrad subsets [31]: Subset P includes tetrads with parental marker configuration at an adjacent reference interval (PD; Figure 6A, left column). Subset N includes tetrads with non-parental marker configuration at the reference interval (TT or NPD; Figure 6A, right column; Table S1).
Map distances derived from subset P are remarkably similar between WT and pch2Δ, with only a single interval exhibiting a significant increase in pch2Δ CO frequencies (interval 5P6; Figure 6A, left panel). In contrast, map distances derived from subset N are strikingly different between WT and pch2Δ (Figure 6A, right panel): In six out of 12 adjacent interval pairs, map distances are significantly increased in pch2Δ compared to WT. Thus, Pch2 has no detectable effect on CO frequencies along an interval when the adjacent interval is parental, but suppresses CO formation in the same interval when the adjacent interval is recombinant. Notably, total map distance increases in intervals 1 and 3 in pch2Δ can entirely be attributed to subset N tetrads, while map distances in subset P tetrads are indistinguishable between WT and pch2Δ (Figure 5C; Figure 6A, left panel).
Numerous mutants defective for CO interference also exhibit intermediate to severe defects in CO assurance, as suggested by frequent occurrence of tetrads with two viable or zero viable spores due to homolog nondisjunction (e.g. [28],[32]). Such patterns of spore viability are frequently associated with elevated levels of non-exchange chromosomes [32]. WT-like patterns of spore viability in pch2Δ provide no indication of increased homolog nondisjunction: Overall spore viability in pch2Δ at 33°C is 84.0% compared to WT viability of 82.6%, consistent with normal homolog disjunction in pch2Δ. WT like levels of spore viability are also observed in pch2Δ at 30°C (see Figure 7 and Figure 8B: SPO11/”, black bars; [33]).
Low levels of non-exchange homolog pairs could be rescued by a backup system that mediates disjunction of non-exchange chromosomes reducing the reliability of spore viability as a measure for CO assurance [45]. To directly evaluate whether chromosomes receive similar numbers of COs in WT and pch2Δ, tetrads formed at 33°C were therefore individually inspected for the number of COs along three pairs of homologs. Tetrads with no CO in the monitored interval are only marginally increased in pch2Δ, by 1%, 6% and 8%, suggesting that similar numbers of COs are formed along a given interval in WT and pch2Δ (Figure S3). Taken together, patterns of spore viability and WT-like levels of COs across the chromosome segments examined suggest that CO assurance is functional in pch2Δ. These findings raise the possibility that CO assurance and CO interference can be separated (discussion).
Non-Mendelian marker segregation during meiosis (e.g., 3∶1 or 1∶3) occurs due to gene conversion of markers in association with meiotic recombination [31]. Gene conversion frequencies are increased in pch2Δ at eight markers, 1.2- to 2.0 fold over WT (Figure 5D; Table S2). Thus, Pch2 plays a role in suppressing gene conversion. A gene conversion may be flanked by parental or recombined chromosome arms, suggesting association with a NCO or a CO, respectively. In WT, at the assayable six central markers, gene conversions are associated with COs and NCOs at similar frequencies. In pch2Δ, at markers where gene conversion is substantially increased (ade2, met13, cyh2), such events are also flanked by COs and NCOs with similar frequencies (Figure S4). Thus, pch2Δ increases occurrence of gene conversions in both CO and NCO interactions. Increased gene conversion could be due to changes in the length of heteroduplex in recombination intermediates, repair defects, and/or region-specific changes in DSB levels (see discussion). Notably, pch2Δ does not affect DSB levels at a hotspot of recombination and does not change global DSB patterns along the majority chromosomal loci (A. Hoachwagen, personal communication; [35],[37]).
Meiotic phenotypes in several mutants are dramatically modulated by incubation temperature, with prominent effects on processing of recombination intermediates and formation of CO products (e.g., [11],[37]). In pch2Δ, temperature modulates defects in recombination progression, but not those in chromosome domain organization [37]. To examine whether the interference defect observed in pch2Δ at 33°C is affected by temperature, we investigated crossover formation and interference also at 30°C. Surprisingly, this minor temperature change results in a drastic improvement in CO interference in pch2Δ. At 30°C, map distances along three chromosomes are similar between pch2Δ and WT (i) for total tetrads, without increases in NPDs (Figure S5) and (ii) for subset P tetrads (Figure 8A, left panel). Also, and in sharp contrast to observations at 33°C, map distances in subset N tetrads exhibit only minor differences between pch2Δ versus WT. Only interval pairs 8N9 and 9N8 exhibit significantly higher CO frequencies in pch2Δ (Figure 8A, right panel). Modified coincidence analysis suggests that in pch2Δ at 30°C, interference is lost in only two interval pairs (Figure 8A; interval pairs 2-1 and 5-6). NPD frequencies further indicate loss of interference in pch2Δ at 30°C in only one interval (Table S4). Thus, defects in crossover interference can be suppressed by incubation at lower temperatures.
The role of Pch2 in meiosis when DSBs are limiting was examined in hypomorphic spo11 strain backgrounds. In WT meiosis, normal homolog segregation is maintained despite reduction of initiating DSBs to ∼20%, of normal levels, likely due to preferential processing of DSBs into COs versus NCOs [29]. Levels of initiating DSBs are reduced to ∼80%, ∼30% or ∼20% of normal levels in strains homozygous for spo11-HA, heterozygous for alleles spo11yf-HA ( = spo11yf) and spo11-HA or homozygous for spo11da-HA ( = spo11da), respectively [29].
Patterns of tetrad viability in PCH2 and pch2Δ strains were determined following sporulation at 30°C on solid medium (n≥97 tetrads). Frequencies of four spore-viable tetrads in WT indicate normal chromosome segregation in >58% of cells despite DSB reduction to ∼20% of WT levels consistent with earlier findings (Figure 7; Table S3) [29]. Frequency of four spore-viable tetrads decreases dramatically in pch2Δ strains hypomorphic for spo11, in particular when DSBs occur are reduced below levels occurring in a homozygous spo11-HA strain. Chromosome segregation occurs normally in only 18% and 8% of meioses in spo11yf/spo11-HA and homozygous spo11da/” strains, respectively, and >50% of meioses in the same strains generate zero spore-viable tetrads. Such viability patterns can occur when≥two homolog pairs missegregate. We conclude that Pch2 plays a critical role for spore viability when DSBs are reduced. Thus, although Pch2 does not play a role in spore viability at normal DSB levels, it is essential under conditions of reduced DSB formation.
Following observation of temperature-modulated interference in pch2Δ, we next examined whether incubation conditions also affect spore viability in pch2Δ at reduced DSB levels. Examining effects of hypomorphic spo11 on spore viability at 33°C on solid medium, we find, surprisingly, that spore viabilities are high in the pch2Δ strain, a drastic deviation from observations at 30°C (compare Figure 7 and Figure 8B). Notably, in spo11yf/spo11-HApch2Δ at 33°C, 75% of tetrads undergo normal meiotic chromosome segregation as suggested by levels of 4 spore-viable tetrads, compared to 18% 4 viable spore tetrads in the same strain sporulated at 30°C in parallel (see Figure 7, orange bars). Similar results are obtained for pch2Δ strains homozygous for spo11da/” (compare Figure 7 and Figure 8, yellow bars) or for spo11da/spo11yf (data not shown): At 33°C, these strains give rise to 37% and 81% four spore-viable tetrads, compared to frequencies of 8% and 2%, respectively, at 30°C. (In spo11daPCH2/spo11yfPCH2 ∼48% of tetrads give rise to four viable spores at both 33°C and 30°C). In subsequent experiments, we also discovered that spore viability patterns in pch2Δ strains hypomorphic for spo11 are also affected by culture conditions (see below).
In summary, higher versus lower temperatures oppositely modulate pch2Δ defects in CO interference and spore viability at reduced DSB levels. Conditions that improve spore viability weaken or eliminate interference and vice versa. Together, these results have several implications: (i) CO interference and factors affecting spore viability at reduced DSB levels can be uncoupled in pch2Δ. (ii) Effects of temperature on CO interference and the process that mediates normal spore viability at reduced DSB levels suggest linkage via Pch2 between both processes. (iii) Pch2 stabilizes both CO interference and spore viability over a wide range of DSB levels, temperatures, and possibly other environmental conditions (see below).
Additional effects of incubation conditions in pch2Δ were revealed during our investigation of recombination defects at reduced DSB levels. Physical recombination analysis is routinely performed in liquid medium, while spore viability is determined following sporulation on solid medium. To ascertain correspondence between these conditions, asci from parallel cultures incubated at 30°C with solid or liquid medium were dissected and viability patterns were compared. Surprisingly, WT and pch2Δ strains carrying spo11-da/” formed four viable spore tetrads at much higher levels when sporulated at 30°C in liquid versus solid medium (Figure 9A, compare pink and yellow bars).
To examine CO formation in pch2Δspo11da/” under conditions that result in low levels of four viable spore tetrads, sporulation in liquid medium was analyzed at 27°C and 30°C: In pch2Δspo11da/”, incubation of parallel cultures results in a dramatic decrease in spore viability at 27°C versus 30°C, while viability patterns are similar under both conditions in PCH2 (Figure 9A, pink and blue bars). Accordingly, in pch2Δ hypomorphic for spo11, spore viability is modulated not only by temperature, but also by the exact nature of the sporulation medium.
Final CO levels were examined in PCH2 and pch2Δ in a spo11da/” background at the HIS4LEU2 hotspot of recombination (see ref. [11] for details). Surprisingly, CO levels in PCH2 and pch2Δ were extremely similar; both at 27°C and 30°C, in four independent WT and pch2Δ strains (Figure 9B and data not shown). Thus, differences in CO formation, at least at the HIS4LEU2 recombination hotspot, are not responsible for the loss in spore viability in pch2Δ at reduced DSB levels.
In summary, pch2Δ is defective in ensuring normal spore viability when overall DSB levels are reduced, with viability patterns suggestive of homolog disjunction defects. pch2Δ may affect genome-wide levels or distribution of initiating DSBs, the efficient designation of DSBs as future COs in genomic regions outside of the HIS4LEU2 hotspot or the formation of functional chiasmata (see Discussion). Importantly, our results identify Pch2 as a protein that ensures normal homolog segregation at reduced DSB levels. Reduction of DSBs or absence of Pch2 alone have marginal or no effects on homolog segregation, yet both mutant conditions combined synergistically affect spore viability.
The present work provides novel insights into the question of how chiasma distribution is controlled along the genome of sexually reproducing organisms. We demonstrate a spatial association between CO-designated sites and structurally differentiated chromosome axes: Early during meiotic prophase, axis proteins Hop1 and Red1 preferentially associate with sites designated to become COs. A key implication of this finding is that local modifications of chromosome axis and recombination site selection are coordinated, and are possibly controlled by the same determinants. Pch2 controls the association of Hop1 with designated CO sites, the number of designated CO sites along the genome and CO interference. Identification of pch2Δ as a mutant that affects CO/chiasma formation in response to remote recombination events without playing a major role in overall CO levels suggests that CO control is mechanistically distinct from and likely superimposed on basic recombination events. Based on these data, we propose a model of CO control in which meiotic chromosomes become organized into multiple modules of assured CO formation, with concurrent imposition of crossover interference. Pch2 is proposed to function as a size determinant for these modules of chiasma assurance and interference.
Pachytene chromosomes in yeast display a domainal organization, where Hop1/Red1 or Zip1-enriched regions occur in an alternating pattern [37]. Our analysis supports the idea that hyperabundance domains are layered over a base level of Hop1 (and Zip1) along the lengths of synaptonemal complexes. Correspondence between Hop1 hyperabundance domains and CO-designation marker Zip3 establishes a link between domainally modified chromosome axes and positions of future COs [21]–[23]. Crossovers and CO designation markers Zip2/Zip3 occur at different sites in each cell and exhibit interference distribution [22],[46]. By implication, Hop1 hyperabundance domains likely also form at different positions in different cells along a given chromosome. Together, these findings suggest that chromosome axes undergo a differentiation process that is spatially coordinated with crossover placement.
During WT meiosis, association between Zip3 and Hop1 reaches maximum levels in pachytene nuclei. High levels of Zip3-Hop1-association are also detected in ndt80Δ cells arrested at the pachytene stage (Figure 4L), indicating that association is established prior to and independent of double Holliday junction resolution, a recombination step blocked in ndt80Δ [10]. Thus, Hop1-Zip3 association is completed prior to and independent of completion of the majority of COs.
Our findings further suggest that association between Hop1/Red1 and Zip3 is established during the zygotene stage and can occur independently of stable strand invasion. Zip3 is the earliest known marker for designated CO sites [22]. In pre-zygotene cells, Zip3 localizes to paired centromeres of yeast chromosomes [24]. During the zygotene stage, Zip3 appears to localize abundantly to additional sites, a process that is completed at the pachytene stage when Zip3 is found at interference-distributed CO designation sites, while it is absent from centromeres [24]. In the population of zygotene nuclei analyzed here, ∼50 Zip3 foci are detected, suggesting that Zip3 occupies multiple non-centromeric positions at this stage. Zip3 foci colocalize at high levels with Red1 in the same zygotene cells (Figure 3I–3M). Thus, association between Red1/Hop1 and Zip3 at designated CO sites appears to be established during the zygotene stage.
In dmc1Δ, a mutant defective for strand invasion, Zip3 localizes to chromosomes with numbers substantially lower than those observed in normal zygotene nuclei. A subset of these cells, however, exhibit Zip3 localization with WT-like numbers and patterns (Figure 4C). A high percentage of Zip3 foci is associated with Hop1 in such nuclei, raising the possibility that Zip3-Hop1 association can occur independent of Dmc1-mediated strand invasion. Whether Zip3 localizes to its normal sites in dmc1Δ cells is presently unknown.
Hop1/Red1 localize to meiotic chromosomes prior to and independent of DSB formation, consistent with functions at earlier stages (G.V.B., unpublished; [13],[17]). Several scenarios can explain the transition between early, pre-DSB association of Red1/Hop1 with meiotic chromosomes and their association with Zip3 marked designated CO sites at later stages: Red1/Hop1 may (i) become associated with future CO sites as an outcome of CO designation, following relocalization from a more dispersed (pre-)leptotene localization pattern; (ii) initially be present at all nascent recombination interactions and later undergo selective stabilization at future CO sites; (iii) preferentially localizes to future CO sites prior to CO designation, possibly participating in CO designation itself. (iv) Finally, it is possible that Zip3 preferentially localizes to Hop1/Red1 hyperabundance domains due to preferential recombination initiation in such domains [17]. Further work is required to determine whether Hop1/Red1 hyperabundance domains assemble at CO designated chromosomal positions before or after CO designation, and whether CO designation is a requirement for association between Hop1 and Zip3.
Spatial association between structural axis modifications and markers of nascent recombination interactions have also been observed in other organisms. (i) In Sordaria, cohesin associated protein Spo76/Pds5 is depleted from Msh4-marked recombination sites in a mutant deficient for the meiotic cohesin Rec8. Local splitting of sister chromatids at the corresponding sites also occurs in the WT, and it was proposed that intersister connections may become destabilized as part of the normal process of chiasma formation [16]. (ii) In an ATM−/− mouse, SC proteins Sycp3 and Sycp1 are absent from sites of ongoing recombination [47]. (iii) In C. elegans, SC component SYP-1 and axis proteins HTP-1/2 are removed from reciprocal chromosome arms directed by and dependent on designation of a given recombination interaction as future CO [48]. These data indicate that locally weakened sister cohesion and enhanced interhomolog interactions may both contribute to preferential interhomolog recombination. Such modifications may be especially important for CO formation which entails long-lived strand invasion intermediates [10],[11]. In yeast, axis ensemble Hop1/Red1 plays a central role in directing meiotically induced DSB processing towards homologous chromosomes, with Mec1-dependent Hop1 phosphorylation constituting a key event in establishing this interhomolog bias [19].
Our results further provide insights into likely dynamics of SC assembly. SC initiation occurs at Zip3 foci, some of which correspond to CO-designated sites [21] (N.J., unpublished data). Conversely, domains enriched for Zip1 alternate with Hop1/Red1 (and Zip3) enriched domains when SC assembly is complete [37]. We propose that in early zygotene nuclei, Hop1/Red1 and Zip1 are present at Zip3-marked sites (see Figure 3I–3L). During SC polymerization, Zip1 is preferentially deposited in axis regions distal from Zip3, giving rise to the alternating Zip1/Hop1 pattern of pachytene chromosomes. Our results are not compatible with a model of SC assembly where Hop1 is displaced from chromosome axes as Zip1 polymerizes [13]. Hop1 in association with Zip3, is present at substantial levels along pachytene chromosomes, both in WT and in ndt80Δ, indicating that Hop1 is an integral component of pachytene chromosomes (this work; [37]).
Pch2 plays key roles in establishing and/or maintaining the distribution of at least two proteins along meiotic chromosome axes. First, Pch2 controls overall levels and localization patterns of Hop1. Accordingly, Hop1-enriched domains appear as multiple discrete foci along WT pachytene chromosomes, while in pch2Δ, Hop1 spreads into fewer, more extended structures (Figure S2B, S2C). Importantly, changes in Hop1 loading in pch2Δ are not an indirect consequence of e.g. a delay in meiotic progression: In ndt80Δ arrested cells, Hop1 forms distinct foci, while in the ndt80Δpch2Δ double mutant, Hop1 loads at increased levels and localizes uniformly along chromosomes. Pch2 may control Hop1 association with chromosome axes via its chromosomal localization. Consistent with this idea, in zip1Δ, a mutant condition that eliminates Pch2 specifically from chromosome arms, Hop1 and Red1 load in a continuous pattern along chromosome axes, reminiscent of patterns observed in pch2Δ [13],[21],[33].
Increased numbers of Zip3 foci along pch2Δ pachytene chromosomes further identify a function of Pch2 in controlling association of Zip3 with meiotic chromosome axes. Increased numbers of Zip3 foci in pch2Δ may indicate defects in CO designation, possibly indicating an increase in the number of CO designation sites with associated inefficiencies to form functional chiasmata. Notably, Zip3 represents a CO designation marker in WT. In several mutants exhibiting reduced CO levels and loss of CO interference, Zip2 foci form with apparently normal numbers and distribution, indicating that CO designation on the cytological level can be uncoupled from the execution of CO formation [22].
Chromosome axis defects in pch2Δ are indicated by increased numbers of Zip3 foci and a failure to undergo appropriate axis shortening. Axis shortening, too, may be an outcome of normal CO designation. Coordinate increases in axis length and number of CO designation sites as well as loss of interference in pch2Δ support a linkage between axis length and CO control. SC length and CO numbers are closely correlated in many taxa, including mammals [41],[49], consistent with the idea that CO number and distribution are controlled via the status of the chromosome axis. In a mutant situation such as pch2Δ, changes in axis status may indicate defects in chromosome axis status, with possible effects on CO placement and/or formation of functional chiasmata.
Pch2 suppresses COs in adjacent chromosome regions without being required for normal CO formation. In pch2Δ, map distances tend to be increased when flanking intervals are nonparental, yet are at WT levels when neighboring intervals are parental. The major implication of these results is that functions in CO interference can be separated from those in CO formation. While Pch2 is required for timely CO formation, COs form at normal levels in a pch2Δ mutant, both at a hotspot of recombination and in the genetic intervals examined here [35],[37, this work]. Conversely, we consider it as unlikely that Pch2 changes the overall distribution of COs rather than affecting CO interference: An overall change in CO distribution should change individual map distances independent of the presence or absence of a CO in an adjacent interval.
Pch2 further does not exert a general inhibitory effect on recombination: (i) DSBs form at normal levels in pch2Δ when analyzed at a recombination hotspot or by a genome-wide approach (A. Hochwagen, personal communication; [35],[37]). (ii) Absence of Pch2 does also not compensate for low DSB levels in hypomorphic spo11 mutants, e.g. by improving spore segregation. Thus, Pch2 performs a function in CO control without playing a role in overall CO levels.
While CO levels in pch2Δ are normal in the intervals examined here, the number of Zip3 foci is substantially increased. We interpret this discrepancy as indicating that CO designation is increased in pch2Δ, yet does not result in a corresponding increase in completed COs. At the same time, we cannot exclude that somewhat different incubation conditions result in actual increases in CO levels in pch2Δ. Cytological studies presented here were performed in liquid medium, but CO levels along the three chromosomes were determined following sporulation on solid medium. Such minor differences may have major effects on CO levels and distribution in pch2Δ. Notably, an independent study from the Alani lab observed increased CO levels in pch2Δ (see accompanying paper).
Two classes of COs, one that exhibits interference and the other that does not exhibit interference, have been proposed to contribute to total CO levels in the WT [e.g., 11],[32],[50],[51]. Accordingly, in certain mutants, CO reduction is accompanied by defective interference among residual COs [32]. Identification of pch2Δ as a mutant that forms COs at normal levels but is defective for interference suggests that interference is superimposed on basic recombination pathways, and that it can be eliminated without loss of COs.
Absence of Pch2 further results in increased levels of gene conversion. Association of increased gene conversion levels with parental and non-parental configuration of flanking chromosome arms suggests that Pch2 affects this process prior to bifurcation of the CO and NCO pathways. Increased gene conversion without increases in DSB levels could occur due to an increased length of heteroduplexes in ongoing recombination interactions and/or mismatch repair defects in recombination intermediates. Such defects could be an outcome of spatial changes in axis juxtaposition, or due to elimination of other factors.
The biological function of interference is presently mysterious. Several ideas have been put forward to explain this conserved phenomenon [e.g., 52]: (i) Closely spaced double COs provides insufficient sister cohesion resulting in chromosome missegregation [53]. (ii) Interference is a byproduct of the CO assurance system [54]. Linkage between defects in interference, loss of CO assurance and/or normal homolog segregation and reduced spore viability in most yeast mutants with interference defects has complicated our understanding of the function of interference [e.g., 28],[32],[55],[56]. The current study indicates that short range interference is not required for normal chromosome segregation and/or the formation of functional gametes. Notably, all intervals tested along chromosome III exhibit interference defects here, yet chromosome segregation (including chromosome III) is normal in pch2Δ, as suggested by high spore viability patterns. Based on this finding, we present a model postulating that interference is a byproduct of the CO assurance system (see below; [54]).
Full levels of interference appear dispensable for meiotic chromosome segregation, yet we demonstrate that a mechanism compensating for reduced DSBs is critical for viable gamete formation. Segregation of the 16 homolog pairs is mostly normal during WT meiosis, even when DSBs are reduced to <20% of WT levels (in spo11da; this work; [4],[29]). By contrast, in a pch2Δ background, DSB reduction to <80% of WT levels (in spo11yf/spo11-HA), results in catastrophic reduction of spore viability, identifying essential functions for mechanisms that compensate for reduced DSBs during yeast meiosis.
In a pch2Δ mutant hypomorphic for spo11, two- and zero viable spore tetrads are highly abundant, a pattern suggestive of defects in homolog disjunction. Such defects are frequently attributed to a failure of homolog pairs to acquire sufficient COs for homolog disjunction. Our analysis of CO levels at the HIS4LEU2 recombination hotspot does not provide evidence for substantial CO defects in pch2Δ at reduced DSB levels. CO levels at HIS4LEU2 may not be representative for genome-wide CO levels in pch2Δspo11da. Notably, unlike other genome regions, HIS4LEU2 does not exhibit CO homeostasis [29]. Alternatively, COs may form efficiently along the entire genome in pch2Δspo11da, but fail to undergo appropriate chiasma maturation. Such defects could affect intersister connections near crossovers resulting in a failure to maintain cohesion along chromosome arms until onset of anaphase. Severe defects in spore viability despite substantial CO formation have also been demonstrated for pch2Δrad17Δ double mutants, and may be related to the results reported here [35].
Importantly, results presented here define a Pch2-dependent mechanism that assures homolog segregation at reduced DSB levels. In yeast, on average>five COs/chiasmata form per homolog pair (90 COs distributed among 16 homolog pairs). Accordingly, stabilizing functions in homolog segregation and/or CO assurance may only manifest themselves when initiating DSBs are reduced. In organisms with lower wild-type COs levels, similar defects in chiasma function may result in homolog nondisjunction at normal DSB levels due to a failure to acquire sufficient COs (e.g. the XY pair in mammals [47]).
Unexpectedly, pch2Δ defects in CO interference and in spore viability at reduced DSB levels are partially rescued under certain conditions. For example, at 33°C, despite loss of most short range CO interference, some long-range interference (>100 kb) is retained. Furthermore, at 30°C, pch2Δ is mostly proficient for interference, yet reduced DSB levels result in formation of inviable spores. Pch2 appears to stabilize CO control and spore viability over a wide range of conditions, including different temperatures and low DSB levels. In the absence of Pch2, a temperature decrease of only 3°C results in catastrophic chromosome missegregation, with no comparable effects in WT highlighting the necessity of Pch2-mediated stabilization of CO control. Oppositely stabilizing and destabilizing effects of temperature on interference and viable spore formation are consistent with the idea that interference and homolog disjunction assuring chiasma formation are the outcome of two antagonistically-acting pathways. Thus, these functions are separable based on their different dependence on incubation temperature. One explanation is that temperature oppositely modulates two chromosome components, e.g. chromosome axes and chromatin fiber (see below).
We infer the existence in the absence of Pch2 of one or several default systems that provide partially functional CO control and/or mechanisms for maintaining high levels of spore viability. Backup systems for CO control may e.g. utilize basic organizational features shared with mitotic chromosomes. Roles in CO control of general structural chromosome components have been demonstrated in C. elegans [57].
We note that Pch2-independent backup systems appear to function independent of a properly structured chromosome axis: Incubation conditions modulate pch2Δ defects in crossover placement and spore viability/chromosome segregation, but not chromosome axis defects. More uniform Hop1 association in pch2Δ occurs over a wide range of conditions, at 33°C and 23°C, and is also evident at 30°C in a different strain background [33]. Drastically different defects in interference and CO homeostasis are observed under the respective conditions (this work; [37]). Together, these data indicate that uniform Hop1 association with chromosome axes is a consequence, not a cause of the initial CO control defect.
The current work identifies functions of Pch2 during WT meiosis in chromosome morphogenesis, CO placement and spore viability/homolog segregation when DSBs are reduced. In C. elegans and Drosophila, Pch2 prevents meiotic progression in mutant meiosis when chromosomal events independent of recombination initiation are defective. No role of Pch2 in WT meiosis has been detected in these organisms [34],[38]. In mouse WT meiosis, Pch2 is required for efficient completion of recombination, but no role in mutant meiosis as a checkpoint is apparent [36]. Accordingly, Pch2 has been described as a checkpoint or a factor required for normal meiotic progression.
Pch2 modulates axis status, recombination progression and SC morphogenesis [37, this work]. Thus, Pch2 affects all processes in WT meiosis that it is proposed to monitor as a checkpoint during mutant conditions. We propose that control of chromosome axis status constitutes Pch2's primary function, with secondary effects on recombination progression, CO placement and homolog segregation. Accordingly, changes in axis status may result in destabilized homolog juxtaposition, with downstream effects such as delayed double Holliday junction turnover, delayed CO/NCO formation and aberrantly high levels of non-Mendelian segregation events [37, this work]. pch2Δ induced changes in axis status would likely also affect axis-associated processes under mutant conditions, with possible consequences for checkpoint activation. Modulation of underlying defects rather than compromised monitoring represents an attractive explanation for diverse Pch2 functions in mutant and WT meiosis. Alternatively, Pch2 may perform unrelated functions in meiotic cell cycle control, CO placement and gamete viability.
We note that in yeast, pch2Δ defects in WT meiosis are relatively subtle, and detectable only under certain conditions (see above). Corresponding defects in other organisms may also be difficult to detect and/or become manifest only under certain conditions. Consistent with this idea, in Drosophila, a synergistic effect of pch2Δ on CO levels has recently been demonstrated in combination with another mutant [38].
One key outcome of the current study is that Pch2 appears to suppress or enhance formation of functional chiasmata at normal or reduced DSB levels, respectively. Here, we propose a model integrating these apparently opposing functions of Pch2. Pch2 is proposed to reorganize chromosome axes into long range CO control modules, hereafter referred to as ‘One Crossover Modules’ (OCMs). Key features of OCMs include assurance to undergo one CO, and suppression of additional COs within the same module. Modules are proposed to tile each bivalent, resulting in formation of as many COs as OCMs.
Pch2-mediated CO control is proposed to occur in two-steps: (i) bivalents become organized into a tiling array of OCMs. (ii) CO designation and interference occur. Cytologically, each OCM would correspond to a centrally localized Zip3 focus with associated Hop1 hyperabundance domain extending to both sides into Hop1-poor regions, reflecting the reach of interference. Pch2 is proposed to function as a determinant for OCM installation.
The stress hypothesis of CO control provides a mechanistic explanation of how chiasma assurance/maturation and interference might be linked along each OCM [27],[54]. We propose that each OCM constitutes an independent stress module. Stress and stress relief along the axis are hypothesized to mediate crossover designation and interference, respectively. Specifically, compression stress along the axis would result in localized axis deformation with two important consequences, stress relief and CO designation. Stress relief prevents additional axis deformation events along each OCM, effectively establishing interference.
By setting a module for stress transmission, Pch2 would promote CO progression of a DSB proximal to the deformed axis segment, and coordinately prevent additional DSBs from undergoing the same fate. OCMs may become installed de novo, or, more likely, be specified via modification of preexisting chromosome features. Available data are easily integrated with this model: Pch2 associates with chromosome axes during the early zygotene stage. At the same stage, CO/NCO differentiation is finalized, axis domains associated with future COs appear, and the interference distribution of Zip2/Zip3 becomes established [9],[11],[22]. Programmed axis deformation at CO sites associated with stress relief and CO designation could further contribute to axis shortening. When defective, this may result in aberrantly long axes (this work). Moreover, axis deformation may promote assembly of Red1/Hop1 and Zip3 at CO-designated sites, with aberrant CO designation/axis deformation resulting in uniform Hop1 axis association.
Default modules that provide some CO control in pch2Δ may result in suboptimal CO designation, aberrant CO positioning, or increased sensitivity to incubation conditions. Such effects may be particularly detrimental when DSBs are limiting. Under such conditions, DSBs normally ensured to become COs may now fail to induce steps in chromosome morphogenesis associated with normal chiasma formation.
Control of CO numbers via one-crossover modules provides an attractive way how recombination frequencies can be controlled in different organisms and even different sexes. Accordingly, in C. elegans, each chromosome would be organized as a single module, while in e.g. mouse, there would be one or two OCMs per homolog pair. In many species, CO levels between males and females differ for identical homolog pairs. Setting differently sized OCMs represent an attractive way to control CO levels in a chromosome-wide manner. Levels and distribution of COs are dramatically modulated by temperature and other environmental factors in many eukaryotes (e.g., [58]). Such sensitivity along WT chromosomes may be related to the mutant sensitivities revealed by the current work.
In summary, we have demonstrated here that chromosome axes undergo programmed changes in their global structure that strikingly parallel the non-random positioning of chiasmata during meiosis. Unlike other cases of cytologically detectable chromosome domain organization, including heterochromatin assembly, such domainal organization is determined individually for each cell, in accordance with non-random meiotic crossover distribution. Close functional and temporal coordination between assured crossover formation and chromosome domain organization identify potential functions for chromosome axis status in faithful meiotic homolog segregation.
Strains were of the SK1 background (Table S5). Markers were introduced by transformation or crossing and were verified by Southern blot. N-terminally HA-tagged PCH2 was transferred from the BR strain background [33] by insertion of the URA3 marker 300 bp upstream of PCH2 followed by PCR amplification of the tagged construct including the marker and transformation into SK1 (strain gift from A. Hochwagen). In the resulting strain, the 3×HA-tag encoding sequence is flanked by 15 and 14 polylinker-encoded amino acids, respectively (N.J., unpublished data).
Haploids mated overnight on supplemented YPD were transferred to identical batches of sporulation medium (0.5% potassium acetate, 0.02% raffinose) and incubated at 33°C or 30°C for 72 hrs. Asci were incubated with zymolyase, dissected on supplemented YPD and replica-printed to appropriate media to determine marker status. Tetrads exhibiting non-Mendelian segregation of ≥5 markers were assumed to be false tetrads and omitted from further analysis. For calculations of map distances and NPD frequencies, see text. Standard error calculations were performed using the Stahl Lab Online Tools. Tetrads with non-Mendelian segregation for either marker of an interval were omitted for calculations for that interval. Chi square values were used to calculate P-values using the Vassar College webpage.
Time courses and meiotic spreads were prepared and immunostained as described [11]. Chromatin was stained using DAPI. Hop1 and Zip1 were stained with rabbit anti-Hop1 (F. Klein) and rabbit anti-Zip1 (S. Keeney) antibodies at 1∶300 to 1∶1000, except for nuclei shown in Figure 4Q–4V in which mouse anti-Zip1 antibody (P. Moens) was used at 1∶500 dilution. GFP- and HA fusion proteins were detected with goat anti-GFP (Rockland) at 1∶400 or mouse anti-HA (Covance) antibodies at 1∶1000 dilution, followed by incubation secondary antibodies conjugated to Alexa 488-, Alexa 594-, or Alexa 680 (Molecular Probes) at 1∶2500 dilution. All antibodies were tested for epitope specificity using appropriate deletion/untagged strains. Images were captured by a computer-assisted fluorescence microscope system (DeltaVision, Applied Precision). The objective lens was an oil-immersion lens (100×, NA = 1.35). Image deconvolution was carried out using an image workstation (SoftWorks; Applied Precision). In double staining experiments, real colocalization was scored by counting the number of overlapping foci in composite images. Fortuitous colocalization was evaluated by the misorientation method where one of the two images is rotated by 180°, ensuring maximum nucleus overlap, and colocalization is determined by counting the number of overlapping foci [59].
ImageJ was used for processing and quantitative analysis of images saved as 16bit TIFF files in SoftWorks. To analyze Hop1 distribution patterns, threshold levels were visually adjusted to maximize signal detection within the DAPI staining area, followed by measurements of the Hop1 positive area and the mean signal intensities at above background levels. Subsequently, threshold levels were set to the mean signal intensity for each image, and a mask was generated for the Hop1 positive signals exhibiting≥average signal intensities. Masks were transferred into MicroMeasure to determine the number and maximum lengths of individual Hop1 signals (below). Total SC length in WT and pch2Δ strains were measured by a “blind” observer, using MicroMeasure. MStat 5.1 was used for data plotting and statistical analysis.
Rasband, WS, ImageJ, U. S. National Institutes of Health, Bethesda, Maryland, USA, http://rsb.info.nih.gov/ij/, 1997–2008
Drinkwater N Mstat Statistical Software: http://mcardle.oncology.wisc.edu/mstat/
Stahl Lab Online Tools: http://www.molbio.uoregon.edu/~fstahl/
MicroMeasure, version 3.3: (http://www.colostate.edu/Depts/Biology/MicroMeasure
|
10.1371/journal.pntd.0000925 | Ecology and Geography of Plague Transmission Areas in Northeastern Brazil | Plague in Brazil is poorly known and now rarely seen, so studies of its ecology are difficult. We used ecological niche models of historical (1966-present) records of human plague cases across northeastern Brazil to assess hypotheses regarding environmental correlates of plague occurrences across the region. Results indicate that the apparently focal distribution of plague in northeastern Brazil is indeed discontinuous, and that the causes of the discontinuity are not necessarily only related to elevation—rather, a diversity of environmental dimensions correlate to presence of plague foci in the region. Perhaps most interesting is that suitable areas for plague show marked seasonal variation in photosynthetic mass, with peaks in April and May, suggesting links to particular land cover types. Next steps in this line of research will require more detailed and specific examination of reservoir ecology and natural history.
| We analyzed the spatial and environmental distributions of human plague cases across northeastern Brazil from 1966-present, where the disease is now only rarely transmitted to humans, but persists as a zoonosis of native rodent populations. We elucidated environmental correlates of plague occurrences by way of ecological niche modeling techniques utilizing advanced satellite imagery and geospatial datasets to better understand the ecology and geography of the transmission of plague. Our analysis indicates that plague foci in Brazil are indeed insular as previously suggested. Furthermore, distribution of such foci are likely not directly dependent on elevation, and rather are contigent on climate and vegetation. Seasonality of zoonotic plague transmission is linked to variations of these ecological parameters- particularly the increase in precipitation and primary production of the rainy season. Spatial analysis of transmission events afford a broad view of potential plague foci distributions across northeastern Brazil and indicate that the epidemiology of plague is driven by a dynamic array of environmental factors.
| Plague arrived in Brazil during the Third Pandemic, in October 1899, imported by ship traffic to Santos, in São Paulo state, and was rapidly diffused to other coastal cities. By 1906, it had dispersed by means of land and sea commerce more broadly, and had become established in native rodent populations, particularly in the northeastern sector of the country [1], [2]. Nonetheless, records of plague in Brazil are sparse through the 1920s, making detailed tracking of the pattern of spread of the disease in the region difficult or impossible [1], [3].
Only by around 1936 were data on plague and its control in Brazil regularly collated and archived. Based on analyses of these data (for 1936–1966), Baltazard [1] identified numerous distinct plague foci occurring in different environmental contexts, a viewpoint that was updated by Vieira & Coelho [4], based principally on elevation. These foci appear to exist independently of one another in time and space [1], and overall numbers of human cases varied from 20 to 100 until the 1970s. Since that time, all of the foci entered a period of relative inactivity, with few or no human cases [5], [6], [7], [8]. The last significant outbreak in Brazil was in the late 1980s in Paraíba [9].
The purpose of this contribution is to present a first range-wide analysis of the geography and ecology of plague transmission in northeastern Brazil using tools drawn from the emerging field of ecological niche modeling, which is beginning to see application to plague biology [10], [11]. Although no recent plague transmission to humans has been recorded in this region, plague remains as a zoonosis across much of northeastern Brazil [12], making a thorough understanding of its geographic distribution an ongoing priority. Here, we marshal new tools from quantitative biogeography in the form of ecological niche modeling approaches, which related known points of occurrence to raster geospatial GIS data layers to estimate the ecological niche of a species or other biological phenomenon, such as transmission of a disease [11]. The result is both a spatial prediction of areas of potential transmission and a first-order evaluation of environmental correlates of plague transmission in northeastern Brazil.
Methods and approaches for estimating ecological niches from species' occurrence data have seen considerable exploration in recent years [23], [24]. Outcomes of these tests have been mixed, with some serious criticisms of the algorithm used herein, the Genetic Algorithm for Rule-Set Prediction (GARP) [25]—these criticisms [23], [26], however, have been based either on misunderstandings of how to use the algorithm [27] or on artifactual differences in performance measures [28], [29]. In reality, and when properly used and evaluated, GARP offers estimates of species' ecological niches that are highly robust to small sample size and to broad gaps in spatial coverage of landscapes in terms of input data [28], [29]—for this reason, we used this approach throughout this study.
GARP is an evolutionary-computing method that estimates niches based on non-random associations between known occurrence points for species and sets of GIS coverages describing the ecological landscape. Occurrence data are used by GARP as follows: 50% of occurrence data points are set aside for an independent test of model quality (extrinsic testing data), 25% are used for developing models (training data), and 25% are used for tests of model quality internal to GARP (intrinsic testing data). Distributional data are converted to raster layers, and by random sampling from areas of known presence (training and intrinsic test data) and areas of ‘pseudoabsence’ (areas lacking known presences), two data sets are created, each of 1250 points; these data sets are used for rule generation and model testing, respectively.
The first rule is created by applying a method chosen randomly from a set of inferential tools (e.g., logistic regression, bioclimatic rules). The genetic algorithm consists of specially defined operators (e.g. crossover, mutation) that modify the initial rules, and thus the result are models that have “evolved”—after each modification, the quality of the rule is tested (to maximize both significance and predictive accuracy) and a size-limited set of best rules is retained. Because rules are tested based on independent data (intrinsic test data), performance values reflect the expected performance of the rule, an independent verification that gives a more reliable estimate of true rule performance. The final result is a set of rules that can be projected onto a map to produce a potential geographic distribution for the species under investigation.
Because each GARP run is an independent random-walk process, following recent best-practices recommendations [30], for each environmental data set (see above), we developed 100 replicate random-walk GARP models, and filtered out 90% based on consideration of error statistics, as follows. The ‘best subsets’ methodology consists of an initial filter removing models that omit (omission error = predicting absence in areas of known presence) heavily based on the extrinsic testing data, and a second filter based on an index of commission error ( = predicting presence in areas of known absence), in which models predicting very large and very small areas are removed from consideration. Specifically, in DesktopGARP, we used a “soft” omission threshold of 20%, and 50% retention based on commission considerations; the result was 10 ‘best subsets’ models (binary raster data layers) that were summed to produce a best estimate of geographic prediction. We took as a final ‘best’ prediction for each species that area predicted present by any, most, or all 10 of these best-subsets models.
Predictive models of disease occurrence may be good or bad, but model quality can be ascertained only via evaluation with independent testing data, preferably which are spatially independent of the training data to avoid problems caused by spatial autocorrelation and nonindependence of points [28]. Because only data documenting presence of plague cases were available for this study (i.e., no data were available to document that plague was absent at particular sites), we used a binomial probability approach to model validation: we compared observed model performance to that expected under a null hypothesis of random association between model predictions and test point distribution. Because such tests require binary (i.e., yes-no) predictions, our first step was to convert raw (continuous) predictions to binary predictions. We considered three distinct thresholds: areas predicted as suitable by any (i.e., ≥1) of the 10 replicate best-subsets models (ANY), areas predicted as suitable by most (i.e., >5) of the 10 replicate best-subsets models (MOST), and areas predicted as suitable by all of the 10 replicate best-subsets models (ALL).
In the binomial tests, the number of test points was used as the number of trials, the number of correctly predicted test points as the number of successes, and the proportion of the study area predicted present as the probability of a success if predictions and points were associated at random [31]. All testing was carried out in a series of spatially stratified tests that are detailed below. These tests evaluated the ability of models to anticipate plague case distributions across unsampled areas, considering a model as validated if it predicts case distributions better than a “model” making random predictions. As such, these tests are considerably more stringent than simple random partitions of occurrence data or cross-validation exercises.
In view of the odd, focal distribution of northeastern Brazilian plague cases (Figure 1), we carried out a series of tests of predictive abilities of models among the five foci that are easily discernable. In each case, we examined model predictivity in a k – 1 framework: with k = 5 foci, we tested all combinations of 4 foci by means of their ability to predict spatial distributions of plague cases in the fifth focus. Tests were developed within two spatial contexts—within 50 km and within 200 km—surrounding the known occurrences within the target focus.
Finally, we wished to develop a single overall model that represents the best-available picture of plague case-occurrence risk across northeastern Brazil, albeit not including statistical testing as above. This model was built using all occurrence data available. To assess uncertainty in these predictions based on all case-occurrence information, we built 100 models each based on a random 50% of the occurrence data chosen at random without replacement. These models thus capture the degree to which plague case-occurrence data availability may drive the results of the analyses, and we consider areas that are predicted consistently in all of these replicate analyses as most certain. We projected this model onto environments across eastern Brazil to provide a broader-extent visualization of the ‘niche’ of plague in northeastern Brazil.
To explore environmental factors associated with positive and negative predictions of suitability for DF transmission, we explored further the environmental correlates of the model based on all points. We plotted 1000 points randomly across areas of the municipalities predicted as absent or present by this model. We then assigned the value of each input environmental and topographic layer to each of the random points, and exported the associated attributes table in DBF format, which was then used for comparisons of environmental characteristics of areas predicted as suitable and unsuitable.
The focal and discontinuous nature of plague case distributions in northeastern Brazil is at once visible in the raw distribution of the occurrence points derived at the outset of this study (Figure 1). The discontinuities that have been assumed based on the clusters of known occurrences are supported by our ecological niche models, many of which show relatively small areas of highly suitable conditions separated by less-suitable areas (see, e.g., Figure 2). What is more, this result is manifested with or without elevation included in the analysis, and thus is not a simple consequence of topographic differences; it is also manifested in analyses based on both surface reflectance (NDVI) and climatic variables. As such, we interpret the discontinuity of plague distributions in northeastern Brazil as dependent on a multidimensional suite of environmental variables.
The model predictions in general performed quite well in anticipating plague case distributions in areas not included in model training. That is, plague not only occurs in discontinuous foci, but it also occurs under predictable and circumscribed environmental conditions, which is the basis for the success of the niche model predictions. The broadest panorama of results shows significant results dominating in the southwestern and northwestern foci (Table 1). However, the frequency of significant results in these tests is clearly and linearly related to sample size on a log10 scale (P<0.05), suggesting that predictivity would be excellent throughout the region were sample size distributions to be more adequate.
Finally, visualizing plague distributions in environmental dimensions (Figure 3), we see clear differences in the seasonal pattern of variation in greenness between areas predicted as suitable (i.e., suitability value of 10) and those predicted as unsuitable (value 0). That is, no marked seasonal variation is notable in areas predicted as unsuitable, whereas areas predicted as suitable show a marked elevation in greenness in April and May, and lower values thereafter, probably corresponding to patterns of rainfall (i.e., rainy season beginning in March, and ending by August).
Extending the model predictions across broader areas—namely all of northeastern and eastern Brazil—yields a picture of potential plague distribution across the region (Figure 4). Because plague transmission to humans in Brazil is currently nil, and no broad-extent data are available regarding circulation among mammals, we have few means of testing the reality of these model projections. However, at least in the case of models based on climatic dimensions, the area predicted as suitable includes the Serra dos Orgãos sites from which plague has been documented [5], [32].
The models that we developed for Brazilian plague offer several intriguing insights into plague distribution, ecology, and natural history in Brazil. However, understanding the limitations of these models is critical, prior to any detailed interpretation or exploration. First and foremost among the limitations of this study are the occurrence data used as input: we relied on human case-occurrence reports accumulated by the Serviço Nacional de Referência em Peste do Centro de Pesquisas Aggeu Magalhães and published in diverse scientific publications [6], [7], [8], [9], [13]. Our use of these data thereby assumes that human case-occurrences are representative of the ecological and environmental situations under which plague is maintained in the zoonotic world, which may not be the case, given the long chain of events necessary for a zoonotic occurrence to be represented in our data set (i.e., transmission to human, correct diagnosis, international reporting). On a finer scale, we also make the not-completely-satisfactory assumption that that place of residence (at the level of the ranch or settlement) is representative of the site of infection, which may be variably true depending on the particular social network and local economy.
One point that became clear in our analyses, confirming previous opinions, is that plague has a highly discontinuous and focal distribution in northeastern Brazil. Our initial suspicions that elevation played a significant role in creating these ‘islands’ were not supported, as analyses with and without elevation included as a predictor variable reconstructed the insular nature of the distribution. The NDVI-based analyses are particularly instructive, as they have no direct, mathematical relation to elevation [as do climate interpolations; 21]—rather, the discontinuous plague distribution in northeastern Brazil appears to reflect multidimensional qualities of the landscape and environment (which of course may be related biologically to elevation), rather than any simple univariate causation.
Previous studies had attributed the cause of plague focality in Brazil to elevation [1]. Baltazard [1] emphasized that Brazilian plague foci are independent—that is, that transmission appears to occur in uncorrelated patterns in different foci. Baltazard [1] also pointed out that these foci are all in elevated areas, and that they are subject to distinct precipitation regimes. Although plague has frequently shown long periods of apparent inactivity (i.e., no human cases), its reappearance at intervals nonetheless indicates its long-term persistence.
The foci are limited geographically, although their footprint can appear to expand during major outbreaks. These expansions appear to correspond to periods of particularly favorable conditions for plague transmission in the highland area, spreading out via valleys into the surrounding lowland areas. If these favorable conditions persist, taking the form of a prolonged winter, rodent host reproduction may be elevated, and plague may be able to spread beyond the limit of the highland areas into the dry sertão per se. This line of thinking led Baltazard [1] to consider the plague foci of Serra da Ibiapaba, Serra do Baturité, Serra do Machado, Serra de Uruburetama, Serra da Pedra Branca, Serra das Matas in northern Ceará (see Figure 1) as a single focus. Vieira and Coelho [4], in contrast, argued that these foci should be treated as isolated and independent. Our analyses suggest that these foci are dependent on a broad suite of conditions, and are not simple or direct correlates of elevation.
Another factor that may play in the picture of focality is the presence of key rodent hosts for plague, including Necromys lasiurus (formerly placed in Bolomys and Zygodontomys). Necromys is the rodent that is most abundant in northeastern Brazilian plague foci, and was considered as responsible for causing epizootics, from which the infection spreads to other species [1]. Given the distribution of this species, other species of rodents must be involved in plague maintenance farther south, for example in the Serra dos Órgãos, Rio de Janeiro state, Brazil. The relative roles of the distribution of the rodent hosts and the fleas (Polygenis spp.) remain to be evaluated in detail.
|
10.1371/journal.pntd.0001257 | Improving the Cost-Effectiveness of Visual Devices for the Control of Riverine Tsetse Flies, the Major Vectors of Human African Trypanosomiasis | Control of the Riverine (Palpalis) group of tsetse flies is normally achieved with stationary artificial devices such as traps or insecticide-treated targets. The efficiency of biconical traps (the standard control device), 1×1 m black targets and small 25×25 cm targets with flanking nets was compared using electrocuting sampling methods. The work was done on Glossina tachinoides and G. palpalis gambiensis (Burkina Faso), G. fuscipes quanzensis (Democratic Republic of Congo), G. f. martinii (Tanzania) and G. f. fuscipes (Kenya). The killing effectiveness (measured as the catch per m2 of cloth) for small targets plus flanking nets is 5.5–15X greater than for 1 m2 targets and 8.6–37.5X greater than for biconical traps. This has important implications for the costs of control of the Riverine group of tsetse vectors of sleeping sickness.
| Sleeping Sickness (Human African Trypanosomiasis) is a serious threat to health and development in sub-Saharan Africa. Currently there are no vaccines or prophylactic drugs available to prevent contraction of the disease. Consequently vector control is the only method of disease prevention. In many areas, especially those lacking high densities of cattle, the only control option for routine use against tsetse flies are insecticide-treated targets or biconical traps. However, these methods in their current form are often too expensive for routine use against the riverine tsetse species that are the major vectors of sleeping sickness. Our aim is to develop a more cost-effective device than those currently available. Working on four species of tsetse fly we have shown that a small 25×25 cm target with adjacent flanking net was up to 38x more cost-effective at killing tsetse flies than existing devices. These findings suggest that this new technology may make vector control in HAT foci an affordable option.
| African sleeping sickness or Human African Trypanosomiasis (HAT) is endemic to 36 countries in sub-Saharan Africa covering 9 million km2 with 60 million of the 400 million inhabitants at risk of the disease. Africa has emerged from a recent sleeping sickness epidemic. In 1997 about 450,000 people were afflicted [1] which has now been reduced to about 70,000 cases per year [2], [3]. Two forms of the disease exists, the Rhodesian (or East African) form being more acute and the Gambian form more chronic. Both these forms of the disease are fatal if left untreated and has an impact of 1.59M DALYs (disability adjusted life years). The related disease (nagana) in domesticated animals causes estimated losses to African agriculture of US$4.5bn per year [4]. In 2000 the African Union recognized trypanosomiasis as “one of Africa's' greatest constraints to socio-economic development” [5]. The trypanosomes causing HAT are transmitted by tsetse flies, particularly those of the Riverine (Palpalis) group. Antigenic variation in the trypanosome makes it unlikely that an effective vaccine will be produced in the foreseeable future. The available drugs are too toxic for prophylactic use. Consequently the only means of preventing the disease is vector control although this is not routinely practiced largely because of the cost.
Drug treatment of HAT is in a parlous state. The drugs available were developed many years ago and their toxicity and consequent human mortality allied to the increasing resistance to the drugs is a great worry [6]. Recent introduction of Nifurtimox Eflornithine Combination Therapy (NECT) has improved the situation but there is serious concern that no other drug for stage II treatment is in reserve should this fail. Vector control is essential for control of the Rhodesiense form of the disease [7] and can play a valuable role in support of case detection and treatment programmes for the Gambiense form of the disease especially in areas of high tsetse challenge when case detection and treatment alone is insufficient for control to be achieved [8], [9]. Given worries about the sustainability of case detection and treatment it is essential that effective vector control measures are available.
A major obstacle in control programmes against Riverine tsetse is cost. Consequently, for the reasons given above, cheaper control techniques are needed. A standard method for control of Riverine tsetse is to use biconical traps, treated or untreated with insecticide or large insecticide-treated targets [9], [10], [11], [12]. Because of their size both are expensive to make and deploy at the high densities required (10–30+/km2). Our aim is to develop a more cost-efficient device than the standard biconical trap or 1 m2 targets. Work is underway on developing artificial odour attractants to improve device efficiency [13]. Other studies have looked for improvements in the colour and shape of targets and traps [14], [15], [16]. However, few studies have focused on reduction of size of targets as a way to achieve better cost efficiency. Recent work on G. f. fuscipes [17] has shown the potential for a dramatic reduction in target size promising a considerable cost saving in control programmes against Riverine tsetse.
Crudely combining data for the number of HAT cases by country [18] and maps of potential distribution of tsetse flies [19] suggests than >90% of current HAT transmission is being caused by a small number of tsetse flies especially G. fuscipes fuscipes (Uganda, Sudan, Congo Brazzaville, Central African Republic), Glossina fuscipes quanzensis (DRC, Angola, Congo Brazzaville) with smaller number being transmitted by G. palpalis gambiensis and G. p. palpalis on the coast of West Africa. In this work we have expanded studies on target size to four other species of Riverine tsetse including the very important vectors G. f. quanzensis and G. p. gambiensis. In addition we have investigated the efficacy of the common practise of insecticide-treating biconical traps in the belief that this increases the number of tsetse they kill beyond those actually trapped by the device [20].
We conducted studies in each country during periods considered to be most appropriate in terms of fly abundance, accessibility, time and logistics. In doing so we could not investigate the effect of long-term seasonality on the efficiency of the different devices, nor was this the object of the current study.
Studies were undertaken on Glossina tachinoides and G. palpalis gambiensis along the lower Comoe river at Folonzo (09° 54′ N, 04° 36′ W) in southern Burkina Faso, between January and May 2009. The two species are sympatric here, along with G. m. submorsitans and G. medicorum. Additional studies on G. p. gambiensis were conducted along the Mouhoun river near Solenzo (12°14′ N, 04°23′ W), in western Burkina Faso, from January to June and in November 2009. See [21] for further details of the site.
Studies were undertaken on G. fuscipes quanzensis in July 2009 near the Lukaya river (4° 29′ S, 15° 18′ E), ∼20 km south east of Kinshasa, Democratic Republic of Congo. See [13] for further details of the site.
Studies were undertaken on G. f. martinii in November 2009 in the Gombe National Park (4° 38′ S, 29° 37′ E) on the shore of Lake Tanganyika, Tanzania. The area receives an annual rainfall of 760–1200 mm and is in a protected area of tropical rain and highland forest. There are several game species in the research area, including bushpigs (Potamochoerus porcus), monitor lizards (Varanus niloticus), bushbuck (Tragelaphus scriptus), olive baboons (Papio anubis), chimpanzees (Pan troglodytes) and various species of monkey and snake. G. brevipalpis also occurs here.
Studies were performed on Glossina f. fuscipes from September to November 2010 on the 0.5 km2 of Chamaunga Island (00 25′ S, 34013′ E), Lake Victoria, Kenya. See [13], [17] for further details of the site.
Square black targets (1×1 m) were compared for their ability to kill tsetse flies with targets 1/16th the size (0.25×0.25 m) and with a standard biconical trap [12] (Fig. 1). Targets were made from black cotton cloth. Electrocuting grids were fitted over fine black netting and these were placed next to targets and traps where they intercepted flies in flight – these devices are called flanking nets. The fine black polyester net (Quality no. 166, Swisstulle, Nottingham, UK) and the blackened 0.2 mm diameter electrocuting wires of the electric net are effectively invisible to tsetse [22], [23]. Electrocuting grids were also placed over the black cloth target. Electrocuted flies fell into trays of soapy water below the grids. All treatments were simultaneously compared with and without flanking nets [14], [17], allowing us to measure efficiency of the devices (i.e., the catch of the black cloth target or the catch inside the trap as a percentage of the total number of flies arriving in the vicinity of the device). The total number of visiting flies was taken to be the catch in the trap or on the target, plus the catch on each flanking net.
Experiments ran for 12 days each and were carried out during peak activity times of each tsetse species during the period of this study: for G. tachinoides and G. p. gambiensis from 08:00–12:00; for G. f. martinii from 10:00–14:00; for G. f. quanzensis from 10:30–14:30. The standard experimental design was a series of Latin-squares of treatments x days x sites, with sites at least 50 m apart. Analyses of variance were performed on log detransformed catches and these are discussed in the text.
Three experiments were conducted to assess the responses of tsetse to 3-dimensional objects. These studies were conducted with G. f. fuscipes only. The first experiment measured both the numbers of G. f. fuscipes caught on electrified 3-dimensional objects (3DO) (e.g. biconical traps) and the numbers of flies circulating but not contacting such objects. Due to difficulty in covering the conical parts of the biconical trap with electrified grids, a comparable 3 dimensional trap (Fig.2) was made which has flat surfaces. The first experiment compared a fully electrified 3DO consisting of three 0.5×1 m electrified grids arranged in a triangular fashion (Fig. 2) and killing all flies coming into contact with the grid, with a similar 3DO (not electrified) but with an adjacent electrified flanking net which intercepted and killed all circling flies. Each of the grids in the 3DO had a blue cotton cloth insert with a central oblong (15×25 cm) piece of black cloth (to simulate the entrance of a biconical trap). This experiment allowed us to compare numbers of flies attracted to and directly landing on a 3-D object against those flies attracted to, but only circling the object and getting caught on the flanking net. The experiment ran for 12 days in a 2×2 Latin square, from 09:00–12:00.
A second experiment was done with a single flanking net (0.5×1 m) adjacent to a biconical trap to intercept circling flies (Fig 1 image on right), compared against a single biconical trap and against a small blue cotton target (25×25 cm) with an adjacent flanking net (25×25 cm). The small target was also used to compare the efficiency of this small device compared to biconical traps. The reason a blue cotton target was used for these experiments and not black as in the size reduction study, is because blue proved to be a better attractant than black for G. f. fuscipes [14], [17] and this type of tiny target is being considered for control purposes.
For the third experiment, we compared a biconical trap closely surrounded with four flanking nets (0.5×1 m) to intercept all flies coming close to the trap as if to land. This was compared against a normal biconical trap as well as a small blue cotton target (25×25 cm) with an adjacent flanking net (25×25 cm). This experiment indicated the number of flies attracted to a biconical trap, but killed on the flanking nets before they could enter or land, compared against the numbers of flies in the top-cage of the standard trap. Again the small target with flanking net was used as control.
Four experiments evaluated the optimal flanking net width for use with a small 25×25 cm target. The first experiment investigated how closely G. f. fuscipes circle around a 25×25 cm blue target. This target was used with a 25×100 cm flanking net for 12 days with the collection tray divided into sections 10 cm wide. This determined where flies first touched the flanking net to give an initial indication of the optimal width of a netting panel. Second, we compared flanking nets of 25 cm, 50 cm and 75 cm widths, in a 3×3 Latin square design for 24 days. Third, a 25×25 cm blue target with the same size flanking net was compared with a 12.5×25 cm target with 12.5×25 cm flanking net for 12 days. Finally, a 25×25 cm blue target with 25×25 cm flanking net (all electrified) was compared against an un-electrified 25×25 cm blue target with a 25×25 cm electrified flanking net, in a 2×2 Latin square design for 12 days.
Catches for all four tsetse species from the devices listed are shown in Table 1. Below we expand and emphasise some of the data which we feel are the most important for the production of more cost effective tsetse killing devices.
The small target with flanking net uses 1/8th of the material in the large 1 m2 target. From Table 1 it can be seen that deploying the available cloth in the form of small rather than large targets will kill more tsetse flies per dollar spent. Female flies are the main target of control operations. If we consider just females from Table 1 then we see that for G. p. gambiensis the catch per m2 for small targets plus flanking nets is between 6.5× (Folonzo), and 8.7× (Solenzo) greater than that for 1 m2 targets. Corresponding figures for G. f. quanzensis are 5.5×, 5.8× for G. tachinoides and 15× for G. f. martinii, although in the last case the samples sizes are small. Figures for male flies show even greater potential for small targets. These findings are in agreement with those from a previous study on G. f. fuscipes [17].
The small target with flanking net uses 1/24th of the material in the biconical trap. From Table 1 we can see that deploying the available cloth in the form of a small target rather than a biconical trap will kill more tsetse flies per dollar spent. Female flies are the main target of control operations. If we consider just females from Table 1 then we see that for G. p. gambiensis the catch per m2 for small targets plus flanking nets is between 22.8× (Solenzo) and 37.5× (Folonzo) greater than that for biconical traps. Corresponding figures for G. f. quanzensis are 22×, for G. tachinoides 8.6× and for G. f. martinii it was impossible to determine as the biconical trap failed to catch any flies. Figures for male flies show even greater potential for small targets. Again, these data are in agreement with those from a previous study on G. f. fuscipes [17].
Investigations into the behaviour of G. f. fuscipes towards a rectangular blue and black 3-D object showed that 2.6x more G. f. fuscipes females circled (mean = 10.9) around the object than landed (mean = 4.2, s.e.d. = 0.09, P = 0.001; for ANOVA see Table S1, experiment 1). For male G. f. fuscipes catches of landing flies (mean = 4.5) were roughly equal to the circling flies (mean = 4.9, P = 0.6, s.e.d. = 0.05). When using the biconical trap as a 3-D object the majority of G. f. fuscipes circle around the trap but do not enter as can be seen below. Compared to the standard trap, a trap surrounded with four flanking nets caught 4.5x more female G. f. fuscipes (mean = 18.1, s.e.d. = 0.11, P<0.001; for ANOVA see Table S1, experiment 2) and a trap with a single adjacent flanking net caught 2.9x more females (mean = 12.2, s.e.d. = 0.06, P<0.001; for ANOVA see Table S1, experiment 3). Male flies also circled more around the trap, with 3.6x more males caught on the single flanking net (mean = 12.2, s.e.d. = 0.06, P = <0.001) and 2.6x more caught on the four flanking nets closely surrounding the trap (mean = 9.2, s.e.d. = 0.1, P<0.001), than were caught inside the standard trap. These data showed that up to 80% of G. f. fuscipes, especially females, are circling the trap and not landing or entering giving the biconical trap only about a 20% efficiency.
Comparing the efficiency of the devices for inducing landing and entering responses, the biconical trap again performed poorly (Fig. 3). Only 26% of the G. tachinoides and 32% G. p. gambiensis attracted to the trap actually entered it. Trap efficiency was even lower for G. f. quanzensis (18%), with the majority of flies circling around but not entering. Catches of G. f. martinii were too low to allow for analysis of its landing and trap-entry responses. In contrast, the efficiency of the large target (i.e. landing response) was much better. Fifty-five percent of G. tachinoides, 38% of G. f. quanzensis and 45% to 58% of G. p. gambiensis that were attracted to the target landed on the black cloth. The small black target with flanking net also induced a poor landing response on the black cloth (Fig. 3), indicating the importance of a flanking net to maintain the killing efficiency of the small target. For example, catches of G. tachinoides declined by 88% and G. f. quanzensis by 83% in the absence of this netting (i.e. catches on the 0.25×0.25 m black target alone), while G. p. gambiensis were 50–90% lower without the small flanking net (Table 1).
Studies to optimize the flanking net width showed that G. f. fuscipes circled closely around the small blue target. Sixty one percent (n = 32, s.e.d. = 0.1) of females and 77% (n = 24, s.e.d. = 0.07) of males were caught on the first 30 cm of flanking net adjacent to the target. A further 23% (n = 12) female and 21% (n = 7) male flies circled up to 50 cm away from the target. The remaining few flies were caught 50–80 cm away, with no flies caught between 80–100 cm. The subsequent experiments with flanking nets of various width showed no difference in catches between the standard 25 cm flanking net (mean = 14.4, sed = 0.05, P = 0.07 for difference between means), the medium 50 cm flanking net (mean = 16.7), or the long 75 cm flank net (mean = 20.3). A smaller flanking net of 12.5×25 cm resulted in a 66% decrease in catches. Equal numbers of flies were caught by the electrified flanking net (mean = 10.4, s.e.d. = 0.09, P = 0.9) adjacent to the un-electrified small blue target, as were caught by the completely electrified target and flank net (mean = 10.2) . This suggests that savings could be made by putting insecticide only on the flanking net.
The catch of tsetse increases with target size but the increase is not in proportion to the increase in surface area. So, paradoxically, it is more cost efficient to deploy the available cloth in the form of small rather than large targets . Tiny targets plus flanking nets use 1/8 and 1/24 the amount of materials required respectively for the large 1 m2 targets or biconical traps which are currently used in control programmes. Despite this they are comparable or superior to these much larger devices in killing G. p. gambiensis, G. f. quanzensis, G. f. martini, (Table 1) and G. f. fuscipes [17]. Clearly this means that considerable cost efficiencies are possible in using these new devices as reflected in the tsetse killed per unit area of cloth (Table 1). For example, concentrating only on female tsetse which are the major target of control programmes, the killing effectiveness measured as the catch per m2 of cloth for small targets plus flanking nets is 5.5–15× greater than that for 1 m2 targets. In comparison to biconical traps, the killing efficiency of small targets plus flanking nets is 8.6–37.5X greater . The tsetse species studied here are responsible for the transmission of virtually all gambiense-form HAT, which represents >90% of all cases of HAT. Hence, the cost savings implied by the above are available to most sleeping sickness control programmes.
Comparison with other tsetse species on which the effects of target size has been studied, is limited to the savannah tsetse. For G. pallidipes and G. morsitans, a target much less than about 1 m2 is strongly contra-indicated [24], [25], [26] due to low attractiveness. This is in strong contrast to our results on Riverine species shown here and in a previous study on G. f. fuscipes [17]. The underlying behavioural differences between Riverine and Savanna tsetse which underpin these findings remain to be explained.
An essential part of the small target is the flanking net, as catches of G. tachinoides, G. p. gambiensis and G. f. quanzensis declined by 88%, 67–91% and 83% respectively, in the absence of netting. This illustrates the importance of small panels of fine, insecticide-treated net attached to the side of the small cloth targets to intercept the flies that circle around the cloth. This principle has been used as part of control targets for savannah species [24] and recommended for control of G. p. gambiensis and G. tachinoides [18]. However, large panels of netting are prone to damage which renders large 1 m2 devices fixed with a netting panel inefficient. With the tiny targets recommended by this work, the small flanking net is much less likely to be damaged. In addition, suitable netting now available on the market, particularly insecticide pre-impregnated polyethylene netting, is stronger and more durable.
A common practice in the control community has been to use insecticide-treated traps in the belief that many more flies will land on the outside of traps than are caught by them [27]. However, there are scant direct data supporting this practice and hence it is not universally accepted. Observations of G. morsitans and G. pallidipes showed that only 47–30% of tsetse approaching a trap landed on it or entered it, i.e.the majority (53–70%) of tsetse visiting a trap did not contact it [28]. Our data show that the efficiency (proportion of the total flies attracted to the trap which are actually caught by it) is low (e.g. 26% G. tachinoides; 32% G. p. gambiensis; 10% G. f. quanzensis). Let us assume for the sake of argument that 100% of the flies circulating the biconical trap in our experiments land on it and collect a lethal dose of insecticide. Even then, using the data from Table 1, the flies killed per 1 m2 of cloth will be greater for small targets plus a flanking net than for biconical traps (2.2X G. tachinoides; 12X and 7.3X G. p. gambiensis; 2.2X G. f. quanzensis). In fact, the results show that the catch from the 3-D target with a flanking net was 1.8x that of the target alone (15.9 tsetse/day vs. 8.7 tsetse/day) suggesting that not all tsetse approaching the object landed on it. The efficiency of the trap-like object (55%) is slightly greater than a trap (31%) suggesting that marginally more flies may land on a trap than are captured by it. If that figure is common to all species it would roughly double the kill per m2 figures given above in this paragraph. Clearly small targets plus flanking nets are a more efficient means of killing tsetse than using either 1 m2 targets or biconical traps whether the latter are treated with insecticide or not.
This work clearly demonstrates the potential savings for tsetse control operations in terms of reduced costs of materials and insecticide associated with the manufacture of small targets. In addition these devices are likely to offer two further advantages. First, the small targets would be considerably easier and cheaper to transport to the field [28] offering further considerable cost savings to control campaigns. For example, the tiny targets can be carried in a backpack and deployed rapidly by a single person. Second, while large targets and their associated doses of insecticide have been shown to have little impact on ecology [29] and to be unobtrusive in national parks [30], the small targets could be expected to be even better in both these respects. A potential problem for small targets is that they may be easily obscured by vegetation which may reduce their efficiency. Further work is underway to look at the importance of this and the indications are that this is very much smaller problem for Palpalis group flies than for Morsitans group flies (Esterhuizen et al., in preparation).
In conclusion, it appears that the use of small targets demands a full scale field trial while further research should be performed to refine them and to explore their applicability against a wider range of tsetse species and in other areas.
|
10.1371/journal.pmed.1002136 | Clonal Evolutionary Analysis during HER2 Blockade in HER2-Positive Inflammatory Breast Cancer: A Phase II Open-Label Clinical Trial of Afatinib +/- Vinorelbine | Inflammatory breast cancer (IBC) is a rare, aggressive form of breast cancer associated with HER2 amplification, with high risk of metastasis and an estimated median survival of 2.9 y. We performed an open-label, single-arm phase II clinical trial (ClinicalTrials.gov NCT01325428) to investigate the efficacy and safety of afatinib, an irreversible ErbB family inhibitor, alone and in combination with vinorelbine in patients with HER2-positive IBC. This trial included prospectively planned exome analysis before and after afatinib monotherapy.
HER2-positive IBC patients received afatinib 40 mg daily until progression, and thereafter afatinib 40 mg daily and intravenous vinorelbine 25 mg/m2 weekly. The primary endpoint was clinical benefit; secondary endpoints were objective response (OR), duration of OR, and progression-free survival (PFS). Of 26 patients treated with afatinib monotherapy, clinical benefit was achieved in 9 patients (35%), 0 of 7 trastuzumab-treated patients and 9 of 19 trastuzumab-naïve patients. Following disease progression, 10 patients received afatinib plus vinorelbine, and clinical benefit was achieved in 2 of 4 trastuzumab-treated and 0 of 6 trastuzumab-naïve patients. All patients had treatment-related adverse events (AEs). Whole-exome sequencing of tumour biopsies taken before treatment and following disease progression on afatinib monotherapy was performed to assess the mutational landscape of IBC and evolutionary trajectories during therapy. Compared to a cohort of The Cancer Genome Atlas (TCGA) patients with HER2-positive non-IBC, HER2-positive IBC patients had significantly higher mutational and neoantigenic burden, more frequent gain-of-function TP53 mutations and a recurrent 11q13.5 amplification overlapping PAK1. Planned exploratory analysis revealed that trastuzumab-naïve patients with tumours harbouring somatic activation of PI3K/Akt signalling had significantly shorter PFS compared to those without (p = 0.03). High genomic concordance between biopsies taken before and following afatinib resistance was observed with stable clonal structures in non-responding tumours, and evidence of branched evolution in 8 of 9 tumours analysed. Recruitment to the trial was terminated early following the LUX-Breast 1 trial, which showed that afatinib combined with vinorelbine had similar PFS and OR rates to trastuzumab plus vinorelbine but shorter overall survival (OS), and was less tolerable. The main limitations of this study are that the results should be interpreted with caution given the relatively small patient cohort and the potential for tumour sampling bias between pre- and post-treatment tumour biopsies.
Afatinib, with or without vinorelbine, showed activity in trastuzumab-naïve HER2-positive IBC patients in a planned subgroup analysis. HER2-positive IBC is characterized by frequent TP53 gain-of-function mutations and a high mutational burden. The high mutational load associated with HER2-positive IBC suggests a potential role for checkpoint inhibitor therapy in this disease.
ClinicalTrials.gov NCT01325428
| Inflammatory breast cancer (IBC) is a rare and poorly understood form of breast cancer that grows and spreads very quickly. Fifty percent of IBC cases are HER2-positive.
Afatinib is an investigational drug that showed promise in early-stage trials in the setting of HER2-positive metastatic breast cancer.
Our study was designed to look at how effective and safe afatinib is in treating HER2-positive IBC patients, and to elucidate how afatinib treatment affects the tumours at the genomic level.
We recruited 26 patients for this study and administered afatinib daily, and 10 patients went on to be treated with daily afatinib and weekly vinorelbine, a chemotherapy drug, upon disease progression.
Thirty-five percent (9 of 26) and 20% (2 of 10) of patients had clinical benefit from being treated with afatinib monotherapy and afatinib plus vinorelbine, respectively.
We sequenced tumour biopsies before and after afatinib treatment and found that IBC has a higher mutational load and more frequent mutations in the well-known cancer gene TP53, compared to non-IBC.
We did not identify any single gene or mutation that led to afatinib resistance, and biopsies before and after treatment were very similar genetically.
Afatinib appears to be clinically active in HER2-positive IBC, albeit in a relatively small patient cohort.
The high mutational load in IBC suggests that checkpoint inhibitors, a type of cancer immunotherapy, might potentially be an effective way of treating patients.
| Inflammatory breast cancer (IBC) is a rare, aggressive form of breast cancer that accounts for around 1%–6% of breast cancers [1–4]. IBC tends to affect younger women and has a high risk of local and distant metastasis. Prognosis is poor, with median survival estimated at 2.9 y in IBC patients versus 6.4 y in those with non-inflammatory, locally advanced breast cancer [3,5]. Current management of IBC involves a combination of anthracycline and taxane-based chemotherapy in the neoadjuvant setting, followed by surgery, adjuvant chemotherapy, or radiotherapy [6].
IBC is thought to be a biologically distinct form of breast cancer, commonly lacking oestrogen (ER) and progesterone (PgR) receptor expression [7]. A greater frequency of HER2 and EGFR overexpression among IBC cases has been reported, occurring in 50% and 30% of patients, respectively [8]. Genomic profiling techniques have led to the identification of genes that are potentially involved in disease development [9–11]; however, HER2-positive IBC has not been characterised through deep exome sequencing.
EGFR and HER2 have been shown to be involved in tumour growth and metastasis of IBC, and as such represent therapeutic targets [12]. Afatinib is a small molecule tyrosine kinase inhibitor that irreversibly and selectively blocks signalling from ErbB family members. Clinically, afatinib showed activity in phase II trials with HER2-positive breast cancer patients [13,14]. Most recently, in the phase III LUX-Breast 1 trial, afatinib combined with vinorelbine demonstrated similar progression-free survival (PFS) and objective response (OR) rates to trastuzumab plus vinorelbine in patients with HER2-positive metastatic breast cancer after failure on trastuzumab, but the afatinib-containing regimen was associated with shorter overall survival (OS) and was less tolerable [15].
We performed an open-label, single-arm phase II clinical trial to investigate the efficacy and safety of afatinib alone and in combination with vinorelbine following disease progression in patients with HER2-positive IBC. Recruitment to this trial was terminated early following the results of the LUX-Breast 1 trial. We carried out prospectively planned whole-exome sequencing of tumour biopsies at baseline and after progression on afatinib monotherapy to explore two questions: (1) what is the mutational landscape of HER2-positive IBC, and is it distinct from HER2-positive non-IBC; and (2) how does exposure to HER2 inhibition affect the evolution of IBC?
This was an open-label, phase II, multicentre trial of afatinib for the treatment of HER2-positive IBC (ClinicalTrials.gov NCT01325428, S1 and S2 Texts). Patients were treated with afatinib monotherapy until disease progression (Part A), and then afatinib and vinorelbine until disease progression (Part B).
PFS was assessed separately for Part A and Part B, and over the whole study. OS was only assessed over the whole study period. The primary endpoint was clinical benefit (defined as stable disease [SD] for ≥6 mo, partial response [PR], or complete response [CR]). Secondary endpoints were objective response (OR) and duration of OR and PFS; other endpoints included OS and safety.
Following the results of the LUX-Breast 1 trial, Part B was stopped and recruitment to the whole trial was stopped thereafter. Patients in Part A were informed that they would no longer be able to receive afatinib plus vinorelbine upon progression, and had to agree with the investigator regarding continuation of afatinib monotherapy. Patients in Part B who were deriving benefit from treatment could continue afatinib plus vinorelbine.
PR was considered to be confirmed if the criteria were met at least 4 wk later. SD had to be observed at least 42 days after first study drug administration in the respective part of the study to be considered for best overall response regardless of confirmation, and had to last for more than 182 d to qualify for clinical benefit.
The study was conducted in line with the Declaration of Helsinki, the International Conference on Harmonization Good Clinical Practice Guideline and approved by the local ethics committees (S1 Appendix). All patients provided written informed consent prior to study participation.
Female patients aged ≥18 y with investigator-confirmed IBC characterized by diffuse erythema and oedema (peau d’orange) with locally advanced or metastatic disease and histologically confirmed HER2-positive disease (i.e. immunohistochemistry [IHC] 3+ or IHC 2+ with FISH/SISH positivity) were eligible for the study (S1 Table). Patients were required to have an Eastern Cooperative Oncology Group (ECOG) status of 0–2 and life expectancy of ≥6 mo. Other exclusion criteria for the trial included: radiotherapy, chemotherapy, hormone therapy, immunotherapy, trastuzumab, or surgery (other than biopsy) within 2 wk prior to the first dose of afatinib in Part A, known pre-existing interstitial lung disease, active brain metastases, significant chronic or recent acute gastrointestinal disorders with diarrhoea as a major symptom, any other current malignancy or malignancy diagnosed or relapsed within the past 5 y (other than non-melanomatous skin cancer and in situ cervical cancer), inadequate bone marrow, and renal and liver functions.
In both parts of the study, patients received a single oral dose of afatinib 40 mg once daily until disease progression. The first dose was administered at the trial site, and subsequent doses were taken at home. Afatinib dose reductions were required for any drug-related grade ≥3 adverse events (AEs) and selected grade 2 AEs. The afatinib dose was reduced in 10 mg decrements to a minimum of 20 mg; all dose reductions were permanent. In Part B, patients received previously tolerated afatinib doses and additionally received short infusion (approximately 10 min) intravenous vinorelbine at a weekly dose of 25 mg/m2 in a 4-weekly course until disease progression. Vinorelbine treatment was administered at the trial site under the supervision of the investigator; treatment was withheld if platelet count was <100,000 cells/mm3 or absolute neutrophil count was <1,500 cells/mm3.
Tumour assessments were performed by computed tomography or magnetic resonance imaging at screening and every 8 wk after the first dose of afatinib. Investigators evaluated response according to Response Evaluation Criteria in Solid Tumors (RECIST) version 1.1. AEs were graded using Common Terminology Criteria for Adverse Events (CTCAE) version 3.0.
Tumour biopsies were obtained before afatinib treatment and on disease progression in Part A, snap frozen, and optimal cutting temperature compound (OCT)-embedded. Venous blood samples were obtained and genomic DNA was extracted. Whole-exome sequencing was performed on pre-treatment tumour biopsies, matched germline genomic DNA and post-treatment tumour biopsies according to the manufacturer’s protocol (Agilent SureSelect Human All Exon 50Mb Kit). Tumour and germline DNA were sequenced at the Beijing Genomics Institute on the Illumina HiSeq 2000 to an average depth of 396x and 157x, respectively (S2 Table).
Raw sequencing data were aligned to human genome sequence version hg19 using bwa (v0.5.9) [16], duplicates marked using Picard (v1.54), and indel realignment performed with GATK IndelRealigner (v1.0.6076) [17]. Somatic single nucleotide variant (SNV) calling was performed using VarScan2 (v2.3.7) [18], MuTect (v1.1.7) [19], Virmid (v1.1.0) [20], and Strelka (v1.0.14) [21]. SNVs called by ≥2 tools were further filtered for variant allele frequency (VAF) ≥5%. Small indels were identified using Pindel (v0.2.5a7) [22] and VarScan2 (v2.3.7). Indels called by both tools were further filtered for VAF ≥5%. Mutations in genes of interest were visualized with Oncoprints [23].
Tumour copy number aberrations, ploidy, and purity were determined using ASCAT 2 [24], which allows for exome sequencing data as input (available at https://github.com/Crick-CancerGenomics/ascat) (S3 Table). Some samples were excluded from copy number analysis due to low tumour content. Segmented copy number data were divided by sample mean ploidy and log2 transformed for GISTIC2.0 analysis [25]. Copy number segments were defined relative to ploidy as previously described [26]: amplification, gain, and loss were defined as log2(4/2), log2(2.5/2), and log2(1.5/2), respectively.
Genome-doubling status was determined as previously described [27]. Briefly, each sample, s, was represented as an aberration profile of major and minor allele copy numbers at chromosome arm resolution. The total number of aberrations (relative to diploid), Ns, and the probabilities of loss/gain for each allele at each chromosome arm, Ps, was calculated. Ten thousand simulations were run for each sample s, where Ns sequential aberrations, based on Ps, were applied to a diploid profile. A p-value for genome doubling was obtained by counting the percentage of simulations in which the proportion of chromosome arms with a major allele copy number ≥2 was higher than that observed in the sample.
The weighted Genomic Instability Index (wGII) was used to assess chromosomal instability [28]. Briefly, the percentage aberrant regions for each autosome was calculated separately and mean percentage aberration then calculated across all 22 chromosomes to account for variation in chromosome size, so that large chromosomes do not have a greater effect on the GII score than small chromosomes.
Mutational signatures were determined using the R package deconstructSigs [29]. Using this tool, the fraction of mutations in each of the 96 trinucleotide contexts was calculated, and the weighted combination of published signatures from [30] identified to most closely reconstruct the mutational profile of the sample.
The mutation copy number and cancer cell fraction of each mutation were calculated by integrating ASCAT-derived integer copy number and tumour purity estimates with the variant frequency as described in [31]. This was used as input for PyClone [32], which uses a hierarchical Bayesian Dirichlet process in order to infer clonal population structure. A modified version of PyClone was used as described in [33]; clusters with 3 or fewer SNVs were excluded.
HLA typing was performed with OptiType [34]. Nonsynonymous mutations were extracted from each tumour sample and translated into mutant peptide 9–11mers long [33]. Using the patient-specific HLA type, we used NetMHC (v2.8) [35] to predict the binding strength of each mutant and wildtype peptide to the respective MHC class I molecules. Somatic mutations that gave rise to peptides with a binding affinity of ≤500 nM were considered to be putatively neoantigenic.
The comparison to the HER2-positive non-IBC cohort is based upon data generated by The Cancer Genome Atlas (TCGA) Research Network: http://cancergenome.nih.gov/ [36]. Tumour samples were filtered for positive HER2 IHC status (n = 131, S4 Table). For the matched cohort, a subset was selected by matching cases based on age (±10 years), ER, and PgR status; this allowed only for a one-to-one matching due to the limited number of TCGA cases available.
Analyses of efficacy and safety in this trial were descriptive and exploratory. A sample size of 40 patients was selected for this study; assuming an underlying clinical benefit rate of 50%, 40 patients would provide more than a 90% probability of observing a clinical benefit rate of at least 40%. Analyses of clinical benefit rate (CBR) and OR rate were planned for the following subgroups: hormone receptor (ER and PgR), EGFR status, new brain metastases, patients presenting with target lesions only versus those with non-target lesions only versus those with both, and prior trastuzumab therapy. Exploratory analyses using genomic data were planned to search for predictive markers of response and resistance to afatinib.
MutSigCV (v1.3) [37] and GISTIC2.0 [25] were used to determine mutational significance of somatic SNVs and somatic copy number alterations (SCNAs). Multiple-testing corrections in these tests were carried out using the Benjamini-Hochberg false discovery rate method. Mann-Whitney and Fisher’s exact test were used for comparison between two groups. Survival curves were estimated using the Kaplan-Meier method, and the log-rank test was used to test for significance.
The study was performed at 14 centres in seven countries between December 2011 and November 2014. Twenty-nine patients were screened, and 26 received afatinib monotherapy; of these, 10 patients continued into Part B of the study (Fig 1). Twenty-four of 26 patients had metastatic disease at study inclusion. Patient demographics at baseline are shown in Table 1.
Nine (35%) of 26 treated patients had confirmed clinical benefit with afatinib monotherapy (eight PRs and one SD of ≥6 months; Table 2, S5 Table). Three patients had an unconfirmed PR, resulting in an overall response rate (ORR) of 42% (n = 11). Twenty (77%) patients progressed or died on afatinib monotherapy; median PFS was 110.5 days (95% CI 58.0–386.0). In total, there were three on-treatment and one post-study deaths. Planned subgroup analyses were performed (S1 Fig). Clinical benefit with afatinib monotherapy was achieved in 0 of 7 trastuzumab-treated patients and 9 of 19 (47%) trastuzumab-naïve patients. Median PFS with afatinib monotherapy was apparently shorter in trastuzumab-treated patients versus trastuzumab-naïve patients (64 versus 151 days, p-value = 0.099, log-rank test; S2 Fig).
Following progression on afatinib monotherapy, ten patients received afatinib plus vinorelbine (Part B). Confirmed clinical benefit was achieved in two (20%; Table 2, S5 Table, S1 Fig); a further two patients had an unconfirmed PR, with an overall CBR rate of 40% (n = 4). Eight (80%) patients progressed or died. Median duration of PFS was 106.0 d (95% CI 36.0–190.0).
OS was analysed across the whole study. Eleven (42%) patients died during the study, and median OS was 713.0 d. Median PFS across the whole study was shorter in trastuzumab-treated patients versus trastuzumab-naïve (136 versus 395 d, p-value = 0.024, log-rank test, S3 Fig).
Median duration of exposure to treatment was 111 d (range: 17–700) in Part A and 84.5 d (range: 42–237) in Part B. All patients had treatment-related AEs (Table 3).
Twenty-two of 26 patients in Part A had tumour biopsy material suitable for whole-exome sequencing analysis (Fig 2). Overall, we identified an average of 134.5 (range: 30–468) somatic coding mutations (Fig 2A, S6 Table). The most commonly mutated gene was TP53 (MutSig q-value = 1.68x10-11); 86.4% (19/22) of the tumours harboured a somatic mutation in TP53 (S7 Table). Strikingly, five patients had gain-of-function TP53 mutations at hot-spot residue p.R248 (S4 Fig). Planned exploratory analyses showed that OS was non-significantly shorter in patients carrying TP53 p.R248 mutations pre-treatment versus those with loss-of-function (nonsense, frame-shift, splice site) mutations (398 versus 652 d, p-value = 0.626, log-rank test). Patients IBC007 and IBC001 had much higher numbers of somatic SNVs compared to the rest of the cohort (468 and 393, respectively) but did not have mutations in known DNA mismatch repair genes.
Mutations in the PI3K/AKT/mTOR pathway are frequent in breast cancer, and activation of this pathway via molecular aberrations in PIK3CA, PIK3CB, PIK3R1, AKT, TSC1/2, and PTEN promotes resistance to HER2-targeted therapies [39–41]. Seven patients harboured PIK3CA mutations, including four with hotspot mutation p.H1047R [42]. IBC007 harboured an activating AKT1 p.E17K mutation, and IBC025 carried an activating ERBB2 p.V777L mutation (Fig 2B) [43,44]. No other somatic mutations in this pathway were identified.
In order to gain insight into the mutational processes shaping the IBC landscape, we utilized previously extracted mutational signatures and applied them to the IBC cohort (Fig 2A, S5A Fig) [29,30]. Signature 1A, which was previously associated with age of diagnosis [45], accounted for the majority of mutations (64.2% ± 26.1). Signatures 2 and 13, attributed to activity of the APOBEC family of cytidine deaminases, were together present in 64% (14/22) of tumours (16.4% ± 19.7% of somatic mutations). In particular, the excess of mutations in IBC001 and IBC007 appear to be driven by APOBEC mutagenesis (S5B Fig), which was observed in both clonal and subclonal mutations for these samples (S5C Fig).
SCNA calling was possible in 20 of 22 tumours (S8 Table). Seventy percent (14/20) of the tumours had undergone whole-genome doubling and had higher genomic instability scores (wGII) compared to non-genome-doubled tumours (0.54 ± 0.18 versus 0.31 ± 0.06, p-value = 4.6x10-3, Mann-Whitney) (Fig 2A). Even though all IBC patients were HER2-positive via IHC or FISH (S1 Table), only 16 of 20 tumours were called as having ERBB2 amplification; two tumours (IBC011 and IBC029) harboured gains and two tumours (IBC007 and IBC028) had neither amplification nor gain of ERBB2 (Fig 2B). It is possible that these four tumours could represent false negatives due to reasons such as sampling bias caused by intra-tumour heterogeneity or normal tissue contamination. Sixty percent (12/20) of tumours had EGFR gains (11 gains, 1 amplification), consistent with previous reports [8]; 10 tumours had PTEN loss (Fig 2B). GISTIC [25] analysis revealed recurrent focal amplifications across 6 loci, including 17q12 (q-value = 9.22x10-13), 8q24.21 (q-value = 4.89x10-3), and 1q32.1 (q-value = 5.80x10-2) containing ERBB2, MYC, and MDM4, respectively (Fig 2C, S9 Table). Recurrent focal losses were identified across 12 chromosomal regions, including 11p5.15 (q-value = 2.09x10-2) containing SIRT3 and PHRF1.
We carried out planned exploratory analyses to identify predictive markers of response and resistance to afatinib. We did not identify an association between EGFR gains or HER2 amplifications and response to afatinib. Since activation of PI3K/Akt signalling is thought to impact the efficacy of HER2-targeted treatment [46–48], we focused on mutations in this pathway to explore any potential impact on PFS. We observed that somatic activation of this pathway (i.e. PIK3CA activating mutation or gain, ERBB2 activating mutation, PTEN deletion, AKT1 activating mutation) was significantly associated with shorter PFS in trastuzumab-naïve patients (p-value = 0.03, S6 Fig). Although activating mutations of the PI3K pathway have been reported as occurring more frequently in ER-positive breast tumours [40], we did not observe a difference in this small cohort (6/10 ER-positive versus 7/12 ER-negative). Unexpectedly, a trastuzumab-naïve patient (IBC024) harbouring a gain overlapping PIK3CA and PTEN heterozygous deletion at baseline showed a PR for 48 wk before disease progression.
To determine if there were significant differences in mutational profiles between HER2-positive IBC and HER2-positive non-IBC, we compared our results against a cohort of TCGA patients with HER2-positive breast cancer (n = 131, S4 Table) [36]. We observed that the average number of somatic protein-changing mutations per patient was higher in IBC than non-IBC patients (102.4 ± 89.4 versus 71.9 ± 115.3; p-value = 0.0107, Mann-Whitney) (Fig 2D).
Given that TP53 was the only significantly mutated gene identified in IBC, we compared the mutation burden of this gene between the 2 cohorts. We observed that TP53 mutations were significantly enriched in the IBC cohort compared to non-IBC (19/22 versus 53/131; p-value = 5.76x10-5, Fisher’s exact) [49], as were TP53 hotspot p.R248 mutations (5/19 versus 3/53; p-value = 0.026, Fisher’s exact) (Fig 2E). Consistent with the higher mutational load, IBC tumours also had a higher number of predicted neoantigens compared to non-IBC (49.59 ± 37.9 versus 31.0 ± 41.8, p-value = 8.39x10-4, Mann-Whitney) (Fig 2F). Similar to IBC, the most prevalent mutational processes among the non-IBC cohort were age and APOBEC-related, with similar distributions of these mutational signatures between the 2 cohorts (S5D Fig).
There were no significant differences in the proportion of genome-doubled tumours (14/20 versus 75/131, p-value = 0.34, Fisher’s exact) or wGII scores (0.47 versus 0.51, p-value = 0.3843, Mann-Whitney) between IBC and non-IBC tumours. Applying GISTIC to the non-IBC tumours, 5 of 6 recurrently amplified regions and all 12 recurrently deleted regions in IBC had wide-peak boundaries that overlapped with those of non-IBC tumours (S9 Table). Only the 11q13.5 amplification in IBC did not overlap with non-IBC, which includes PAK1, an oncogene that activates MAPK and MET signalling and regulates cell motility; interestingly, previous reports have associated IBC with MAPK hyperactivation [50,51].
Utilizing an age and ER/PgR status matched cohort (Methods), the results were concordant, with a higher burden of somatic protein-changing mutations, neoantigens, and TP53 mutations in IBC versus non-IBC (S7 Fig).
Among 13 tumour biopsies obtained following disease progression, we identified an average of 181.4 (range: 50–505) somatic mutations, of which 79.1% ± 12.0% were shared with baseline tumours (Fig 3A, S10 Table). The overall mutation burden in tumours following treatment was higher in post-treatment samples compared to pre-treatment samples (172.5 ± 136.7 versus 156.1 ± 151.9, p-value = 0.030, paired t test). No recurrent mutations were identified among newly arising mutations post-treatment, and no new mutations in PI3K/Akt pathway genes were identified, aside from a MTOR p.K30N mutation (variant of unknown significance) in IBC021.
Nine of 13 matched pairs had copy number data (S11 Table); all tumours had the same genome-doubling status pre- and post-treatment, and there was no difference in ploidy (2.9 versus 2.9, p-value = 0.80) or wGII scores (0.42 versus 0.46, p-value = 0.55) (S8 Fig, S3 Table) between pre- and post-treatment samples. ERBB2 amplification status appeared to change in two of nine patients, from gain to copy-neutral in IBC029 and from copy-neutral to gain in IBC007 (Fig 3A). Overall, SCNAs between paired samples (n = 9) were highly concordant, and unsupervised hierarchical clustering showed that tumour biopsies clustered by patient rather than treatment stage (S9 Fig).
Drug resistance may arise as a consequence of an evolutionary bottleneck, where a resistant subclone is selectively enriched during therapy [52]. We utilized previously described methods to compare the clonal architecture of tumours before and after treatment [32,53]. Of these nine patients, IBC021 was the only patient with confirmed clinical benefit. We observed in all patients a cluster of variants that was clonal in both pre- and post-treatment biopsies (cancer cell fraction [CCF] around 1.0 on both x and y axis in S10 Fig); all gain-of-function PIK3CA and TP53 mutations, when present in the tumour, belonged to this cluster. In eight of nine patients, we observed some evidence of branching evolution, with new clones identifiable in the post-treatment samples and others declining in frequency or disappearing (Fig 3B, S11 Fig). Interestingly, the majority of mutations identified after treatment were detected in the pre-treatment tumour biopsy at a similar CCF (S10 and S11 Figs), and the overall clonal composition in all 8 tumours remained largely similar between the two time points with little evidence of bottlenecking, consistent with the lack of confirmed benefit in these patients, aside from IBC021. In one patient (IBC007), new clones were not observed, but there were distinct clonal shifts; there was clonal expansion of two subclones from 2% to 38% and 22% to 81%, and the major clone decreased slightly from 96% to 78%; no known drivers were identified in the subclones (S10 Table). Importantly, we cannot exclude the possibility that the observed dynamics could be due to tumour sampling bias between pre- and post-treatment samples.
Longitudinal analysis of the genomic evolution of tumours during therapy can inform drug resistance mechanisms and the changing landscape of disease over time. Here, we report the first prospectively planned clinical trial in IBC with genomic analysis, and the first assessment of afatinib with or without vinorelbine in patients with HER2-positive IBC.
Afatinib monotherapy demonstrated activity in patients with HER2-positive IBC, with nine (35%) patients achieving clinical benefit and median PFS of 110.5 d. This is concordant with data from a phase II trial assessing lapatinib 1500 mg daily in 126 patients with relapsed or refractory HER2-positive IBC, in which no patients had a CR but 49 (39%) had a PR and median PFS was 102.2 d [54]. Following progression on afatinib monotherapy, two (20%) patients achieved clinical benefit with addition of vinorelbine, and median PFS in Part B was 106.0 d.
The most common treatment-related events reported during the trial were diarrhoea, rash, and decreased appetite in Part A, and neutropenia, diarrhoea, nausea, and anaemia in Part B. Overall, the safety profile observed was generally consistent with previously published data on afatinib and vinorelbine. Importantly, this trial included pre-planned exome analysis of tumour biopsies at two time-points: before treatment and at disease progression. To our knowledge, this is the first report characterising IBC through exome sequencing. We identified a high incidence of TP53 mutations, as reported previously [49], and an enrichment of p.R248 hotspot DNA-contact mutations that promote nuclear accumulation of p53 [55–57]; cellular and animal studies indicate that these gain-of-function mutations induce increased invasion, chemoresistance and decreased survival [58–60]. Our results showed a non-significant reduction in OS in IBC patients carrying TP53 p.R248 mutations, consistent with previous analysis [61] and reports of nuclear p53 overexpression representing an adverse prognostic marker in IBC [62–64].
We identified recurrent focal gains across 6 loci and losses across 12 regions, including 11p5.15 containing SIRT3 and PHRF1 (also identified in the non-IBC cohort). SIRT3 is deleted in 40% of human breast tumours, and loss of SIRT3 increases reactive oxygen species production and HIF-1a stabilization [65]. PHRF1 functions as a tumour suppressor by promoting the TGF-beta cytostatic programme [66]; a recent transcriptomic study identified reduced TGF-beta signalling as a specific gene expression signature of IBC compared to non-IBC [67]. Comparing IBC to non-IBC, the only different recurrent focal SCNA was the amplification of 11q13.5 containing PAK1 in IBC; PAK1 is an oncogene that activates MAPK and MET signalling and regulates cell motility, and previous reports have associated IBC with MAPK hyperactivation [50,51].
We compared tumours before and after afatinib monotherapy to investigate potential drivers of resistance. The tumour pairs displayed a high degree of genetic relatedness, both in terms of point mutations and large-scale genomic aberrations. We did not observe changes in ERBB2 amplification status in the majority (7/9 or 78%) of our tumours, consistent with previous reports of loss of HER2-positivity occurring in only 12%–32% of patients undergoing anti-HER2 therapy [68–71]. In the two patients who appeared to undergo a change in amplification status, we are unable to conclude if the lack of ERBB2 amplification (in the pre-treatment biopsy for IBC007 and in the post-treatment biopsy for IBC029) was due to technical limitations of exome sequencing, sampling bias, or selection of a HER2-negative subclone during therapy (in the case of IBC029).
In contrast to EGFR mutant lung adenocarcinomas, in which the T790M gatekeeper mutation is commonly selected following EGFR inhibitor exposure [72], there was no evidence of selection for mutations in specific genes in the post-treatment IBC tumours. Eight out of 9 tumour pairs displayed branching evolution, with new clones emerging and others disappearing after treatment, possibly reflecting the differential effect that afatinib monotherapy had on the different subclones; it is worth noting that only 1 of 8 of these patients (IBC021) derived confirmed clinical benefit from afatinib monotherapy. It is also possible that subclones detected only in the pre- or post-treatment tumour biopsy in this study could be related to sampling bias or caused by the “illusion of clonality” derived from a single-region biopsy. Regardless, the majority of mutations in these tumours were shared between the two time points and possessed largely similar clonal compositions, concordant with previous reports in pre- and post-treatment samples of multiple myeloma and high-grade serous ovarian carcinoma [53,73]. IBC007 was the only tumour with an apparent shift in clonal structure, possibly reflecting random drift of tumour clones or sampling bias, given that this patient did not respond to afatinib monotherapy [32,53]. Immune checkpoint inhibitors have been shown to provide clinical benefit in a variety of cancers, including melanoma and lung cancer [74–77]. In particular, a high mutational load (>100 somatic nonsynonymous coding mutations) was reported as significantly correlated with improved OS in patients with metastatic melanoma treated with ipilimumab or tremelimumab [78]. Several clinical trials investigating the efficacy of checkpoint inhibitors have already been initiated in HER2-positive breast cancer (NCT02734004, NCT02605915, NCT02318901, NCT02403271) and HER2-positive gastric cancer (NCT02689284). The mutational burden in our study revealed an average of 102.4 nonsynonymous mutations in baseline HER2-positive IBC, above the threshold indicated for clinical benefit with anti-CTLA4 therapy [78]. The high mutational and neoantigenic load associated with HER2-positive IBC suggests a potential role for checkpoint inhibitor therapy in this disease.
Following the results of the LUX-Breast 1 trial, recruitment to this study was terminated early. As such, a limitation of this study is the relatively small sample size of HER2-positive IBC patients, making it difficult to draw robust conclusions regarding clinical efficacy of afatinib in this disease. Furthermore, single-region biopsies could be leading to underestimation of tumoural heterogeneity and clonal dynamics.
In conclusion, this phase II trial demonstrated that afatinib, with or without vinorelbine, showed activity in patients with HER2-positive IBC in trastuzumab-naïve patients, albeit in a small patient cohort. This is one of the first clinical trials to fully and prospectively integrate longitudinal exome sequencing with drug development. HER2-positive IBC is characterised by a higher mutational and neoantigenic burden and greater incidence of TP53 mutations compared to HER2-positive non-IBC. PI3K pathway activation was associated with poorer outcomes on afatinib therapy. Analysis of pre- and post-afatinib monotherapy tumour biopsies did not identify major dynamics of tumour sublcones or recurrent somatic mutations driving resistance. Epigenetic and tumour microenvironmental changes [79,80] may contribute to drug resistance in IBC and should be investigated further in future trials.
This study provides a proof of principle that prospective planning of genomic analysis in clinical trials is feasible in advanced breast cancer, and provides insight into the dynamics of cancer genome evolution through therapy.
|
10.1371/journal.pcbi.1004226 | Sparse and Compositionally Robust Inference of Microbial Ecological Networks | 16S ribosomal RNA (rRNA) gene and other environmental sequencing techniques provide snapshots of microbial communities, revealing phylogeny and the abundances of microbial populations across diverse ecosystems. While changes in microbial community structure are demonstrably associated with certain environmental conditions (from metabolic and immunological health in mammals to ecological stability in soils and oceans), identification of underlying mechanisms requires new statistical tools, as these datasets present several technical challenges. First, the abundances of microbial operational taxonomic units (OTUs) from amplicon-based datasets are compositional. Counts are normalized to the total number of counts in the sample. Thus, microbial abundances are not independent, and traditional statistical metrics (e.g., correlation) for the detection of OTU-OTU relationships can lead to spurious results. Secondly, microbial sequencing-based studies typically measure hundreds of OTUs on only tens to hundreds of samples; thus, inference of OTU-OTU association networks is severely under-powered, and additional information (or assumptions) are required for accurate inference. Here, we present SPIEC-EASI (SParse InversE Covariance Estimation for Ecological Association Inference), a statistical method for the inference of microbial ecological networks from amplicon sequencing datasets that addresses both of these issues. SPIEC-EASI combines data transformations developed for compositional data analysis with a graphical model inference framework that assumes the underlying ecological association network is sparse. To reconstruct the network, SPIEC-EASI relies on algorithms for sparse neighborhood and inverse covariance selection. To provide a synthetic benchmark in the absence of an experimentally validated gold-standard network, SPIEC-EASI is accompanied by a set of computational tools to generate OTU count data from a set of diverse underlying network topologies. SPIEC-EASI outperforms state-of-the-art methods to recover edges and network properties on synthetic data under a variety of scenarios. SPIEC-EASI also reproducibly predicts previously unknown microbial associations using data from the American Gut project.
| Genomic survey of microbes by 16S rRNA gene sequencing and metagenomics has inspired appreciation for the role of complex communities in diverse ecosystems. However, due to the unique properties of community composition data, standard data analysis tools are likely to produce statistical artifacts. For a typical experiment studying microbial ecosystems these artifacts can lead to erroneous conclusions about patterns of associations between microbial taxa. We developed a new procedure that seeks to infer ecological associations between microbial populations, by 1) taking advantage of the proportionality invariance of relative abundance data and 2) making assumptions about the underlying network structure when the number of taxa in the dataset is larger than the number of sampled communities. Additionally, we employed a novel tool to generate biologically plausible synthetic data and objectively benchmark current association inference tools. Finally, we tested our procedures on a large-scale 16S rRNA gene sequencing dataset sampled from the human gut.
| Low-cost metagenomic and amplicon-based sequencing promises to make the resolution of complex interactions between microbial populations and their surrounding environment a routine component of observational ecology and experimental biology. Indeed, large-scale data collection efforts (such as Earth Microbiome Project [1], the Human Microbiome Project [2], and the American Gut Project [3]) bring an ever-increasing number of samples from soil, marine and animal-associated microbiota to the public domain. Recent research efforts in ecology, statistics, and computational biology have been aimed at reliably inferring novel biological insights and testable hypotheses from population abundances and phylogenies. Classic objectives in community ecology include, (i) the accurate estimation of the number of taxa (observed and unobserved) from microbial studies [4] and, related to that, (ii) the estimation of community diversity within and across different habitats from the modeled population counts [5]. Moreover, some microbial compositions appear to form distinct clusters, leading to the concept of enterotypes, or ecological steady states in the gut [6], but their existence has not been established with certainty [7]. Another aim of recent studies is the elucidation of connections between microbes and environmental or host covariates. Examples include a novel statistical regression framework for relating microbiome compositions and covariates in the context of nutrient intake [8], observations that microbiome compositions strongly correlate with disease status in new-onset Crohn’s disease [9], and the connections between helminth infection and the microbiome diversity [10].
One goal of microbiome studies is the accurate inference of microbial ecological interactions from population-level data [11]. ‘Interactions’ are inferred by detecting significant (typically non-directional) associations between sampled populations, e.g., by measuring frequency of co-occurrence [12, 13]. Microbiota are measured by profiling variable regions of bacterial 16S rRNA gene sequences. These regions are amplified, sequenced, and the resulting reads are then grouped into common Operational Taxonomic Units (OTUs) and quantified, with OTU counts serving as a proxy to the underlying microbial populations’ abundances. Knowledge of interaction networks (here, a measure of microbial association) provides a foundation to predictively model the interplay between environment and microbial populations. A recent example is the successful construction of a dynamic differential equation model to describe the primary succession of intestinal microbiota in mice [14]. A commonly used tool to infer a network is correlation analysis; that is computing Pearson’s correlation coefficient among all pairs of OTU samples, and an interaction between microbes is assumed when the absolute value of the correlation coefficient is sufficiently high [15, 16].
However, applying traditional correlation analysis to amplicon surveys of microbial population data is likely to yield spurious results [9, 17]. To limit experimental biases due to sampling depth, OTU count data is typically transformed by normalizing each OTU count to the total sum of counts in the sample. Thus, communities of microbial relative abundances, termed compositions, are not independent, and classical correlation analysis may fail [18]. Recent methods such as Sparse Correlations for Compositional data (SparCC) [17] and Compositionally Corrected by REnormalization and PErmutation (CCREPE) [9, 11, 19] are designed to account for these compositional biases and represent the state of the art in the field. Yet, it is not clear that correlation is the proper measure of association. For example, correlations can arise between OTUs that are indirectly connected in an ecological network (we expand on this point below).
Dimensionality poses another challenge to statistical analysis of microbiome studies, as the number of measured OTUs p is on the order of hundreds to thousands whereas the number of samples n generally ranges from tens to hundreds. This implies that any meaningful interaction inference scheme must operate in the underdetermined data regime (p > n), which is viable only if additional assumptions about the interaction network can be made. As technological developments lead to greater sequencing depths, new computational methods that address the (p > n) challenge will become increasingly important.
In the present work, we present a novel strategy to infer networks from (potentially high-dimensional) community composition data. We introduce SPIEC-EASI (SParse InversE Covariance Estimation for Ecological ASsociation Inference, pronounced speakeasy), a new statistical method for the inference of microbial ecological networks and generation of realistic synthetic data. SPIEC-EASI inference comprises two steps: First, a transformation from the field of compositional data analysis is applied to the OTU data. Second, SPIEC-EASI estimates the interaction graph from the transformed data using one of two methods: (i) neighborhood selection [20, 21] and (ii) sparse inverse covariance selection [22, 23]. Unlike empirical correlation or covariance estimation, used in SparCC and CCREPE, our pipeline seeks to infer an underlying graphical model using the concept of conditional independence.
Informally, two nodes (e.g. OTUs) are conditionally independent if, given the state (e.g. abundance) of all other nodes in the network, neither node provides additional information about the state of the other. A link between any two nodes in the graphical model implies that the OTU abundances are not conditionally independent and that there is a (linear) relationship between them that cannot be better explained by an alternate network wiring. In this way, our method avoids detection of correlated but indirectly connected OTUs, thus ensuring parsimony of the resulting network model (for more detail, see Materials and Methods and Fig 1). This model is an undirected graph where links between nodes represent signed associations between OTUs. The use of graphical models has gained considerable popularity in network biology [24–26] and, more recently, in structural biology [27], particularly to correct for transitive correlations in protein structure prediction [28].
To properly benchmark our inference scheme and compare its performance with other state-of-the-art schemes [9, 17], SPIEC-EASI is accompanied by a synthetic data generation routine, which generates realistic synthetic OTU data from networks with diverse topologies. This is significant because, to date, (i) no experimentally verified set of “gold-standard” microbial interactions exists, (ii) previous synthetic benchmark data do not accurately reflect the actual properties of microbiome data [17], and (iii) theoretical and empirical work from high-dimensional statistics [29–31] suggests that network topology can strongly impact network recovery and performance and thus must be considered in the design of synthetic datasets.
We show that SPIEC-EASI is a scalable inference engine that (i) yields superior performance with respect to state-of-the-art methods in terms of interaction recovery and network features in a diverse set of realistic synthetic benchmark scenarios, (ii) provides the most stable and reproducible network when applied to real data, and (iii) reliably estimates an invertible covariance matrix which can be used for additional downstream statistical analysis. In agreement with statistical theory [29], inference on the synthetic datasets demonstrates that the degree distribution of the underlying network has the largest effect on performance, and this effect is observed across all methods tested. SPIEC-EASI network inference applied to actual data from the American Gut Project (AGP) shows (i) that clusters of strongly connected components are likely to contain OTUs with common family membership and (ii) that actual gut microbial networks are likely composites of archetypical network topologies. In the Materials and Methods section, we present statistical and computational aspects of SPIEC-EASI. We then benchmark SPIEC-EASI, comparing it to current inference schemes using synthetic data. We then apply SPIEC-EASI to measurements available from the AGP database. The SPIEC-EASI pipeline is implemented in the R package [SpiecEasi] freely available at https://github.com/zdk123/SpiecEasi. All presented numerical data is available at http://bonneaulab.bio.nyu.edu/.
SPIEC-EASI comprises both an inference and a synthetic data generation module. Fig 2 summarizes the key components of the pipeline. In this section, we introduce all statistical and computational aspects of the inference scheme and then describe our approach for generating realistic synthetic datasets.
For this discussion, a table of OTU count data, typical output of 16S rRNA gene sequencing data curation pipelines (e.g., mothur [32], QIIME [33]) are given. The OTU data are stored in a matrix W ∈ ℕ 0 n × p where w ( j ) = [ w 1 ( j ) , w 2 ( j ) , … , w p ( j ) ] denotes the p-dimensional row vector of OTU counts from the jth sample, j = 1, …, n, with total cumulative count m ( j ) = ∑ i = 1 p w i ( j ); ℕ0 denotes the set of natural numbers {0, 1, 2, …}. As described above, to account for sampling biases, microbiome data is typically transformed by normalizing the raw count data w(j) with respect to the total count m(j) of the sample [10]. We thus arrive at vectors of relative abundances or compositions x ( j ) = [ w 1 ( j ) m ( j ) , w 2 ( j ) m ( j ) , … , w p ( j ) m ( j ) ] for sample j. Due to this normalization OTU abundances are no longer independent, and the sample space of this p-part composition x(j) is not the unconstrained Euclidean space but the p-dimensional unit simplex 𝕊 p ≐ { x ∣ x i > 0 , ∑ i = 1 p x i = 1 }. Thus, OTU compositions from n samples are constrained to lie in the unit simplex, X ∈ 𝕊n×p. This restriction of the data to the simplex prohibits the application of standard statistical analysis techniques, such as linear regression or empirical covariance estimation. Covariance matrices of compositional data exhibit, for instance, a negative bias due to closure effects.
Major advances in the statistical analysis of compositional data were achieved by Aitchison in the 1980’s [18, 34]. Rather than considering compositions in the simplex, Aitchison proposed log-ratios, log [ x i x j ], as a basis for studying compositional data. The simple equivalence log [ x i x j ] = log [ w i / m w j / m ] = log [ w i w j ] implies that statistical inferences drawn from analysis of log-ratios of compositions are equivalent to those that could be drawn from the log-ratios of the unobserved absolute measurements, also termed the basis.
Aitchison also proposed several statistically equivalent log-ratio transformations to remove the unit-sum constraint of compositional data [18]. Here we apply the centered log-ratio (clr) transform:
z≐clr(x)=[log(x1/g(x)),…,log(xp/g(x)]=[log(w1/g(w)),…,log(wp/g(w))] (1)
where g(x)=[∏pi=1xi]1/p is the geometric mean of the composition vector. The clr transform is symmetric and isometric with respect to the component parts. The resulting vector z is constrained to a zero sum. The clr transform maps the data from the unit simplex to a p − 1-dimensional Euclidean space, and the corresponding population covariance matrix Γ = Cov[clr(X)] ∈ ℝp×p is also singular [18]. The covariance matrix Γ is related to the population covariance of the log-transformed absolute abundances Ω = Cov[logW] via the relationship [34]:
Γ = G Ω G (2)
where G = I p − 1 p J, Ip is the p-dimensional identity matrix, and J = [j1, j2, …, ji, …, jp], ji = [1, 1, …, 1] the p-dimensional all-ones vector. For high-dimensional data, p > > 0, the matrix G is close to the identity matrix, and thus we can assume that a finite sample estimator Γ ^ of Γ serves as a good approximation of Ω ^. This approximation serves as the basis of our network inference scheme. Finally, because real-world OTU data often contain samples with a zero count for low-abundance OTUs, we add a unit pseudo count to the original count data to avoid numerical problems with the clr transform.
Our key objective is to learn a network of pairwise taxon-taxon associations (putative interactions) from clr-transformed microbiome compositions Z ∈ ℝn×p. We represent the network as an undirected, weighted graph 𝓖 = (V, E), where the vertex set V = {v1, …, vp} represents the p taxa (e.g., OTUs) and the edge set E ⊂ V × V the possible associations among them. Our formal approach is to learn a probabilistic graphical model [35] (i) that is consistent with the observed data and (ii) for which the (unknown) graph 𝓖 encodes the conditional dependence structure between the random variables (in our case, the observed taxa). Graphical models over undirected graphs (also known as Markov networks or Markov Random Fields) have a straightforward distributional interpretation when the data are drawn from a probability distribution π(x) that belongs to an exponential family [36, 37]. For example, when the data are drawn from a multivariate normal distribution π(x) = 𝓝(x∣μ, Σ) with mean μ and covariance Σ, the non-zero elements of the off-diagonal entries of the inverse covariance matrix Θ = Σ−1, also termed the precision matrix, defines the adjacency matrix of the graph 𝓖 and thus describes the factorization of the normal distribution into conditionally dependent components [35]. Conversely, if and only if an entry in Θ: Θi, j = 0, then the two variables are conditionally independent, and there is no edge between vi and vj in 𝓖. We seek to estimate the inverse covariance matrix from the data, thereby inferring associations based on conditional independence. This is fundamentally distinct from SparCC and CCREPE (see S1 Table), which essentially estimate pairwise correlations (though other pairwise metrics could be considered for CCREPE). We highlight this key difference in Fig 1. For an intuitive introduction to graphical models in the context of biological networks see Bühlmann et. al, 2014 [38].
Inferring the exact underlying graph structure in the presence of a finite amount of samples is, in general, intractable. However, two types of statistical inference procedures have been useful in high-dimensional statistics due to their provable performance guarantees under assumptions about the sample size n, dimensionality p, underlying graph properties, and the generating distribution [29, 39]. The first approach, neighborhood selection [20, 39], aims at reconstructing the graph on a node-by-node basis where, for each node, a penalized regression problem is solved. The second approach is the penalized maximum likelihood method [22, 23], where the entire graph is reconstructed by solving a global optimization problem, the so-called covariance selection problem [40]. The key advantages of these approaches are that (i) their underlying inference procedures can be formulated as convex (and hence tractable) optimization problems, and (ii) they are applicable even in the underdetermined regime (p > n), provided that certain structural assumptions about the underlying graph hold. One assumption is that the true underlying graph is reasonably sparse, e.g., that the number of taxon-taxon associations scales linearly with the number of measured taxa.
Graphical model inference. The SPIEC-EASI pipeline comprises two types of inference schemes, neighborhood and covariance selection. The neighborhood selection framework, introduced by Meinshausen and Bühlmann [20] and thus often referred as the MB method, tackles graph inference by solving p regularized linear regression problems, leading to local conditional independence structure predictions for each node. Let us denote the ith column of the data matrix Z by Zi ∈ ℝn and the remaining columns by Z¬i ∈ ℝn×p−1. For each node vi, we solve the following convex problem:
β ^ i , λ = arg min β ∈ ℝ p - 1 ( 1 n ∥ Z i - Z ¬ i β ∥ 2 + λ ∥ β ∥ 1 ) , (3)
where ‖ a ‖ 1 = ∑ i = 1 p − 1 ∣ a i ∣ denotes the L1 norm, and λ ≥ 0 is a scalar tuning parameter. This so-called LASSO problem [41] aims at balancing the least-square fit and the number of necessary predictors (the non-zero components βj of β) by tuning the λ parameter. We define the local neighborhood of a node vi as N i λ = { j ⊂ { 1 , … p } \ i : β ^ j i , λ ≠ 0 }. The final edge set E of 𝓖 can be defined via the intersection or the union operation of the local neighborhoods. An edge ei, j between node vi and vj exists if j ∈ N i λ ∩ i ∈ N j λ or j ∈ N i λ ∪ i ∈ N j λ. For edges in the set j ∈ N i λ ∩ i ∈ N j λ, the edge weights, ei, j and ej, i, are estimated using the average of the two corresponding β entries. From a theoretical point of view, both edge selection choices are asymptotically consistent under certain technical assumptions [20]. The choice of the λ parameter controls the sparsity of the local neighborhood, which requires tuning [42]. We present our parameter selection strategy at the end of this section.
The second inference approach, (inverse) covariance selection, relies on the following penalized maximum likelihood approach. In the standard Gaussian setting, the related convex optimization problem reads:
Θ ^ = arg min Θ ∈ PD ( - log det ( Θ ) + tr ( Θ Σ ^ ) + λ ∥ Θ ∥ 1 ) , (4)
where PD denotes the set of symmetric positive definite matrices {A: xT Ax > 0, ∀x ∈ ℝp}, Σ ^ the empirical covariance estimate, ‖⋅‖1 the element-wise L1 norm, and λ ≥ 0 a scalar tuning parameter. For λ = 0, the expression is identical to the maximum likelihood estimate of a normal distribution 𝓝(x∣0, Σ). For non-zero λ, the objective function (also referred as the graphical Lasso [22]) encourages sparsity of the underlying precision matrix Θ. The non-zero, off-diagonal entries in Θ define the adjacency matrix of the interaction graph 𝓖 which, similar to MB, depends on the proper choice of the penalty parameter λ. Originally, this estimator was shown to have theoretical guarantees on consistency and recovery only under normality assumptions [43]. However, recent theoretical [29, 44] work shows that distributional assumptions can be considerably relaxed, and the estimator is applicable to a larger class of problems, including inference on discrete (count) data. In addition, nonparametric approaches, such as sparse additive models, can be used to “gaussianize” the data prior to network inference [45]. We thus propose the following estimator for inferring microbial ecological associations. Given clr-transformed OTU data Z ∈ ℝn×p, we propose the modified optimization problem:
Ω ^ - 1 = arg min Ω - 1 ∈ P D ( - log det ( Ω - 1 ) + tr ( Ω - 1 Γ ^ ) + λ ∥ Ω - 1 ∥ 1 ) , (5)
where Γ ^ is the empirical covariance estimate of Z, and Ω−1 is the inverse covariance (or precision matrix) of the underlying (unknown) basis. As stated above, Γ ^ will be a good approximation for the basis covariance matrix Ω ^ because p > > 0. The resulting solution is constrained to the set of PD matrices, ensuring that the penalized estimator has full rank p. The non-zero off-diagonal entries of the estimated matrix Ω−1 define the inferred network 𝓖, and their values are the signed edge weights of the graph. To reduce the variance of the estimate, the covariance matrix Γ ^ can also be replaced by the empirical correlation matrix R ^ = D Γ ^ D, where D is a diagonal matrix that contains the inverse of the estimated element-wise standard deviations.
The covariance selection approach has two advantages over the neighborhood selection framework. First, we obtain unique weights associated with each edge in the network. No averaging or subsequent edge selection is necessary. Second, the covariance selection framework provides invertible precision and covariance matrix estimates that can be used in further downstream microbiome analysis tasks, such as regression and discriminant analysis [10].
Model selection. For both neighborhood and covariance selection, the tuning parameter λ ∈ [0, λmax] controls the sparsity of the final model. Rather than inferring a single graphical model, both methods produce a λ-dependent solution path with the complete and the empty graph as extreme networks. A number of model selection criteria, such as Bayesian Information Criteria [46] and resampling schemes [47], have been used. Here we use a popular model selection scheme known as the Stability Approach to Regularization Selection (StARS) [48]. This method repeatedly takes random subsamples (80% in the standard setting) of the data and estimates the entire graph solution path based on this subsample. For each subsample, the λ-dependent incidence frequencies of individual edges are retained, and a measure of overall edge stability is calculated. StARS selects the λ value at which subsampled non-empty graphs are the least variable (most stable) in terms of edge incidences. For the selected graph, the observed edge frequencies indicate the reproducibility, and likely the predictive power, and are used to rank edges according to confidence.
Theoretical and computational aspects. Learning microbial graphical models with neighborhood or inverse covariance selection schemes has important theoretical and practical advantages over current methods. A wealth of theoretical results are available that characterize conditions for asymptotic and finite sample guarantees for the estimated networks [20, 29, 39, 43, 46]. Under certain model assumptions, the number of samples n necessary to infer the true topology of the graph in the neighborhood selection framework is known to scale as n = O(d3 log(p)), where d is the maximum vertex (or node) degree of the underlying graph (i.e. the maximum size of any local neighborhood). Additional assumptions on the sample covariance matrices reduce the scaling to n = O(d2 log(p)) [39]. This implies that graph recovery and precision matrix estimation is indeed possible even in the p > > n regime, and that the underlying graph topology strongly impacts edge recovery. The latter observation means that, even if the number of interactions e is constant, graphs with large hub nodes, perhaps representing keystone species in microbial networks, or, more generally, scale-free graphs with, a few highly connected nodes, will be more difficult to recover than networks with evenly distributed neighborhoods. In addition to these theoretical results, a second advantage is that well-established, efficient, and scalable implementations are available to infer microbial ecological networks from OTU data in practice. Thus, SPIEC-EASI methods will efficiently scale as microbiome dataset dimensions grow (e.g., due to technological advances that increase the number of OTUs detected per sample). The SPIEC-EASI inference engine relies on the R package huge [49], which includes algorithms to solve neighborhood and covariance selection problems [20, 22], as well as the StARS model selection.
Estimating the absolute and comparative performance of network inference schemes from biological data remains a fundamental challenge in biology. In the context of gene regulatory network inference, recent community-wide efforts, such as the DREAM (Dialogue for Reverse Engineering Assessments and Methods) Challenges (http://www.the-dream-project.org/), have considerably advanced our understanding about feasibility, accuracy, and applicability of a large number of developed methods. In the DREAM challenges, both real data from “gold standard” regulatory networks (e.g., networks where the true topology is known from independent experimental evidence) and realistic in-silico data (using, e.g., the GeneNetWeaver pipeline [50]) are included. In the context of microbiome data and microbial ecological networks, neither a gold standard nor a realistic synthetic data generator exist. SPIEC-EASI is accompanied by a set of computational tools that allow the generation of realistic synthetic OTU data. As outlined in Fig 2, real taxa count data serve as input to SPIEC-EASI’s synthetic data generation pipeline. The pipeline enables one to: (i) fit the marginal distributions of the count data to a parametric statistical model and (ii) specify the underlying graphical model architecture (e.g., scale-free).
The NorTA approach. The parametric statistical model and network topology are then combined in the ‘Normal To Anything’ (NorTA) [51] approach to generate synthetic OTU data that resemble real measurements from microbial communities but with user-specified network topologies. NorTA [51] is an approximate technique to generate arbitrary continuous and discrete multivariate distributions, given (1) a target correlation structure R with entries ρi, j and (2) a target univariate marginal distribution Ui. To achieve this task, NorTA relies on normal-copula functions [51–53]. A n × p matrix of data U is sampled from a normal distribution with zero mean and a p × p correlation matrix RN. For each marginal Ui, the Normal cumulative distribution function (CDF) is transformed to the target distribution via its inverse CDF. For any target distribution P with CDF Ξ, we can thus generate multivariate correlated data via
U P i = Ξ - 1 ( Φ ( U N i ) ) , (6)
where UN ∼ 𝓝(0, RN) and Φ ( U ) = ∫ − ∞ U 1 2 σ 2 e − u 2 2 d u, the CDF of a univariate normal. In Fig 3a, we illustrate this process for bivariate Poisson and negative binomial data (n = 1000 and correlation ρij = 0.7).
In the original NorTA approach, an element-wise monotone transformation cU(⋅) with RN = cU(R) is applied to account for slight differences in correlation structure between normal and target distribution samples [51]. Here we neglect this transformation step because we observe that the log-transformed data from exponential count distributions, such as the Poisson and Negative binomial, are already close to R, provided that the mean is greater than one, particularly when the counts data are log-transformed (S1 Fig). In practice, SPIEC-EASI relies on routines from base R and VGAM packages [54, 55] to estimate the uniform quantiles of the normal data and to fit the desired CDF with estimated parameters.
Fitting marginal distribution to real OTU data. Prior to fitting marginal distributions to real data, several commonly used pre-proccesing steps are applied. For any given OTU abundance table of size n × K, we first select p non-zero columns. To account for experimental differences in sample sequencing, we then normalize samples to a median sequencing depth by multiplying all counts by the ratio of minimum desirable sampling depth to the total sum of counts in that sample and rounding to the nearest whole number, which is preferable to rarefaction [56]. These filtered and sequencing-depth-normalized data serve as the marginal counts, which are fit to a parametric distribution Ui and used as input to the NorTA approach. The concrete target marginal distribution depends on the actual microbiome dataset. For gut microbiome data (e.g. from HMP or APG), the zero-inflated Negative Binomial (ziNB) distribution is a good choice, as it accounts for both overdispersion [56, 57] and the preponderance of zero-count data points in microbial count datasets. The fitting procedure is done within a maximum likelihood framework. The corresponding optimization problem is solved with the Quasi-Newton methods with box-constraints, as implemented in the optim function in R [54]. In S1 Text, we use quantile-quantile plots to compare ziNB to several other candidate distributions (e.g., lognormal, Poisson, NB) and show that ziNB has superior fit.
Generation of network topologies and correlation matrices. Under normality assumption, the non-zero pattern of the precision matrix corresponds to the adjacency matrix of the underlying undirected graph. We use this property to generate target covariance (correlation) matrices originating from different graph topologies. The pipeline to generate a network structure for simulated data proceeds in three steps: (i) Generate an undirected graph, in the form of an adjacency matrix, with a desired topology and sparsity, (ii) convert the adjacency matrix to a positive-definite precision matrix by assigning positive and negative edge weights and appropriate diagonal entries, and (iii) invert Θ and convert the resulting covariance matrix Σ to a correlation matrix (R = DΣD, where, D is a diagonal matrix with diagonal entries 1 / σ i).
Among many potential graph structures, we focus on three representative network structures: band-like, cluster, and scale-free graphs (see Fig 3b for graphical examples). Maximum network degree strongly impacts network recovery, and thus our choice of network topologies spans a range of maximum degrees (band < cluster < scale-free). In addition, cluster and scale-free lend themselves to hypothetical ecological scenarios. Cluster graphs may be seen as archetypal models for microbial communities that populate different disjoint niches (clusters) and have only few associations across niches. Scale-free graphs, ubiquitous in many other facets of network biology (such as gene regulatory, protein-protein and social networks), serve as a baseline model for a microbial community that comprises (1) a few “keystone” species (hub nodes with many partners) that are essential for coordinating/stabilizing the community and (2) many dependent species that are sparsely connected to each other. The sparsity of the networks is controlled by the number of edges, e < p(p − 1)/2, in the graph. The topologies are generated according to the following algorithms, starting with an empty p × p adjacency matrix:
Band: A band-type network consists of a chain of nodes that connect only their nearest neighbors. Let e = eused + eavailable, the number of edges already used and available, respectively. Fill the next available off-diagonal vector with edges if and only if eavailable ≥ number of elements in the off-diagonal.
Cluster: A cluster network comprises h independent groups of randomly connected nodes. For given p and e we divide the set of nodes into h components of (approximately) identical size and set the number of edges in each component to ecomp = e/h. For each component, we generate a random (Erdös-Renyi) graph for which we randomly assign an edge between two nodes in the cluster with probability p = e comp h ( h − 1 ) / 2 ⋅
Scale-free: The distribution of degrees, the number of edges per node, in a scale-free network is described by a power law, implying that the central node or nodes (potentially keystone species in an ecological network) have proportionally more connections. We use the standard preferential attachment scheme [58] until p − 1 edges are exhausted.
After generation of these standard adjacency matrices, we randomly remove or add edges until the adjacency matrix has exactly e edges. All schemes generate symmetric adjacency matrices that describes a graph, with entries of 1 if an edge exists and 0 otherwise.
From the adjacency matrices, we generate precision matrices by uniformly sampling non-zero entries Θi, j ∈ [−Θmax, −Θmin]∪[Θmin, Θmax], where Θmin, Θmax > 0 are model parameters and describe the strength of the conditional dependence among the nodes. To ensure that the precision matrix is positive definite with tunable condition number κ = cond(Θ), we scale the diagonal entries Θi, i by a constant c using binary search. The precision matrix Θ is then converted to a correlation matrix R to be used as input to the NorTA approach.
Given that no large-scale experimentally validated microbial ecological network exists, we use SPIEC-EASI’s data generator capabilities to synthesize data whose OTU count distributions faithfully resemble microbiome count data. By varying parameters known to influence network recovery (network topology, association strength, sample number) and quantifying performance on resulting networks, we rigorously assess SPIEC-EASI inference relative to state-of-the-art inference methods, SparCC [17] and CCREPE [9], as well as standard Pearson correlation.
Thus far we have used the n1 = 304 first-round AGP samples as a means to construct realistic synthetic microbiome data sets with SPIEC-EASI’s data generation module. In this section, we apply SPIEC-EASI inference methods to construct ecological association networks from the AGP data directly. To do this, we first filter out rare OTUs by selecting only the top 205 OTUs (to match the dimensionality of the synthetic data) in the combined AGP dataset (by frequency of presence) before adding a pseudo-count and total-sum normalization. Although there is no independent means to assess the accuracy of these hypothetical networks, we can assess their reproducibility and consistency. For each method, we first infer a single representative network of taxon-taxon interactions from Round 1 AGP abundance data. For SPIEC-EASI, the StARS model selection approach is used to select the final model network. For SparCC, we use a threshold ρt = 0.35 to construct a relevance network from the SparCC-inferred correlation matrix; i.e. an edge between nodes vi, vj is present in the SparCC network if ∣ρi, j∣ > ρt [17]. Similarly, we use a q-value cut-off of 10−24 to create an interaction network from CCREPE-corrected significance scores of Pearson’s correlation coefficient [9]. For each method, we thus arrive at a reference network that can be considered the hypothetical gold standard. We then use the n2 = 254 Round 2 AGP samples as an independent test set and learn a new model network from these data alone. We measure consistency between the two network models by computing the Hamming distance between the reference and new network models, i.e., the difference between the upper triangular part of the two adjacency matrices. For the present data, the Hamming distance can vary between p(p − 1)/2 = 20910 (no edges in common) and a minimum of 0 for identical networks. Confidence intervals for Hamming distances can be obtained by combining Round 1 and 2 samples into a unified dataset, repeatedly subsampling these data into two disjoint groups of size n1 and n2, and repeating the entire inference procedure.
Fig 6a shows network reproducibility for SPIEC-EASI methods, SparCC, and CCREPE. The S-E(MB) has smallest the Hamming distance, followed by S-E(glasso), SparCC, and CCREPE. In S-E(MB), the edge disagreement is roughly 50 with very small error bars. At the other extreme, CCREPE edge disagreement is 250 edges and highly variable.
These numerical experiments clearly demonstrate that SPIEC-EASI networks are more reproducible than other current methods.
Finally, we use each inference method to construct a candidate American Gut microbiome association network from the unified dataset of size n1 + n2 = 558 (Fig 6c). We analyze the differences between the reconstructed networks by quantifying the number of unique and shared predicted edges (Fig 6b). All four methods agree on a core network that consists of 127 edges. These edges are mostly found within OTUs of the same taxonomic group. This phenomenon, termed assortativity, has also been observed in other microbial network studies [11]. Assortativity is one of the most salient features of the AGP-derived networks, and, for all networks, the assortativity coefficients for each network are close to unity (e.g., maximum assortativity, S11 Fig). The SparCC network comprises about twice as many edges as the SPIEC-EASI networks. SparCC infers 147 distinct edges; these additional edges correspond to negative associations between OTUs of Ruminococcaceae (genus Faecalibacterium) and Enterobacteriacae families (various genera) and a dense web of correlations within Enterobacteriacae OTUs. Similarly, CCREPE identified 152 edges uniquely, with many negative edges between Enterobacteriaceae and Lachnospiraceae (genera: Blautia, Roseburia and unknown); additionally, CCREPE uniquely predicted positive edges between the Lachnospiraceae and Ruminococcaceae (genus: Faecalibacterium). Both SPIEC-EASI methods produce relatively sparse networks by comparison. S-E(glasso) infers a total of 271 total edges (with one unique edge), and S-E(MB) infers 206 edges with 25 unique edges. In scale with edge predictions, both CCREPE and SparCC infer networks with large maximum degree (33 and 30, respectively), while the S-E(MB) and S-E(glasso) networks have a maximum degree of sixteen and eight, respectively (S11 Fig). However, even though CCREPE and SparCC predict a similar number of total edges, the global network properties are distinct. CCREPE predicts a higher maximum betweenness centrality and a larger number of nodes in the largest connected component (100).
In summary, analysis of the AGP networks suggests that the SPIEC-EASI inference schemes construct more reproducible taxon-taxon interactions than SparCC and CCREPE and infer considerably sparser model networks than the other two methods. These observations may be explained as follows: SparCC and CCREPE aim to recover correlation networks, which contain both direct edges as well as indirect (e.g., spurious) edges (due to correlation alone). SparCC and CCREPE may recover indirect edges less robustly than direct edges, an explanation that would be consistent with the Hamming distance reproducibility analysis. In addition, all methods’ resulting networks suggest that the topology of the American Gut association network cannot be attributed to a specific network class. Instead, these networks are a mixture of band, scale-free, and cluster network type, and they exhibit high phylogenetic assortativity within highly connected components.
Inferring interactions among different microbial species within a community and understanding their influence on the environment is of central importance in ecology and medicine [19, 60]. An ever increasing number of recent amplicon-based sequencing studies have uncovered strong correlations between microbial community composition and environment in diverse and highly relevant domains of life [1, 9, 10, 61–63]. These studies alone underscore the need to understand how the microbial communities adapt, develop, and interact with the environment [5]. Elucidation of species interactions in microbial communities across different environments remains, however, a formidable challenge. Foremost, available high-throughput experimental data are compositional in nature, overdispersed, and usually underdetermined with respect to statistical inference. In addition, for most microbes few to no ecological interactions are known, thus the ecological interaction network must be constructed de novo, in the absence of guiding assumptions and a set of “gold standard” interactions for validation.
To overcome both challenges, we present SPIEC-EASI (Sparse InversE Covariance Estimation for Ecological Association Inference), a computational framework that includes statistical methods for the inference of microbial ecological interactions from 16S rRNA gene sequencing datasets and a sophisticated synthetic microbiome data generator with controllable underlying species interaction topology. SPIEC-EASI’s inference engine includes two well-known graphical model estimators, neighborhood selection [20] and sparse inverse covariance selection [22, 23, 46] that are extended by compositionally robust data transformations for application to the specific context of microbial abundance data.
The synthetic data pipeline was used to generate realistic-looking gut microbiome datasets for a controlled benchmark of SPIEC-EASI’s inference performance relative to two state-of-the-art methods, SparCC [17] and CCREPE [9]. We showed that neighborhood selection (S-E(MB)) outperforms SparCC and CCREPE in terms of recovery of taxon-taxon interactions and global network topology features under almost all tested benchmark scenarios, while covariance selection (S-E(glasso)) performs competitively with and sometimes better than SparCC and CCREPE.
Through our simulation study, we demonstrate that several other factors, in addition to total number of samples, affect network recovery. Foremost and in agreement with theoretical results from high-dimensional statistics [29, 30, 39], network topology has a significant impact, as network recovery performance is nearly doubled from scale-free to cluster to band (Fig 4) for fixed sample size, number of taxa, and condition number. We also demonstrated dependence to strength of direct interactions (and thus strength of correlations) within a given network. Our simulation study provides the community with rough guidelines for requisite sample sizes, given state-of-the-art network inference and basic assumptions about the underlying network. This is of obvious importance to experimental design and the estimation of statistical power. Here, we used the synthetic data pipeline to generate datasets characteristic of the gut microbiome. However, the SPIEC-EASI data generator is generic and therefore enables researchers to generate synthetic datasets that resemble microbiome samples in terms of taxa dispersion and marginal distributions from their field of research, such as soil or sea water ecosystems [1].
Our application study on real American Gut Project (AGP) data revealed that inference with SPIEC-EASI produced more consistent and sparser interaction networks than SparCC and CCREPE. In addition, our AGP network analysis revealed several biologically relevant observations. Specifically, we observed that OTUs were more likely to interact with phylogenetically related OTUs (Fig 6c and S11 Fig). In addition, our gut microbial interaction networks appear to be a composite of network types, as we find evidence for scale-free, band-like, and cluster subnetworks.
An important advantage of neighborhood and covariance selection as underlying inference frameworks is their ability to include prior knowledge about the underlying data or network structure from independent scientific studies in a principled manner. For example, in the neighborhood selection scheme, the standard LASSO approach can be augmented by a group penalty [64] that takes into account a priori known group structure. The assortativity observed in our gut microbial interaction networks suggests that such a grouping of OTUs based on phylogenetic relationship might improve inference. Moreover, if verified species interactions are available for a certain microbial contexts, this knowledge can be included in covariance and neighborhood selection by relaxing the penalty term on these interactions. This strategy has already been fruitfully applied to inference of similarly high-dimensional transcriptional regulatory networks [65]. Finally, in agreement with theoretical and empirical work in high-dimensional statistics, our synthetic benchmark results confirmed that networks with scale-free structures elude accurate inference even if the underlying network is globally sparse. Recent modified neighborhood [30] and covariance selection [31] schemes improve recovery of scale-free networks and can be conveniently included into SPIEC-EASI.
Finally, although the main focus of this work is inference of microbial interaction networks, estimation of the regularized inverse covariance matrix with S-E(glasso) will be key to addressing several other important questions arising from microbiome studies. For example, statistical methods to infer which taxa are responsive to design factors in 16S gene amplicon studies is an active area of research. Most methods test each taxon independently one-at-a-time (see [56] and references therein) even though taxa are actually highly correlated and thought to ecologically interact. Inference of taxa responses from 16S rRNA gene sequencing datasets could be improved by modeling this correlation structure through incorporation of the inverse covariance matrix into the statistical model [66].
Other, more complex questions are motivated by a desire to understand why and how ecosystems evolve with time. In the dynamic modeling setting, association networks have already been successfully used as an underlying structure to fit a differential-equation-based model of gut microbiome development in mice [14]. Thus, association networks provide the underlying topology for dynamic models, which can be used to develop hypotheses about how the ecosystems might respond to specific perturbations [5].
In conclusion, SPIEC-EASI is an improvement over state-of-the-art methods for inference of microbial ecological networks from microbiome composition datasets. We demonstrate this through rigorous benchmarking with synthetic networks and also through application to a true biological dataset. In addition, the LASSO underpinnings of the SPIEC-EASI inference methods provide a flexible and principled mathematical framework to incorporate additional information about microbial ecological association networks as it becomes available, thereby improving prediction. We anticipate that SPIEC-EASI network inference will serve as a backbone for more sophisticated modeling endeavors, engendering new hypotheses and predictions of relevance to environmental ecology and medicine.
|
10.1371/journal.pbio.1001290 | Mediator Acts Upstream of the Transcriptional Activator Gal4 | The proteasome inhibitor MG132 had been shown to prevent galactose induction of the S. cerevisiae GAL1 gene, demonstrating that ubiquitin proteasome-dependent degradation of transcription factors plays an important role in the regulation of gene expression. The deletion of the gene encoding the F-box protein Mdm30 had been reported to stabilize the transcriptional activator Gal4 under inducing conditions and to lead to defects in galactose utilization, suggesting that recycling of Gal4 is required for its function. Subsequently, however, it was argued that Gal4 remains stably bound to the enhancer under inducing conditions, suggesting that proteolytic turnover of Gal4 might not be required for its function. We have performed an alanine-scanning mutagenesis of ubiquitin and isolated a galactose utilization-defective ubiquitin mutant. We have used it for an unbiased suppressor screen and identified the inhibitor Gal80 as a suppressor of the transcriptional defects of the ubiquitin mutant, indicating that the protein degradation of the inhibitor Gal80, and not of the activator Gal4, is required for galactose induction of the GAL genes. We also show that in the absence of Gal80, Mdm30 is not required for Gal4 function, strongly supporting this hypothesis. Furthermore, we have found that Mediator controls the galactose-induced protein degradation of Gal80, which places Mediator genetically upstream of the activator Gal4. Mediator had originally been isolated by its ability to respond to transcriptional activators, and here we have discovered a leading role for Mediator in the process of transcription. The protein kinase Snf1 senses the inducing conditions and transduces the signal to Mediator, which initiates the degradation of the inhibitor Gal80 with the help of the E3 ubiquitin ligase SCFMdm30. The ability of Mediator to control the protein degradation of transcriptional inhibitors indicates that Mediator is actually able to direct its own recruitment to gene promoters.
| The expression levels of proteins are tightly regulated, not only via their production but also via their degradation. Genes are transcribed only if their encoded proteins are required by the environmental or developmental conditions of a cell, and once a certain protein is no longer needed, it is rapidly degraded by the ubiquitin proteasome system (UPS). Transcriptional activators appeared to contradict this simple economic principle, as it had been claimed that they had to be degraded in order to function. The claim was based upon a correlation: if the degradation of an activator was prevented by drugs or mutations in the UPS, the activator became stable but also nonfunctional. We have now shown that it is not the activator itself but its inhibitor that is the functionally relevant target of the UPS. Furthermore, we have found that the degradation of the inhibitor is controlled by a protein complex called Mediator. The activator is known to recruit Mediator to gene promoters, where Mediator assists RNA polymerase in initiating transcription. Mediator was always considered to be completely under the control of the activator; however, we observe that by regulating the degradation of the inhibitor, Mediator is also able to control the activator and thereby to orchestrate its own recruitment to gene promoters.
| Cells regulate the expression of their genes according to requirement [1]. Activators recruit chromatin-remodeling or chromatin-modifying complexes that change the structure of chromatin to promote transcription [2],[3], while repressors recruit chromatin-modifying complexes that change the structure of chromatin to prevent transcription [4],[5]. Repressors also bind directly to activators and prevent the recruitment of the transcription machinery [6]. According to the reverse recruitment hypothesis [7], the transcription factors do not move to the highly transcribed genes, but the highly transcribed genes move to the gene expression machines (GEMs), which are protein complexes with fixed locations in the nuclear periphery. GEMs, which host all transcription factors that are required for gene expression from RNA Polymerase to RNA capping, splicing, poly-adenylation, and export factors [8], are associated with the nuclear pores, and the mature mRNAs, once produced at the GEM, are immediately exported out of the nucleus to be translated at the ribosomes of the rough endoplasmic reticulum [7].
The Saccharomyces cerevisiae GAL genes are a paradigm for transcriptional regulation in eukaryotes [9]. In cells grown with glucose, Gal80 binds to Gal4 and blocks its activation function [10], while Mig1 binds to an upstream silencer and recruits the general repressor Tup1 to prevent gene expression [11]. Upon the switch to galactose media, Snf1 phosphorylates Mig1, causing its translocation from the nucleus to the cytoplasm [12], while Gal80 dissociates from Gal4 [13] and is sequestered in the cytoplasm by Gal3 [14], leaving Gal4 free to activate the GAL genes, which are required for galactose utilization [7].
Proteolytic stability of transcription factors offers an intriguing possibility for the eukaryotic cell to control gene expression [15]. Ubiquitin proteasome-dependent degradation (UPD) of activators and repressors plays an important role in gene regulation [16], and treatment of S. cerevisiae cells with the proteasome inhibitor MG132 abolished galactose induction of the GAL1 gene [17]. Ubiquitin is a small protein of 76 amino acids that is transferred by E3 ubiquitin ligases to proteins to be targeted for degradation by the 26S proteasome [18]. F-box proteins confer substrate specificity to SCF (Skip1-Cullin-F-box protein) E3 ubiquitin ligases [19]. When cells are grown with galactose, an SCF E3 ubiquitin ligase containing the F-box protein Mdm30, SCFMdm30, ubiquitinates Gal4 [20]. The deletion of MDM30 stabilizes Gal4 under inducing conditions and leads to defects in galactose utilization, suggesting that recycling of Gal4 is required for its transcriptional activator function [20]. Subsequently, however, it was argued that Gal4 remains stably bound to the enhancer under inducing conditions, suggesting that proteolytic turnover of Gal4 might not be required for its function [21]–[23]. Previously, it had been shown that mono-ubiquitination protected Gal4 from the promoter-stripping activity of proteasomal ATPases [24]–[26], suggesting a role for ubiquitin in transcriptional activation other than protein degradation. Recently, it has been reported that the proteolytic stability of Mediator subunits is inversely correlated with their ability to activate transcription when fused to a DNA-binding domain [27].
Mediator is a complex of more than 20 proteins that is conserved from yeast to man [28]. It was discovered by its ability to respond to transcriptional activators in vivo and in vitro [29]. Genome-wide gene expression studies with temperature-sensitive alleles have shown that Mediator is required for the transcription of nearly all RNA Polymerase II–dependent genes in yeast [30]. Mediator interacts directly with activators, General Transcription Factors, and RNA Polymerase II [31]. In higher eukaryotes, Mediator facilitates a DNA loop between enhancer and basal promoter via its interaction with cohesin [32]. In addition, Mediator affects steps that are downstream of the recruitment of RNA Polymerase II to the core promoter, as Med26-containing metazoan Mediator switches RNA Polymerase into the productive transcription elongation mode by an interaction of Med26 with TBP (TATA-binding protein) and the CTD (C-terminal domain of RNA Polymerase II) kinase P-TEFb [33]. Mediator also modifies chromatin via its own CDK8 subunit, which phosphorylates histone H3S10, and by its interaction with histone acetyl- and methyltransferases [34],[35]. Metazoan Mediator plays important roles in neurogenesis, cancer formation, and stem cell proliferation [31]. All of these reported functions of Mediator are genetically downstream of transcriptional activators. Here, we have found that Mediator additionally is able to act upstream of the transcriptional activator Gal4 by controlling the ubiquitin-mediated protein degradation of the inhibitor Gal80. In the absence of Gal80, Gal4 is free to recruit Mediator to the promoter of the GAL genes. Therefore, Mediator actually orchestrates its own recruitment to the GAL promoters upon galactose induction.
The role of ubiquitin proteasome-dependent protein degradation in the transcriptional regulation of the GAL genes has been controversial [19]–[22]. We performed an alanine-scanning mutagenesis of ubiquitin in order to isolate galactose-utilization defective (gal−) mutant strains and use these for unbiased multi-copy suppressor screens. However, no ubiquitin single point mutant displaying the gal− phenotype was isolated (Figure S1, even lanes; Figure S2). The addition of an N-terminal tag can sometimes enhance the phenotype of point mutants, and so we fused a stretch of 10 N-terminal histidines to all ubiquitin mutant proteins. S. cerevisiae cells expressing H10UbF4A, H10UbK6A, H10UbI13A, H10UbR42A, H10UbF45A, H10UbD58A, and H10UbT66A in the place of endogenous ubiquitin displayed growth defects on galactose plates containing the respiration inhibitor Antimycin A (AA; Figure S1, lanes 5, 11, 23, 85, 87, 105, 119; Figure S2). The presence of the respiration inhibitor AA requires the cells to metabolize more galactose molecules in order to form colonies, which serves to translate defects in the transcriptional activation of the GAL genes into stronger growth defects on galactose plates. The H10UbD58A mutant strain was also unable to grow on galactose plates in the absence of AA (Figure 1A, line 5), and it was transformed with a multi-copy library of S. cerevisiae genomic DNA fragments [36]. Gal3 was isolated by its ability to confer growth to the H10UbD58A mutant strain on galactose plates upon over-expression (Figure 1A, line 6). The over-expression of Gal3 also dosage-compensated the gal− phenotype of the other H10Ub mutant strains (Figure S3, compare odd and even lanes; the H10UbF4A mutant strain was barely viable and was excluded from further studies). Gal3 sequesters Gal80 in the cytoplasm upon galactose induction [10], and our finding that the over-expression of Gal3 suppressed the gal− phenotype of the H10Ub mutant strains indicated that ubiquitin-mediated protein degradation of Gal80 could be required for galactose induction of the GAL genes and that the gal− phenotype of these H10Ub mutant strains might have been caused by excess Gal80. Consistently, the additional gene deletion of GAL80 suppressed the gal− phenotype of the H10UbD58A mutant (Figure 1A, line 7) and of the other H10Ub mutant strains (Figure S4). Reverse transcription coupled with real-time PCR quantification revealed that galactose induction of GAL1 mRNA relative to ACT1 mRNA was abolished in the H10UbD58A strain and that the over-expression of Gal3 and the additional gene deletion of GAL80 (partially) restored galactose induction (Figure 1B). We performed chase assays with the protein biosynthesis inhibitor cycloheximide and found that HA-Gal80 was indeed degraded in galactose-induced H10Ub cells (Figure 1C, lanes 5 to 8; Figure 1D, white bars). Importantly, HA-Gal80 had become stable in galactose-induced H10UbD58A mutant cells (Figure 1C, lanes 13 to 16; Figure 1D, black bars) as well as in the other gal− H10Ub mutants strains (Figure S5), suggesting that the galactose-stimulated protein degradation of Gal80 is necessary for transcriptional activation of the GAL genes. Our finding that the additional gene deletion of GAL80 suppressed the gal− phenotype of the H10Ub mutant strains provides genetic evidence that the failure to degrade Gal80 had been the cause (and not the consequence) of the gal− phenotype of the H10Ub mutant strains.
E3 ubiquitin ligases add ubiquitin to proteins that are targeted for degradation by the 26S proteasome [18], and Skp1 is an essential component of all SCF E3 ubiquitin ligases [19]. Previously, we had found that the Skp1 derivative Nub-HA-Skp1V90A,E129A (Skp1dM) causes the gal− phenotype when expressed in place of endogenous Skp1 [37]. We had isolated α2 as a multi-copy suppressor and shown that galactose-induced protein degradation of the repressor Mig2, which—like α2 [38]—uses the co-repressor Tup1, was abolished in the skp1dM strain [37]. The most likely explanation was that over-expression of α2 had titrated Tup1 away from GAL1 promoter-bound Mig2, which—like Mig1 [39]—activated transcription in the absence of Tup1. The additional gene deletion of MIG2, however, had only partially suppressed the gal− phenotype of the skp1dM strain [37], suggesting that Skp1 mediated the galactose-induced protein degradation of additional transcription factors. Therefore, we wanted to see if Gal80 was a functionally relevant target of SCF E3 ubiquitin ligases. HA-Gal80 protein was degraded in galactose-induced SKP1 wild-type cells (Figure 2A, lanes 5 to 8; Figure 2B, white bars), while it was stable in galactose-induced skp1dM cells (Figure 2A, lanes 13 to 16; Figure 2B, black bars), indicating that wild-type Skp1 was required for the galactose-induced protein degradation of Gal80. We transformed the skp1dM mutant strain with multi-copy plasmids expressing Sgt1 (which is required for Skp1-dependent cyclin degradation [40]), α2, Ubp3 (which dosage-compensates the gal− phenotype of cells expressing the proteolytically instable Tbp1E186D [41]), and Gal3. The over-expression of Sgt1 suppressed the temperature sensitivity of the skp1dM strain (Figure 2C, line 3). The over-expression of Gal3 and α2 suppressed the gal− phenotype, but not the temperature sensitivity, of the skp1dM mutant strain (Figure 2C, lines 4 and 6), while the over-expression of Ubp3 had no effect (Figure 2C, line 5). Real-time PCR quantification revealed that galactose induction of GAL1 mRNA relative to ACT1 mRNA was abolished in the skp1dM mutant strain (Figure 2D) and that it was restored to some 550-fold in the presence of excess Gal3 and almost fully in the absence of Gal80 (Figure 2D), providing genetic evidence that galactose-stable Gal80 had been the main cause for the gal− phenotype of the skp1dM strain.
F-box proteins provide the substrate specificity to SCF E3 ubiquitin ligases [19], and the deletion of the gene encoding the F-box protein Mdm30 causes a gal− phenotype [20]. Cycloheximide chase assays demonstrated that HA-Gal80 was degraded in galactose-induced BY4741ΔW wild-type cells (Figure 3A, lines 5 to 8; Figure 3B, white bars), while it was stable in galactose-induced ΔMDM30 cells (Figure 3A, lines 13 to 16; Figure 3B, black bars), suggesting that SCFMdm30 targets Gal80 for galactose-induced protein degradation. Importantly, and consistent with a recent report [42], the additional gene deletion of GAL80 suppressed the gal− phenotype of the ΔMDM30 strain (Figure 3C, line 4). Gal80 was still degraded in galactose-induced ΔGAL11 cells (Figure 3A, lanes 21 to 24; Figure 3B, grey bars) and the additional gene deletion of GAL80 did not suppress the gal− phenotype of cells lacking Gal11 (Figure 3C, line 6), confirming that the suppression of the gal− phenotype of the ΔMDM30 strain by the additional gene deletion of GAL80 was gene-specific and that the F-box protein Mdm30 acts genetically upstream of the repressor Gal80, while the Mediator component Gal11 (Med15; which is a target of Gal4 [43]) acts genetically downstream of the repressor Gal80. Real-time PCR quantification of GAL1 mRNA relative to ACT1 mRNA confirmed that the additional gene deletion of GAL80 fully suppressed the transcriptional defect of the ΔMDM30 strain (Figure 3D), suggesting that Mdm30 targets mainly Gal80 for galactose-induced protein degradation. Consistently, GST-Gal80, but not GST, pulled down HA-tagged Mdm30 and Skp1 from yeast extracts (Figure 3E, lanes 8 and 9; Figure S6). Coomassie staining demonstrated that Gal80 and Mdm30 interacted at approximately equal amounts (Figure 3E, lanes 8 and 9). However, Gal80 interacted with Mdm30 (and Skp1) not only in galactose-induced but also in glucose-grown cells (Figure 3D, compare lanes 8 and 9; Figure S6), possibly reflecting the (slower) protein degradation of Gal80 in cells grown with glucose (Figure 3A, lane 4; Figure 3B, white bars). The half-life of Gal80 was calculated to be approximately 3 h in glucose-grown BY4741ΔW cells and approximately 1 h in galactose-induced BY4741ΔW cells. Gal80 had been completely stable in glucose-grown H10Ub cells (Figure 1, lines 1 to 4), indicating that the N-terminal tail of 10 histidines might have interfered with the slow protein degradation of Gal80 in glucose-grown cells. In agreement with the hypothesis that Gal80 is not just degraded in galactose-induced but also in glucose-grown cells (albeit with slower kinetics), Gal80 was poly-ubiquitinated in cells grown with glucose and in cells induced with galactose (Figure 3F, lanes 6 and 7). The amount of poly-ubiquitinated species of Gal80 was only slightly higher in galactose-induced cells as compared to in glucose-grown cells, suggesting that the generation of the poly-ubiquitinated species of Gal80 is rate-limiting, and once generated, poly-ubiquitinated Gal80 is immediately degraded. The ubiquitinated forms of HA-Gal80 are not visible in the input lanes, indicating that only a very small fraction of the Gal80 inside the cell is ubiquitinated at any point in time. The figure further shows that Gal80 was poly-ubiquitinated in wild-type cells as well as in cells lacking Mdm30 (Figure 3F, compare lanes 7 and 9), indicating that Mdm30 is not the only F-box protein targeting Gal80. In order to identify additional SCF E3 ubiquitin ligases targeting Gal80, we tested galactose utilization defective F-box protein gene deletion mutant strains [37] and found that Gal80 was also stable in galactose-induced cells lacking the F-box proteins Das1 and Ufo1 (Figure S7A,B). Importantly, the gal− phenotype of cells lacking Das1 and Ufo1 was suppressed by the additional gene deletion of GAL80 (Figure S7C) and GST-Gal80, but not GST, pulled down Das1, and Ufo1 from yeast extracts (Figure S7D,E), indicating that targeting of Gal80 by at least these three F-box proteins is required for the efficient galactose-induced protein degradation of Gal80. Gal80 interacted with all three F-box proteins in cells grown with glucose and in cells grown with galactose. Consistently, the deletion of MDM30, DAS1, and UFO1 stabilized Gal80 also in glucose-grown cells (Figures 3B and S7B). The signal observed for the pulldown of the F-box proteins with GST-Gal80 was higher in glucose-grown cells than in galactose-induced cells (compare lanes 8 and 9 in Figures 3E and S7D,E). A possible explanation is that in galactose-induced cells, more than in glucose-grown cells, the protein-protein interaction between the F-box proteins and Gal80 resulted in the protein degradation of Gal80, which means that the amount of the F-box protein pulled by GST-Gal80 does not necessarily reflect the strength of the protein-protein interaction. The over-expression of Mdm30 and Ufo1 suppressed the gal− phenotype of cells lacking Das1 (Figure S7F, lanes 3 and 4), indicating that galactose induction requires a critical threshold of Gal80-targeting SCF E3 ubiquitin ligases.
SCF E3 ubiquitin ligases are enzymes that not only target Gal80 for ubiquitin proteasome-mediated protein degradation but also other proteins like Gal4 [20] and Mig2 [37]. It could be argued that defects in the protein degradation of some protein other than Gal80 had caused the gal− phenotype of the H10UbD58A, skp1dM, and ΔMDM30 mutant strains. We have shown that the additional gene deletion of GAL80 suppressed the transcriptional defects of all of these mutants, indicating that Gal80 is the only functionally relevant target, but in order to gain independent evidence that the galactose-induced protein degradation of Gal80 is required for the galactose induction of the GAL genes, we sought to generate a galactose-stable Gal80 derivative that would interfere with transcriptional activation of the GAL genes. Some degraded proteins contain an N-terminal degron, and we performed a series of small N-terminal deletions of Gal80 and tested them for causing defects in galactose utilization. The over-expression of wild-type HA-Gal80 reduced growth on a galactose plate in the presence of the respiration inhibitor Antimycin A (Figure 4A, line 2). The successive deletion of two amino acids increased the growth inhibition, with the deletion derivative lacking the 12 N-terminal amino acids of Gal80 showing the biggest growth inhibition (Figure 4A, line 6). N-terminal deletions of more than 12 amino acids resulted in less inhibition, with the Gal80 deletion derivative lacking the N-terminal 20 amino acids (which removes the first four residues of the Rossmann-fold [44]) having lost the ability to inhibit growth on the galactose plate (Figure 4A, line 10). Real-time PCR quantification of GAL1 mRNA relative to ACT1 mRNA showed that the over-expression of the HA-Gal80 derivative lacking the N-terminal 12 amino acids reduced galactose induction of the GAL1 gene 5- to 3-fold more than the over-expression of wild-type HA-Gal80 (Figure 4B). Cycloheximide chase assays demonstrated that the HA-Gal80 deletion derivative lacking the N-terminal 12 amino acids was indeed stable in galactose-grown cells (Figure 4C, lanes 19 to 24; Figure 4D, black bars), confirming our hypothesis that galactose induction of the GAL1 gene requires protein degradation of the repressor Gal80.
The essential Mediator subunit Srb7 (Med21) plays a pivotal role in the regulation of transcription [45],[46]. In order to identify human proteins interacting with the human Mediator component hSrb7, we fused it to the C-terminal half of ubiquitin that was extended by the RUra3 reporter (Cub-RUra3) and performed a Split-Ubiquitin screen [47],[48] with an expression library of human cDNAs fused to the N-terminal half of ubiquitin (Nub; Figure S8A). The Nub fusion of the human SCF E3 ubiquitin ligase component hSkp1 was isolated by its ability to confer FOA resistance to S. cerevisiae cells expressing hSrb7-Cub-RUra3 (Figure S8B), indicating that both proteins interacted inside the yeast cells. E. coli–expressed GST-hSrb7, but not GST, pulled down Nub-HA-hSkp1 from yeast extract (Figure S8C, lane 6), and E. coli–expressed GST-hSkp1, but not GST, pulled down E. coli–expressed H6-HA-hSrb7 (Figure S8C, lane 3), demonstrating that both proteins interacted directly with each other also in vitro. The human Split-Ubiquitin system (Figure S8D; [49]) was used to demonstrate that both proteins interacted with each other also in vivo (Figure S8E). hSrb7 and hSkp1 are subunits of distinct protein complexes, but the SCF component hSkp1 might play an additional role as a component of Mediator, while the Mediator component hSrb7 might moonlight as a component in an SCF complex. In order to distinguish between these possibilities, we performed co-immunoprecipitations with HeLa extracts and found that hSkp1 pulled down other Mediator components like hMed6 (Figure S9A, lane 3), while hMed6 pulled down other SCF components like hCul1 (Figure S9A, lane 10), indicating that hSrb7 and hSkp1 interacted with each other as components of their own respective complexes. We knocked down hSrb7 and hSkp1 in HeLa cells by RNA interference (Figure S9B), which dramatically reduced the heat-shock induction of the human HSP70B' gene (Figure S9C), indicating that hSrb7 and hSkp1 are functionally relevant for transcription in human cells. Skp1 is a component of the SCF E3 ubiquitin ligases, suggesting that protein degradation could be an important aspect of how Srb7 regulates transcription.
The Split-Ubiquitin assay revealed that also the S. cerevisiae Srb7 and Skp1 proteins interacted with each other in vivo (Figure 5A, line 2). Interestingly, Skp1dM was defective for the protein interaction with Srb7 (Figure 5A, line 4). Our results showed that the Mediator of transcription interacts with SCF E3 ubiquitin ligases, and in order to see if Mediator plays a role in the galactose-induced protein degradation of Gal80, we generated a gal− allele of SRB7 by replacing endogenous Srb7 with a GST fusion to a C-terminal fragment of Srb7 lacking the first 40 amino acid residues (Figure 5B, line 2). The over-expression of Gal3 and the deletion of GAL80 suppressed the gal− phenotype of the GST-Srb7Δ40 strain (Figure 5B, compare lines 1 to 4), indicating that excess Gal80 could have caused the gal− phenotype. The over-expression of Gal3 and the deletion of GAL80 did not suppress the gal− phenotype of cells lacking the Mediator subunit Gal11 (Figure 5B, compare lines 5 to 7), demonstrating that the suppression had been gene-specific and that the Mediator subunit Srb7 acts genetically upstream of Gal80, while the Mediator subunit Gal11 acts genetically downstream of Gal80. The over-expression of α2 and the deletion of MIG2 did not suppress the gal− phenotype of the GST-Srb7Δ40 strain (Figure 5B, compare lines 8 to 11), while the over-expression of α2 and the deletion of MIG2 had suppressed (partially) the gal− phenotype of the skp1dM strain (Figure 2C, line 4 and [37]), suggesting that Skp1 acts genetically upstream of both Gal80 and Mig2, while Srb7 acts genetically upstream of Gal80 only. Cycloheximide chase assays demonstrated that Gal80 was degraded in galactose-induced cells expressing wild-type Srb7 (Figure 5C, lanes 5 to 8; Figure 5D, white bars), but stable in galactose-induced GST-Srb7Δ40 cells (Figure 5C, lanes 13 to 16; Figure 5D, black bars), indicating that Mediator controls the galactose-induced protein degradation of Gal80. Real-time PCR quantification confirmed that galactose induction of GAL1 mRNA relative to ACT1 mRNA was abolished in the GST-Srb7Δ40 strain and that it was almost fully restored by the over-expression of Gal3 and the deletion of GAL80 (Figure 5E), suggesting that the failure of the GST-Srb7Δ40 strain to degrade Gal80 upon galactose induction had been the main cause for the failure to activate the transcription of the GAL1 gene. GST-Srb7, but not GST, pulled down Skp1 from yeast extract, while GST-Srb7Δ40 failed to do so (Figure 5F, lanes 5 and 6), indicating that the protein-protein interaction with Skp1 is mediated by the N-terminus of Srb7, which is the most conserved part of the protein [46]. Our results have shown that the degradation of Gal80 was abolished when endogenous Skp1 was replaced by a mutant Skp1 derivative that failed to interact with Srb7 and when endogenous Srb7 was replaced by a Srb7 mutant protein that failed to interact with Skp1, suggesting that the protein-protein interaction between the Mediator component Srb7 and the SCF component Skp1 is required for the protein degradation of Gal80.
Mediator acts upstream of the activator Gal4 by controlling the galactose-induced protein degradation of the inhibitor Gal80. But how does Mediator know about the switch in carbon source? The protein kinase Snf1 is required for the transcription of glucose-repressed genes in S. cerevisiae, and the deletion of SNF1 resulted in the failure to degrade Gal80 (Figure 6A, lanes 19 to 24; Figure 6B, grey bars), to utilize galactose (Figure 6C), and to activate the GAL1 gene under inducing conditions (Figure 6D). The activating gamma subunit Snf4 is required for the kinase activity of the SNF1 complex and Gal80 was also stable in galactose-induced ΔSNF4 cells (Figure 6A, lanes 31 to 36; Figure 6B, grey bars), indicating that the kinase activity of the SNF1 complex is required for the degradation of Gal80. The additional gene deletion of GAL80 fully suppressed the transcriptional defect of ΔSNF1 and ΔSNF4 cells (Figure 6C, lines 3 and 4; Figure 6D), but no interaction was observed between Snf1 and Gal80 in a pulldown assay (Figure S10), indicating that Snf1 controls GAL1 expression mainly by targeting Gal80 via Srb7 and SCF E3 ubiquitin ligases. The Split-Ubiquitin assay did not reveal an interaction between Snf1 and Srb7 (Figure S11, line 23), however Srb7 is a component of Mediator and Snf1 interacted with the Mediator components Med6 (Figure S11, line 6), Med10 (Figure S11, line 11), Srb6 (Med22; Figure S11, line 21), and Srb11 (CycC; 11, line 27). The protein interaction between the kinase Snf1 and the Mediator component Srb11 had been observed both in vivo and in vitro previously [50],[51]. Srb11 is a cyclin-like cofactor for the protein kinase Srb10 (Cdk8), and the Mediator components Srb10 and Srb11 are both required for the full transcriptional activation of the GAL1 gene [52]. Gal80 was stable in galactose-induced ΔSRB10 and ΔSRB11 cells (Figure S12A, lanes 7 to 12 and 19 to 24; Figure S12B), confirming that Snf1 might transduce the signal to degrade Gal80 via the Mediator subunit Srb11. The additional gene deletion of GAL80 suppressed the galactose utilization defect of cells lacking Srb10 and Srb11 (Figure S12C, lines 3 and 5), providing genetic evidence that galactose-stable Gal80 had caused the gal− phenotype of ΔSRB10 and ΔSRB11 cells.
The SNF1 kinase is activated by the absence of glucose, but transcriptional activation of the GAL genes requires additionally the presence of galactose, as transcription of GAL1 is not activated in cells grown with—for example—raffinose (Figure S13B). Consistently, Gal80 was more stable in cells grown with raffinose than in cells grown with galactose (Figure S14). The half-life of Gal80 in BY4741ΔW cells was calculated to be approximately 3 h when the cells were grown with glucose, 2 h when the cells were grown with raffinose, 1 h when galactose-induced cells had been pre-grown with glucose, and half an hour when the galactose-induced cells had been pre-grown with raffinose. However, our observations also indicate that active SNF1 kinase is necessary but not sufficient for the galactose-stimulated protein degradation of Gal80. An additional transducer that signals the presence of galactose is apparently required. A possible candidate for such a signal transducer is Gal3, as it is known to bind both galactose and Gal80 [14]. Cells lacking Gal3 display a strong gal− phenotype (Figure S15A, lines 3 and 4), which is suppressed by the additional gene deletion of GAL80 (Figure S15A, lines 5 and 6), but the degradation of Gal80 in galactose-induced cells remained unchanged upon the deletion of GAL3 (Figure S15B, lanes 5 to 8; Figure 15C), indicating that Gal3 does not play a role in the galactose-induced protein degradation of Gal80 and that galactose must utilize another transducer to stimulate the protein degradation of Gal80.
Mediator was isolated by its ability to respond to transcriptional activators, and all studies published about Mediator have focused on the role of Mediator past its recruitment to the promoter by the activator [28]. Once recruited, Mediator is required to recruit the General Transcription Factors and RNA Polymerase II and to initiate transcription [29]. Mediator also affects post-initiation steps by affecting transcription elongation and chromatin structure [31]. We have shown here that Mediator additionally acts upstream of the activator Gal4 by controlling the degradation of the inhibitor Gal80. In cells grown with glucose, Gal80 binds to the activation domain of Gal4 and prevents it from activating transcription. Upon galactose induction, Mediator initiates the degradation of Gal80 via its interaction with the SCF E3 ubiquitin ligase component Skp1. Therefore, Mediator actually orchestrates its own recruitment to the GAL1 promoter by regulating the activity of Gal4 (Figure 7).
SCFMdm30 targets not only Gal80 but also Gal4 in galactose-induced cells, leading to the mono-ubiquitination and subsequent poly-ubiquitination and protein degradation of Gal4 [20]. In galactose-induced cells lacking Mdm30, Gal4 is no longer ubiquitinated and no longer degraded [20]. One could argue that changes in the proteolytic stability of Gal4 or in its mono-ubiquitination status might have been the cause for the gal− phenotypes that we have observed for the various H10UbD58A, skp1, mdm30, srb7, and snf1 mutant strains described here. Therefore, it is important to note that our claim that the degradation of Gal80—and not the degradation of Gal4—is required for the transcriptional activation of the GAL genes is not just based on a simple correlation between the proteolytic stability of Gal80 and the inability of the cell to activate transcription of the GAL1 gene, but on functional suppression. The additional gene deletion of GAL80 fully suppressed the transcriptional defects of the H10UbD58A, skp1, mdm30, srb7, and snf1 mutant strains. This means that in the absence of Gal80, Gal4 activated transcription in all these mutant strains just fine, which demonstrates that any effects that these strain mutations might have had on Gal4 were not relevant for Gal4's function as a transcriptional activator. Therefore, while mono-ubiquitination of Gal4 was certainly affected in the H10UbD58A strain (since endogenous wild-type ubiquitin had been replaced with H10UbD58A), Gal4-H10UbD58A fully activated transcription of the GAL1 gene in the absence of Gal80, suggesting that H10UbD58A still protected Gal4 from the UAS-stripping activity of the 19S proteasome [25]. Furthermore, Gal4 fully activated transcription in cells lacking both Mdm30 and Gal80, which argues that Gal4 does not have to be degraded to become transcriptionally active. In addition, we have generated a galactose-stable Gal80 derivative that inhibited galactose induction in otherwise wild-type cells, which means that we have presented evidence for our claim that galactose induction requires Gal80 degradation that did not rely on a mutant strain background.
The deletion of the three F-box protein-coding genes MDM30, DAS1, and UFO1 completely abolished galactose induction of GAL1 mRNA (Figure 3D and [37]). Das1 and Ufo1 (but not Mdm30) also target the repressor Mig2 for galactose-induced protein degradation [37]. However, the additional gene deletion of MIG2 did not increase galactose induction of GAL1 mRNA in the ΔUFO1 strain and had only a very small effect on the galactose induction of the GAL1 mRNA in the ΔDAS1 strain [37]. Therefore, an additional target for Das1 and Ufo1 had been proposed, and we have now shown here that Gal80 is this functionally relevant target, as Gal80—like Mig2 [37]—became stable in galactose-induced cells lacking Das1 and Ufo1 (Figure S7A and S7B), and the additional gene deletion of GAL80 suppressed the gal− phenotype of both the ΔDAS1 and the ΔUFO1 strains (Figure S7C). We are proposing that a critical concentration of the three F-box proteins Mdm30, Das1, and Ufo1 is required for the galactose-stimulated protein degradation of Gal80. If any one of these three F-box proteins is missing, the concentration of the remaining two F-box proteins is insufficient for targeting of Gal80; Gal80 is not degraded and excess Gal80 prevents Gal4 from activating the GAL genes under inducing conditions. In support of this model (Figure 7), we were able to show that the gal− phenotype of ΔDAS1 cells was suppressed by the over-expression of Ufo1 and Mdm30 (Figure S7E).
Gal3 sequesters Gal80 in the cytoplasm upon galactose induction [14]. The gene deletion of GAL3 had caused a gal− phenotype that was suppressed by the additional gene deletion of GAL80, but the protein degradation of Gal80 was still stimulated in galactose-induced cells lacking Gal3 (Figure S15), indicating that instable Gal80 was not sufficient to allow Gal4 to activate transcription in the absence of Gal3. On the other hand, sequestration of Gal80 into the cytoplasm by endogenous levels of Gal3 was not sufficient to allow Gal4 to activate transcription in the presence of stable Gal80. Apparently, sequestration of Gal80 into the cytoplasm by Gal3 and ubiquitin-mediated protein degradation of Gal80 are both required for the galactose induction of the GAL genes.
Contrary to a previous report [20], we found that the deletion of the gene encoding the F-box protein Mdm30 abolished galactose induction of the GAL1 mRNA. We have grown the cells in glucose liquid media prior to the switch to galactose liquid media—which is consistent with the switch in carbon sources conducted for the plate assay—while Muratani et al. grew the cells in raffinose liquid media prior to the switch to galactose liquid media. In order to determine if this difference in protocols was the cause for the difference in results, we performed the galactose induction with cells that had been pre-grown in raffinose, and we found that in this case, galactose-induced protein degradation of Gal80 and galactose induction of GAL1 mRNA relative to ACT1 mRNA were restored in the ΔMDM30 strain (Figures S13B and S14). One other remarkable difference between the two growth protocols is the speed of induction. Galactose induction of GAL1 mRNA relative to ACT1 mRNA takes 4 h if the cells are pre-grown with glucose and only 1 h if the cells are pre-grown with raffinose (Figure S16). Consistently, cycloheximide chase assays demonstrate that Gal80 is more rapidly degraded in galactose-induced cells when the cells had been pre-grown with raffinose instead of with glucose (compare Figures 3B and S14B). The half-life of Gal80 in galactose-induced BY4741ΔW cells was approximately 1 h when the cells had been pre-grown with glucose and only half an hour when the cells had been pre-grown with raffinose. The correlation of the kinetics of galactose-induced Gal80 destruction and GAL1 mRNA production suggests that the degradation of Gal80 is the rate-limiting step for the galactose induction of the GAL1 gene.
The S. cerevisiae strain SUB288 [53] has all chromosomal ubiquitin genes deleted and allows for the expression of plasmid-born ubiquitin derivatives in place of endogenous ubiquitin (see Table S1 for the genotypes of the strains and Table S2 for the sequences of PCR primers). However, the strain fails to grow on galactose plates containing the respiration inhibitor Antimycin A (AA). Transformation of the strain with single-copy vectors expressing Gal3 from its own promoter allowed the strain to grow on galactose AA plates and sequencing of the chromosomal GAL3 gene demonstrated that SUB288 carries a frame shift in the third codon of GAL3. The TRP1 and LEU2 genes were deleted and the defective gal3 gene was repaired by homologous recombination with a wild-type GAL3 PCR fragment followed by selection on a galactose AA plate or with YIplac204-GAL3, a derivative of the TRP1-marked integrative vector YIplac204 [54] containing the GAL3 gene, resulting in SUB288GAL3ΔWL+316-Ub and SUB288GAL3ΔL+316-Ub, respectively. DNA sequencing of PCR fragments derived from genomic DNA was used to confirm that the GAL3 gene had been repaired. The ubiquitin point mutants were generated by two-step PCR with degenerate primers and cloned into the LYS2-marked single-copy vector RS317 [55] containing the ACT1 promoter-terminator cassette and into RS317 expressing 10 histidines from the ACT1 promoter. The 317-Ub and 317-H10-Ub plasmids were transformed into SUB288GAL3ΔWL+316-Ub and 316-Ub was shuffled out on FOA plates. All ubiquitin mutant strains were confirmed by DNA sequencing. The GAL80 gene of SUB288GAL3ΔWL+317-Ub was knocked out with a derivative of NKY51 [56], which carried the hisG-URA3-hisG cassette in the BglII site at nucleotide 612 of GAL80. 317-Ub was replaced by 316-Ub via plasmid loss, and plasmid shuffle was used to generate the 317-H10-Ub and 317-H10-UbD58A strains carrying hisG integrated into GAL80. The essential SRB7 gene of JD52 and JD52ΔGAL80 was knocked out with a PCR fragment containing the HIS3 gene flanked by 50 bp of SRB7 promoter and terminator in the presence of 33-SRB7, a derivative of the URA3-marked single-copy vector YCplac33 [54] that expressed Srb7 from its own promoter. GST-Srb7 and GST-Srb7Δ40 were expressed from the TRP1-marked multi-copy vector YG1μ under the control of the ADH1 promoter. BY4741ΔW and BY4742ΔW and their gene deletion derivatives were obtained from the respective EUROSCARF strains by inserting hisG into the TRP1 gene with the help of NKY1009 [56]. YEp13-GAL3 was isolated from a LEU2-marked multi-copy YEp13-based genomic DNA library [37] as a multi-copy suppressor of the gal− phenotype of the H10UbD58A strain. YEp13-GAL3 contains a 2,643 bp genomic DNA fragment with the entire GAL3 gene, including 842 bp of promoter and 238 bp of terminator DNA. 112-GAL3 is a derivative of the TRP1-marked multi-copy vector YEplac112 [54] containing the genomic GAL3 fragment. 314-Gal3 is a derivative of the TRP1-marked single-copy vector RS314 [55], expressing Gal3 from the ACT1 promoter. 315-Gal3 is a derivative of the LEU2-marked single-copy vector RS315 [55], expressing Gal3 from the ACT1 promoter. 316-HA-Gal80 is a derivative of RS316, expressing Gal80 from the ACT1 promoter. The N-terminal deletion derivatives of Gal80 were cloned into the same vector. 423-HA3-Mdm30, 423-HA3-Das1, and 423-HA3-Ufo1 are derivatives of the HIS3-marked multi-copy vector RS423 [55], expressing Mdm30, Das1, and Ufo1 tagged with three HA epitopes from the ACT1 promoter. 424-GST and 424-GST-Gal80 are derivatives of the TRP1-marked multi-copy vector RS424 [55], expressing GST and GST-Gal80 from the ACT1 promoter. YIplac128-Snf1c-HA3H10 is a derivative of the LEU2-marked integrative vector YIplac128 [54], containing a C-terminal BglII-SalI fragment of SNF1 lacking the stop codon, and YIplac128-Skp1c-HA3H10 is a derivative of YIplac128 containing a C-terminal EcoRI-SalI fragment of SKP1 lacking the stop codon. Snf1-HA3H10 was expressed from the SNF1 promoter following digestion with MluI and integration into the SNF1 locus, while Skp1-HA3H10 was expressed from the SKP1 promoter following digestion with AvaI and integration into the SKP1 locus.
A Clontech library derived from human B-cell cDNAs was partially digested with Sau3A and cloned into the BglII site of PADNX-Nub-IBC [57] in all three reading frames, resulting in 60,000 independent DH5α transformants. hSrb7 was cloned into Pcup1-Cub-RUra314 [57] and transformed together with the Nub library into JD52 [58], resulting in 160,000 transformants, which were plated onto FOA plates containing 10 µM CuSO4. The Nub plasmids from the 10 arising colonies were isolated and transformed back into JD52 containing hSrb7-Cub-Ura314. Only one was plasmid-linked, and it contained the entire hSkp1 open reading frame fused to Nub-HA.
HeLa cells were grown to 80% confluency and transfected with 2 µg of pSuper (OligoEngine) construct and 5 µl of lipofectamin in serum-free DMEM for 5 h before being transferred into regular DMEM. The three constructs used were an empty vector as a negative control, siRNA specific for hSKP1, and siRNA specific for hSRB7. 48 h after transfection, one set of cells was heat-shocked at 45°C for 15 min and allowed to recover for 1 h in a 37°C incubator. A non-heat-shock sample was also incubated at 37°C for an identical length of time. These cells were then harvested by trypsinization and their mRNA was extracted using a Qiagen RNA Easy Kit. 300 nM of mRNA was utilized for reverse transcription primed by random hexamers, and the cDNA was quantified using Sybr-Green in an ABI Prism. Primers for HSP70B' mRNA were 5′-ccccatcattgaggaggttg-3′ and 5′-gaagcagaagaggatgaacc-3′. Primers for hSKP1 mRNA were 5′-gcaaagagaaccagtggtgtga-3′ and 5′-aggtttgggatctgtgctcaa-3′. Primers for hSRB7 mRNA were 5′-aatgtggtcctcctgcctctt-3′ and 5′-ccagaagcatgtctcctcgata-3′. Primers for GAPDH mRNA were 5′-ctctctgctcctcctgttcgac-3′ and 5′-tgagcgatgtggctcggct-3′.
S. cerevisiae cells were cultured in synthetic complete 2% (w/v) glucose medium at 28°C. At OD600 nm = 1, the cells were collected by centrifugation. Galactose induction was performed by resuspending the cells in 2% galactose medium and incubation for the indicated amount of time. Total RNA was isolated using the RNAeasy Mini Kit (Qiagen) according to the manufacturer's protocol. cDNA was generated by reverse transcription PCR using Taqman MicroRNA Reverse Transcription Kit (Roche Applied Biosystems). Quantitative real-time PCR was performed using SYBR Green PCR Master Mix (Applied Biosystems). Primers used for ACT1 mRNA were 5′-gaccaaactacttacaactcca-3′ and 5′-cattctttcggcaatacctg-3′. Primers used for GAL1 mRNA were 5′-acttgcaccggaaaggtttg-3′ and 5′-ttggtacatcaccctcacagaaga-3′. All mRNA quantifications were performed three times, and the error bars represent the standard deviations.
HeLa cells were grown to 80% confluency before they were transfected with 2 µg of pCMV-myc-hSKP1 or pCMV-myc vector and 5 µl of lipofectamin in serum-free DMEM for 5 h before being transferred into regular DMEM. The cells were harvested 48 h after transfection by trypsinization and lysed in 1× PBS by freeze-thaw. The cell lysate was subsequently agitated on a rotor with 2 µl of anti-myc affixed agarose beads (Sigma) in 500 µl of ice cold 1× PBS overnight. The beads were washed four times with 1 ml PBS prior to heat elution at 95°C for 15 min. Proteins were separated on a 12% gel, transferred to a nitrocellulose membrane, which was probed with anti-Med6 rabbit polyclonal antibody (Abcam).
HeLa cells were grown to 80% confluency before they were harvested by trypsinization and lysed in 1× PBS by freeze-thaw. The cell lysate was diluted 1∶5 with RIPA buffer (50 mM Tris-HCl ph 8, 150 mM NaCl, 2 mM EDTA, 1% NP-40, 0.5% Sodium deoxycholate, 0.1% SDS) and incubated with 10 µl of rProtein G Sepharose (GE Healthcare) as well as 5 µl of anti-Med6 rabbit polyclonal antibody (Abcam) or anti-Cul1 mouse monoclonal antibody (Abcam) for 3 h. The sepharose was washed four times with 1 ml PBS and heat eluted at 95°C for 15 min. Proteins were separated on a 12% gel and transferred to a nitrocellulose membrane, which was probed with the reciprocal antibody (anti-Cul1 mouse monoclonal antibody (Abcam) or anti-Med6 rabbit polyclonal antibody (Abcam), respectively).
GST pulldown assays were performed using whole cell S. cerevisiae extracts prepared by bead beating in yeast lysis buffer (100 mM Tris pH 7.5, 50 mM KCl, 1 mM EDTA, 0.1% NP40) and whole cell E. coli extracts prepared by freeze-thaw in PBS (Phosphate-Buffered Saline). 500 µl of whole cell extract was added to equilibrated glutathione beads (Amersham Biosciences) containing 2 mM PMSF and 1 mM DTT. The reaction mixture was incubated at 4°C for 1 h. The sample was centrifuged at 3,000 rpm and the supernatant was removed. The glutathione beads were washed five times before Western Blot analysis.
S. cerevisiae cells were grown in 50 ml synthetic complete 2% glucose medium to OD600 nm = 1 and harvested by centrifugation. The cell pellets were suspended in 1 ml yeast breaking buffer (Triton X-100, 10% SDS, 5 M NaCl, 1 M Tris-Hcl pH 8, 0.5 M EDTA; Figure 5D) or yeast lysis buffer (Figure S6), pipetted into a screw-cap microcentrifuge tube containing acid-washed glass beads (Sigma-Aldrich, USA), and 2 mM PMSF was added. The tubes were then subjected to homogenization with a bead beater for 1 min and then rested on ice for 3 min. This process was repeated for three times. The samples were then centrifuged for 15 min at 13,000 rpm, and the supernatants were incubated with 10 µl of equilibrated nickel beads for 1 h at 4°C. After incubation, the samples were centrifuged at 3,000 rpm for 2 min. The nickel beads were washed with 1 ml yeast breaking/lysis buffer containing 20 mM imidazole. This washing process was repeated five times. The bound protein was eluted from the nickel beads using 100 µl of yeast breaking/lysis buffer with 500 mM imidazole for 30 min. This process was repeated two times. The supernatant was collected and stored at −80°C.
S. cerevisiae cells were grown in liquid drop out media containing 2% glucose or raffinose to OD600 nm = 1. Half of the cultures were induced in liquid media containing 2% galactose for 1 h before the addition of 200 mg/l cycloheximide (Sigma). Aliquots were taken at the indicated time points, and cellular proteins were analyzed by Western Blot with primary antibodies against hemagglutinin (HA; Roche) and carboxypeptidase Y (CPY; Molecular Probes), followed by staining with a horseradish peroxidase-coupled secondary anti-mouse IgG antibody and by Coomassie Brilliant Blue (Sigma) staining. The intensities of the bands were quantified with Image J (rsb.info.nih.gov/ij/index.html). The ratio of the band intensities before the addition of cycloheximide (time = 0) was set as 1, and the error bars represent the deviations between duplicates. Representative Western blots are shown. No significant differences were observed when the HA-Gal80 bands were normalized to CPY or to Coomassie staining. The half-life of Gal80 was calculated using trendline (excel).
|
10.1371/journal.pgen.1000421 | A Genome-Wide Association Study in Chronic Obstructive Pulmonary Disease (COPD): Identification of Two Major Susceptibility Loci | There is considerable variability in the susceptibility of smokers to develop chronic obstructive pulmonary disease (COPD). The only known genetic risk factor is severe deficiency of α1-antitrypsin, which is present in 1–2% of individuals with COPD. We conducted a genome-wide association study (GWAS) in a homogenous case-control cohort from Bergen, Norway (823 COPD cases and 810 smoking controls) and evaluated the top 100 single nucleotide polymorphisms (SNPs) in the family-based International COPD Genetics Network (ICGN; 1891 Caucasian individuals from 606 pedigrees) study. The polymorphisms that showed replication were further evaluated in 389 subjects from the US National Emphysema Treatment Trial (NETT) and 472 controls from the Normative Aging Study (NAS) and then in a fourth cohort of 949 individuals from 127 extended pedigrees from the Boston Early-Onset COPD population. Logistic regression models with adjustments of covariates were used to analyze the case-control populations. Family-based association analyses were conducted for a diagnosis of COPD and lung function in the family populations. Two SNPs at the α-nicotinic acetylcholine receptor (CHRNA 3/5) locus were identified in the genome-wide association study. They showed unambiguous replication in the ICGN family-based analysis and in the NETT case-control analysis with combined p-values of 1.48×10−10, (rs8034191) and 5.74×10−10 (rs1051730). Furthermore, these SNPs were significantly associated with lung function in both the ICGN and Boston Early-Onset COPD populations. The C allele of the rs8034191 SNP was estimated to have a population attributable risk for COPD of 12.2%. The association of hedgehog interacting protein (HHIP) locus on chromosome 4 was also consistently replicated, but did not reach genome-wide significance levels. Genome-wide significant association of the HHIP locus with lung function was identified in the Framingham Heart study (Wilk et al., companion article in this issue of PLoS Genetics; doi:10.1371/journal.pgen.1000429). The CHRNA 3/5 and the HHIP loci make a significant contribution to the risk of COPD. CHRNA3/5 is the same locus that has been implicated in the risk of lung cancer.
| There is considerable variability in the susceptibility of smokers to develop chronic obstructive pulmonary disease (COPD), which is a heritable multi-factorial trait. Identifying the genetic determinants of COPD risk will have tremendous public health importance. This study describes the first genome-wide association study (GWAS) in COPD. We conducted a GWAS in a homogenous case-control cohort from Norway and evaluated the top 100 single nucleotide polymorphisms in the family-based International COPD Genetics Network. The polymorphisms that showed replication were further evaluated in subjects from the US National Emphysema Treatment Trial and controls from the Normative Aging Study and then in a fourth cohort of extended pedigrees from the Boston Early-Onset COPD population. Two polymorphisms in the α-nicotinic acetylcholine receptor 3/5 locus on chromosome 15 showed unambiguous evidence of association with COPD. This locus has previously been implicated in both smoking behavior and risk of lung cancer, suggesting the possibility of multiple functional polymorphisms in the region or a single polymorphism with wide phenotypic consequences. The hedgehog interacting protein (HHIP) locus on chromosome 4, which is associated with COPD, is also a significant risk locus for COPD.
| COPD is expected to be the third leading cause of worldwide mortality and the fifth leading cause of morbidity by the year 2020 [1]. Cigarette smoking is the major risk factor for COPD but smokers show considerable variation in their risk of developing airflow obstruction. Familial aggregation studies suggest a strong genetic component to this risk [2]–[8]. However the only proven genetic risk factor for COPD is severe deficiency of α1-antitrypsin [9], which is present in only 1–2% of individuals with COPD. This suggests that other genes have yet to be identified that predispose smokers to airflow obstruction. We report the first genome wide association study (GWAS) for COPD. Our primary discovery sample was a case-control population collected from Bergen, Norway, and we used three independent study cohorts to further evaluate the top associations emerging from the GWAS analysis.
Baseline characteristics of the subjects used in the GWAS and subsequent replication studies are presented in Table 1.
We used a multi-stage replication design (Figure 1) for this study. The genome-wide association analyses of the COPD case-control status in the Bergen cohort identified several significant associations, including three SNPs on chromosome 5 that reached the level of genome-wide significance (Table S1). The Q-Q plot showing the distribution of observed P values from the discovery cohort is shown in online Figure S1. The top 100 SNPs were then evaluated in the ICGN population and 8 were replicated at a nominal p value of 0.05 (SNP rs11219732 showed inconsistent risk alleles in the Bergen and ICGN population and hence was not considered further, Table 2). Two of the three SNPs (rs7727670 and rs7341022 on chromosome 5) that showed genome-wide significance in the Bergen cohort did not replicate in the ICGN population. The SNPs showing the most definitive evidence for replication were rs8034191 and rs1051730 in the CHRNA3/5 locus.
Several additional SNPs were later analyzed in the CHRNA3/5 region in the Bergen and ICGN populations (Table S2). One non-synonymous polymorphism in CHRNA5 (rs16969968) which coded for the substitution of an asparagine for an aspartic acid at amino acid 398) was associated with COPD in the Bergen (p = 8.8×10−4) and ICGN (p = 2.78×10−6) cohorts (combined p value 5.08×10−8). Since this SNP showed a weaker association than both rs8034191 and rs1051730 it was not considered as a causal variant.
We then tested the 7 SNPs that showed definite or nominal significance in the NETT-NAS case-control population, and the results are provided in Table 2. These results further confirmed the association of two SNPs at the CHRNA3/5 locus with COPD (p = 2.5×10−3, OR = 1.43, combined p value: 1.48×10−10 for rs8034191 and p = 0.017, OR = 1.32, combined p value 5.74×10−10 for rs1051730). Two SNPs (rs1828591 and rs13118928) at the HHIP locus on chromosome 4 also showed consistent replication across the three cohorts, but the combined p values did not reach genome-wide significance (1.47×10−7 and 1.67×10−7 respectively).
The only significant associations in the Boston Early-Onset COPD families were with the rs8034191 and rs1051730 SNPs at the CHRNA 3/5 locus (p = 0.03 and 0.03 respectively) and the rs1828591 and rs13118928 SNPs at the HHIP locus (p = 0.0025 and 0.0014 respectively) with post bronchodilator FEV1. None of the SNPs was significantly associated with a diagnosis of COPD. Since the ICGN cohort had recruited subjects with a wide range of lung function, we also analyzed the association of the CHRNA 3/5 markers with post bronchodilator FEV1 after adjusting for age, height, gender, pack years and smoking status. The results show that CHRNA 3/5 SNPs were associated with FEV1 in the ICGN population (p values 1.04×10−4 and 1.75×10−5 for rs8034191 and rs1051730 respectively).
The COPD associated region on chromosome 15 spans seven genes (Figure 2). Cholinergic nicotinic receptor subtypes α3, α5 and β4; IREB2, PSMA4, NP_001013641.2 (a gene with unknown function) and Q9UD29 (Surfactant protein B (SP-B)-binding protein). A partial map of the region is shown in online Figure S2. SP-B binding protein is a DNA binding protein which binds to the promoter of SP-B and affects its expression [10]. Since SP-B is a critical surfactant in the lungs [11], we sequenced the SP-B binding protein in 30 COPD subjects who are homozygous for the risk allele of rs8034191 but did not identify any polymorphisms in this gene.
The p values reported above were based on the adjusted analyses correcting for smoking exposure. The results from the unadjusted association analyses for COPD status were highly significant (Bergen 2×10−4 and 4×10−4; ICGN 7.46×10−7 and 1.40×10−6; NETT/NAS, 2.0×10−5 and 2.5×10−4 and combined p values of 1.86×10−12 and 6.6×10−11 for rs8034191 and rs1051730 respectively; Table S3). Although the adjustments for smoking exposure attenuated the p values, they still remained highly significant (Table 2). In the Norwegian discovery cohort, a significant genotype-by-environment interaction (P = 0.002, Table 3) was observed with a substantially higher risk of COPD in current smokers carrying the rs8034191 C allele (OR = 2.00) than in former smokers (OR = 1.10). In the overall population, the C allele of rs8034191 was estimated to have a population attributable risk of 12.2% for COPD. This risk was 14.3% in current smokers and 3.1% in former smokers. The p values were attenuated in the ICGN family-based population following adjustment for age, sex, pack-years of smoking and center but remained highly significant (Table 2). We identified a SNP by pack-years interaction (p = 0.0037 for rs8034191), however no significant SNP by current smoking status interaction (p = 0.85) was detected in the ICGN population.
Testing directly for an association between the amount of smoking, measured as pack-years, within cases and controls respectively, with the SNP rs8034191, demonstrated no such association in the Norway discovery cohort (P = 0.63 and 0.47, respectively). We also carried out tests comparing allele frequencies for current and former smokers and heavy and light smokers, (two different definitions, using pack years of smoking and length of smoking history) within cases and controls separately (Table 3). The only significant association observed was in comparing current and former smokers among the controls (p = 0.028). Similarly, the rs8034191 SNP was not associated with pack-years smoked in the NETT cases or in the NAS controls.
We have demonstrated and replicated genetic associations between SNPs at the CHRNA3/5 locus and COPD in four study populations. The estimated population attributable risk from this locus was 12.2% and represents the discovery of a common major locus contributing to COPD in the general population. However, a potential complication with the interpretation of these findings is the possibility that differences in smoking behavior, likely related to nicotine addiction, between COPD cases and controls may drive the observed association. This is similar to the recently reported association of CHRNA3/5 SNPs with lung cancer [12]–[14].
In the current study populations, only limited assessment of nicotine addiction is available: (i) whether subjects were still smoking at the time of study participation, and (ii) their lifetime smoking intensity. Thus, we have limited ability to disentangle a genetic determinant of smoking behavior from a genetic determinant of COPD through an alternative pathway. There are several pieces of evidences to suggest that there could be a direct effect of CHRNA3/5 locus on COPD, independent of smoking behavior. First, to the extent that smoking behavior is captured in pack-years, this effect should be factored out by the statistical design in which the discovery analyses used a logistic regression model incorporating pack-years, age and gender as covariates. The adjustments for pack-years smoked, age and gender were also performed in all the replication analyses. However, pack-years smoked only partially capture smoking behavior. Many other factors, such as depth of inhalation, number of puffs per cigarette, and age of starting smoking are also likely to affect the toxicant exposure and effect. Second, we also tested directly for an association between the amount of smoking, measured as pack-years, within cases and controls with the SNP rs8034191. There was no significant association between the SNPs and pack-years of smoking in the Bergen and NETT/NAS populations. This is consistent with the observed allele frequency among the Norwegian pediatric general population sample (minor allele frequency = 0.326, n = 551) which is between that observed for cases and controls and not significantly different from either.
We observed a genotype-by-environment interaction between the risk of the rs8034191 genotype and current smoking status on COPD in the Norwegian sample (P = 0.002, Table 3), showing a substantially higher risk of COPD in current smokers carrying the rs8034191 C allele (OR = 2.0) than in former smokers (OR = 1.1). There are several possible explanations for this gene-by-environment interaction. First, it could relate to nicotine addiction; smokers that have greater difficulty quitting smoking may be more likely to develop COPD. Alternatively, it could indicate that a subset of individuals is at greater risk for developing COPD if they continue to smoke. A similar interaction with current smoking was not identified in the ICGN families. Since all the COPD patients in the NETT population were former smokers, we could not address this question in the NETT/NAS study.
The association of smoking dependence was explored in the lung cancer report by Hung and colleagues [13] who did not detect any association with individual Fagerstrom indices of nicotine addiction or when comparing controls with a heaviness of smoking index (HSI). Another lung cancer report by Amos and colleagues [12] did show weak evidence for association with smoking behavior, while a further report by Thorgeirsson and colleagues[14] showed very strong association with smoking behavior. Association of this locus with nicotine dependence has been reported in two other studies [15],[16]. Thus, it is reasonable to conclude that the variants may affect smoking behavior, at the same time as having a significant effect on COPD and other smoking related diseases such as lung cancer and peripheral arterial disease [12]–[14].
The CHRNA3/5 SNPs were also associated with lung function (FEV1) in the ICGN and BEOCOPD populations. These SNPs were shown to be associated with FEV1 in the British Birth Cohort (rs8034191 and rs1051730, p = 0.029 and 0.023, respectively (http://www.b58cgene.sgul.ac.uk/, accessed [3/7/2008]). Historically, nicotinic receptors are classified as neuronal or muscle-type, based on their initial site of identification and composite subunits [17]. Cholinergic activity in the airways primarily induces tracheo-bronchial smooth muscle contraction and mucous secretion. However, there is an increasing body of literature showing the importance of extra-neuronal cholinergic signaling [18] in the lung.
The association of the SNPs at the chromosome 4 HHIP (Hedgehog-Interacting Protein) locus is also interesting, though it did not reach the stringent genome-wide significance levels in the populations studied in this manuscript. These SNPs were also associated with FEV1 in the BEOCOPD study (rs1828591 and rs13118928, p = 0.0025 and 0.0014). The same SNPs are also associated with FEV1 in the British Birth Cohort (rs1828591 and rs13118928, p = 0.039 and 0.038, respectively) but were not associated with FEV1 in the ICGN population.
In another manuscript in this issue of the journal, genome-wide association analysis results for FEV1/FVC in the Framingham Heart Study (FHS) are reported (Wilk et al). Due to differences in genotyping platforms, the most significantly associated SNPs on chromosome 15 in our study were not genotyped in FHS. Analysis of the genotyped SNPs in the chromosome 15 region in the FHS indicated no significant association with COPD, but association with FEV1 percent predicted was observed with one SNP in LD with rs8034191 (rs11636431 p value 0.007). Evaluation of the imputed data for the most significantly associated SNPs in our populations did not show association with COPD in FHS. Several factors could contribute to the absence of association to the COPD phenotype in FHS: (1) The FHS cohort is a population-based collection, while our studies evaluated populations ascertained specifically for COPD; (2) The FHS cohort was recruited over three decades, while our cohorts represent more recent recruitments (in the last 5–10 years)-smoking habits have changed over time, and it is also possible that COPD clinical characteristics have changed over this period; (3) Our cohorts include a greater proportion of severe COPD subjects than in FHS; and (4) There could be differences in linkage disequilibrium patterns between study populations. Further studies will be required to define the specific genetic determinants influencing COPD on chromosome 15, the relationship of these genetic factors to smoking behavior, and the characteristics of COPD subjects influenced by these genetic determinants.
The association of the Chromosome 4 region in the FHS cohort was genome-wide significant for the FEV1/FVC ratio and was also associated with COPD. This association was subsequently replicated in the Family Heart Study population. Though the HHIP locus association in our study did not reach genome-wide significance, the additional evidence from the FHS and Family Heart Study underscore the importance of this locus on COPD susceptibility.
We used independent populations with varying COPD severity, independent genotyping platforms and stringent statistical significance criteria to define genome-wide significant associations. We used consensus criteria for replication using a multi-stage replication design with similar phenotypes, the same genetic model and direction of association [19]. The levels of statistical significance of the association for our most significant results in the CHRNA3/5 region were consistent in all of the populations studied and are unlikely to be false positive results. The p values after adjusting for multiple testing using the most conservative Bonferroni correction were 7.3×10−5 and 2.83×10−4 for the SNPs rs8034191 and rs1051730 respectively. Though this can be considered as strength, the conservative approach for SNP confirmation that we have used may lead to larger false negative rates. However, with the inconsistent results of previous complex disease genetic association studies, we contend that a conservative approach is appropriate. We selected only the top 100 SNPs from the GWAS for subsequent replication study and a larger number of significant associations may have been uncovered if more of the most promising SNPs had been followed up. A negative association in the replication studies may not rule out a true association, since the power to detect association in the replication populations may be limited. The primary replication cohort (ICGN) is moderately powered to detect the replicated associations. Though the sample sizes of the NETT/NAS and BEOCOPD studies are relatively low, these studies include a large percentage of severely affected individuals, who may be enriched for COPD susceptibility genes. This likely account for the high rate of replication in these populations. COPD is a heterogeneous disease and we used a spirometry-based definition for COPD in all of the populations. Differences in smoking exposure, current smoking status, entry criteria and geographic origin of the cohorts might contribute to phenotypic heterogeneity and may lead to lack of replication. The fact that the replicated associations holds-up strongly and consistently in all the populations shows that phenotypic heterogeneity likely has little effect on the most significant results.
In summary, we have identified robust evidence of association of COPD with the α-nicotinic receptor (CHRNA 3/5) and HHIP loci. The hedgehog (Hh) gene family encodes signaling molecules that play an important role in regulating morphogenesis and the HHIP locus may play a role lung development. Although there is evidence of association of CHRNA 3/5 locus with nicotine addiction, both this study and recent reports of a lung cancer association [12]–[14] with the same alleles suggest that this region may be involved in more than nicotine addiction, and may potentially have direct functional relevance in the development of COPD, lung cancer, peripheral arterial disease, and other smoking related conditions. The first-degree relatives of both lung-cancer patients and COPD patients have higher rates of impaired forced expiratory flow rates than relatives of patients with non-pulmonary disease [20], suggesting that susceptibility to lung cancer and COPD share common familial components. The association of CHRNA 3/5 locus with COPD, lung cancer, and peripheral arterial disease is powerful enough to make genetic screening of smokers an attractive interventional strategy.
Subjects from a case-control study [21],[22] recruited from Bergen, Norway were used as the discovery cohort in the GWAS. Baseline characteristics of the subjects are shown in Table 1. The entry criteria for COPD cases were post-bronchodilator forced expiratory volume in 1 second (FEV1) <80% predicted and FEV1/FVC (forced vital capacity) <0.7. The controls were selected based on post-bronchodilator FEV1 >80% predicted and FEV1/FVC >0.7. Individuals with Pi ZZ, ZNull, Null-Null or SZ α1-antitrypsin deficiency were excluded. Subjects with chronic pulmonary disorders other than COPD (e.g., lung cancer, sarcoidosis, active tuberculosis, and lung fibrosis) were also excluded. Because of the potential overlap in susceptibility genes for COPD and asthma, and the difficulty of diagnosing COPD vs. asthma in smokers with chronic airflow obstruction, previous asthma diagnosis was not used as an exclusion criterion. Both cases and controls were required to have a minimum of 2.5 pack-years of smoking. A total of 823 COPD cases and 810 controls were included in the present analysis. All of the subjects used in the primary and replication populations were current or former smokers (Table 1). Although the mean number of pack-years smoked was higher in cases (mean 32 SD 18) compared with controls (mean 19 SD 13), subjects with a range of smoking intensities were included in the analysis. The distribution of pack-years of smoking in cases and controls in the Bergen cohort is shown in Figure S3.
Subjects from the International COPD Genetics Network (ICGN) were used as the primary replication population. In the multi-center ICGN study [22],[23] subjects with known COPD were recruited as probands, and siblings and available parents were ascertained through the probands. Inclusion criteria for probands were post-bronchodilator FEV1<60% predicted and FEV1/VC <90% predicted at a relatively early age (45 to 65 years), a≥5 pack-year smoking history, and at least one eligible sibling with a≥5 pack-year smoking history. COPD in siblings was defined by a post-bronchodilator FEV1<80% predicted and FEV1/VC<90% predicted. The same exclusion criteria used in the Bergen study were also applied for the ICGN population. In total, 1891 Caucasian individuals from 606 pedigrees were included in the ICGN family-based association analysis.
The second replication cohort included 389 non-Hispanic white COPD cases from the U.S. National Emphysema Treatment Trial (NETT) [24] and 472 non-Hispanic white control subjects from the Normative Aging Study (NAS) [25]. Subjects in NETT had severe COPD (FEV1 ≤45% predicted) and bilateral emphysema on chest CT; all NETT subjects were former smokers. Control subjects from the NAS had normal spirometry and at least 10 pack-years of cigarette smoking history. Subjects from extended pedigrees in the Boston Early-Onset COPD (BEOCOPD) study were used as an additional family-based replication cohort. BEOCOPD subjects were recruited through COPD probands with age <53 years, FEV1 <40% predicted, and without severe α1-antitrypsin deficiency [26]. The BEOCOPD analysis included 949 individuals from 127 pedigrees.
Finally, to estimate allele frequencies in the general population in Norway we used 551 children (all non-smoking) from the Environment and Childhood Asthma (ECA) birth cohort study in Oslo [27].
All participants provided written informed consent and local institutional review boards approved the study protocols.
All samples in the Bergen discovery cohort were genotyped using Illumina's HumanHap550 genotyping BeadChip (version 3) which contains 561,466 single nucleotide polymorphisms (SNPs). All samples that had a call rate <98%, and all SNPs that had a call frequency <99% were deleted. This resulted in the elimination of 23,436 SNPs from further analysis (See Text S1 for more details). The ICGN subjects were genotyped using Sequenom's iPLEX SNP genotyping protocol developed for measurement with the MassARRAY mass spectrometer [28]. Genotyping in the NETT/NAS and BEOCOPD studies was performed using Sequenom iPLEX or Applied Biosystems TaqMan assays. Genotyping in the Norwegian ECA Birth cohort was done by TaqMan.
For the association analyses COPD affection status in the Norway discovery cohort, we used a logistic regression model to perform single-marker genotype trend tests for the QC-passed SNPs. To control for the possibility of spurious associations resulting from population stratification, we used a modified EIGENSTRAT method [29] (and Text S1). This showed that there were 12 significant principal component axes, all of which were included in the model. We included age and sex, and since smoking effects are known to influence COPD risk, we also included current smoking status and pack-years of smoking as co-variates.
The top 100 SNPs showing the lowest P values in this stage were selected for assessment in replication cohorts. For replication, we used a two stage strategy using three independent cohorts (Figure 1). In the first stage, family-based association analysis for COPD affection status was conducted in the ICGN data using PBAT version 3.6 [30]. Adjustments for age, gender, pack-years of smoking, current smoking status and center were performed in order to take into account the effect of smoking on the association results. Association with FEV1 was also tested using PBAT with age, gender, pack-years of smoking, current smoking status and height as co-variates. Gene-by-environment interaction analyses were also conducted using the PBAT program. Biallelic tests were conducted for SNPs using an additive genetic model. In the second stage the NETT case-control study was analyzed for the presence/absence of COPD using an additive genetic model. An unadjusted analysis and a logistic regression model adjusted for age and pack-years of smoking were conducted; sex was not included as a covariate because all NAS subjects were male, and current smoking was not included because all NETT subjects were ex-smokers. The BEOCOPD family-based study in the validation stage was analyzed using PBAT version 3.6 [30]. COPD was defined by post-bronchodilator FEV1/FVC<0.7 and FEV1<80% predicted (GOLD stage 2 or greater). Because a broad range of FEV1 values were included in the BEOCOPD study, we also analyzed FEV1 as a quantitative outcome in that population. Analysis of post-bronchodilator values of FEV1 was adjusted for ever-smoking status, pack-years of smoking, age, sex, and height.
We assessed genome-wide significance with a Bonferroni correction (p cutoff = 1.013×10−7 considering 493,609 independent tests in the Bergen cohort (see Text S1), 100 tests in the ICGN cohort, 7 tests in the NETT/NAS study and 6 tests in the BEOCOPD study (Total 493,772 tests).
|
10.1371/journal.ppat.1003464 | Structure and Function of a Fungal Adhesin that Binds Heparin and Mimics Thrombospondin-1 by Blocking T Cell Activation and Effector Function | Blastomyces adhesin-1 (BAD-1) is a 120-kD surface protein on B. dermatitidis yeast. We show here that BAD-1 contains 41 tandem repeats and that deleting even half of them impairs fungal pathogenicity. According to NMR, the repeats form tightly folded 17-amino acid loops constrained by a disulfide bond linking conserved cysteines. Each loop contains a highly conserved WxxWxxW motif found in thrombospondin-1 (TSP-1) type 1 heparin-binding repeats. BAD-1 binds heparin specifically and saturably, and is competitively inhibited by soluble heparin, but not related glycosaminoglycans. According to SPR analysis, the affinity of BAD-1 for heparin is 33 nM±14 nM. Putative heparin-binding motifs are found both at the N-terminus and within each tandem repeat loop. Like TSP-1, BAD-1 blocks activation of T cells in a manner requiring the heparan sulfate-modified surface molecule CD47, and impairs effector functions. The tandem repeats of BAD-1 thus confer pathogenicity, harbor motifs that bind heparin, and suppress T-cell activation via a CD47-dependent mechanism, mimicking mammalian TSP-1.
| Work on fungi is of worldwide importance due to the increasing burden of diseases caused by these agents in humans, plants and animals globally and throughout our ecosystem. The human pathogen Blastomyces dermatitidis harbors an essential virulence factor BAD-1. We describe new structural and functional features of BAD-1 that account for fungus' ability to cause disease. BAD-1 harbors a repetitive domain – a tandem repeat - that is present in up to 41 copies in the protein and essential in pathogenesis. We report an example of mimicry between this BAD-1 repeat and mammalian thrombospondin type-1 (TSP-1), a molecule that helps regulate human immunity. BAD-1 tandem repeats shares a tryptophan-rich sequence with TSP-1 that binds to tissue matrix and circulating immune cells. Like TSP-1, BAD-1 binds to this matrix. Through this property, BAD-1 binds and retards the action of white blood cells that fight fungal infection. Thus, we describe how fungi have evolved the means to mimic molecules that mammals commonly use to dampen unwanted immune responses, thus promoting pathogen survival.
| The dimorphic fungus Blastomyces dermatitidis is endemic to the Ohio and Mississippi river valleys, where it is the causative agent of blastomycosis. Blastomycosis is one of the principal systemic mycoses of humans and animals worldwide, and results from the inhalation of spores and/or hyphal fragments released into the air by this soil-dwelling fungus. Pulmonary infections that go undiagnosed or untreated may progress and disseminate, leading to substantial morbidity and mortality even in immunocompetent hosts.
Blastomyces adhesin-1 (BAD-1) is a 120-kDa protein of B. dermatitidis that mediates multiple functions including adhesion, modulation of pro-inflammatory immune responses and virulence [1], [2], [3]. A targeted deletion of BAD-1 attenuates pathogenicity in a murine model of pulmonary infection. BAD-1 expression is yeast-phase specific and is induced during the temperature-driven morphological transition of B. dermatitidis mold to yeast [1], [4]. Following secretion, BAD-1 coats the yeast and mediates binding of yeast to macrophages by CD11b/CD18 (CR3) and CD14 receptors [1]. This binding fosters entry into phagocytes [1], while inhibiting the release of pro-inflammatory cytokines such as TNF-α [2], [3]. Surface bound BAD-1 inhibits TNF-α release in a TGF-β dependent manner, while secreted BAD-1 does so in a manner that is independent of TGF-β [5].
BAD-1 is composed of a short N-terminal region that harbors a secretion signal, an extensive core of tandem repeats that are responsible for adhesion [6], and a C-terminal EGF-like domain that anchors the released protein on the yeast surface by binding chitin [7]. The number of tandem repeats varies between B. dermatitidis strains [8], [9], but they typically comprise over 80% of the protein's primary sequence. The repeats share 20 strongly conserved amino acids, including two cysteines postulated to define a loop structure via disulfide bonding [10]. The tandem repeats bind divalent cations including calcium, zinc and copper [10] (1∶1 stoichiometry) and calcium binding enables the C-terminal EGF domain to fasten itself to exposed yeast cell-wall chitin. The binding of divalent cations and sequence similarities to EF-hand domains present in thrombospondin-1 (TSP-1), which is a multi-functional extracellular matrix protein [10], [11], previously lead us to postulate that the BAD-1 tandem repeats might coordinate divalent cations, triggering a conformational shift [10]. While we observed changes in the peptide mapping patterns of BAD-1 in the presence of calcium, elevated divalent cations unexpectedly failed to impact the secondary structure of the molecule as measured by circular dichroism and tryptophan fluorescence spectroscopy [10]. Those findings prompted questions about the native structure of BAD-1.
We sought here to elucidate the 3-D structure of BAD-1 by NMR to gain deeper insight into the function of its tandem repeats in the pathogenesis of B. dermatitidis infection. NMR structural analysis demonstrated that the repeats adopt a tightly folded 17-amino acid loop conformation, constrained by a disulfide bond between conserved cysteines. We found no evidence for a conformational shift upon interaction of the tandem repeats with divalent cations, nor evidence for an EF-hand structure. Rather, each tandem repeat loop contains a conserved WxxWxxW motif found in TSP-1 type 1 heparin-binding repeats. BAD-1 was confirmed to bind to heparin specifically, saturably and with high affinity. A novel BAD-1 action involved TSP-1-like suppression of T lymphocyte activation and effector function in a manner similarly dependent on heparan sulfate decoration of CD47 on T cells. Our work sheds new light on the structure and TSP-1-like function of BAD-1 in virulence, and offers a striking example of molecular mimicry likely contributing to the pathogenesis of this fungal disease.
Our earlier descriptions of BAD-1 in B. dermatitidis ATCC strain 26199 identified 30 tandem repeats based on their adherence to a consensus sequence [6]. This criterion may have been unduly stringent. Reanalysis of conserved sequences in BAD-1 shows that the tandem repeat domain extends to within 17 residues of the N-terminus of the mature protein (Fig. 1A). This analysis identifies 41 tandem repeats based on conservation of the distance between cysteine pairs and the presence of strictly conserved histidine, tyrosine, and leucine residues (Fig. 1B). This contrasts with prior characterizations of the N-terminus, and emphasizes the extent to which the tandem repeats dominate the primary structure of the protein. Thus, BAD-1 may be characterized as having two principal domains: an exceptionally long tandem repeat domain and a chitin-binding, C-terminal EGF-like domain that fixes it to the yeast cell surface.
Since BAD-1 is an extended series of 41 tandem repeats with a short EGF-like C-terminal domain (Fig. 1), and since the C-terminal domain has proven to be dispensable for pathogenicity [12], we formally tested the role of the tandem repeats in virulence. To do so, we engineered a recombinant form of BAD-1 harboring only half the normal complement of tandem repeats and expressed this construct in a strain of B. dermatitidis (ATCC 26199) from which native BAD-1 had been deleted. Two independently engineered strains (TrepeatΔ20-Y and TrepeatΔ20-AE) were selected for their capacity to display amounts of surface BAD-1 similar to strains expressing the full-length protein (BAD1-6H-J and -AC) (Fig. S1). Each of these strains was compared in parallel for pathogenicity in a murine model of lethal pulmonary blastomycosis.
The transformed strains expressing the truncated forms of BAD-1 (TrepeatΔ20) were significantly less virulent than each strain expressing BAD-1 with the full complement of 41 repeats (Fig. 2). In contrast, the presence or absence of the C-terminal region (bearing all 41 repeats but no EGF-like C-terminal domain) had no significant impact on virulence in this model of infection as previously reported [12]. Thus, the tandem repeats of BAD-1 are required for pathogenicity in a murine model of lethal pulmonary infection.
Because of the functional significance of the tandem repeats, we sought structural insight into these domains in full length BAD-1 via NMR. The large size of the protein and peak overlap resulting from minor sequence variations between repeats made it difficult to elucidate the 3-D structure of the full-length protein. We therefore expressed a set of representative tandem repeats of identical sequence in E. coli (their sequence reflects the most prevalent amino acid in each position). The shortest recombinant protein that we could express in quantity contained four tandem repeats (TR4). Initially, TR4 displayed a random assortment of disulfide linkages, determined by variations in mobility by non-reducing PAGE (Fig. S2). To correct this, TR4 was reduced, associated with an NiNTA column and then slowly refolded under a glutathione gradient (see Methods). After refolding, variations in mobility resolved into a single, predominant band (Fig. S2). Tryptic digests of TR4 in both refolded and reduced states were examined by LC-MS. Digests of refolded TR4 lacked the reduced versions of cysteine-containing peptides, confirming that the cysteines in TR4 are fully disulfide-linked. This was corroborated in NMR studies (below).
15N HSQC NMR analysis of this refolded TR4 molecule produced a pattern of peaks amenable to interpretation, but which otherwise corresponded closely to the pattern of peaks derived from 15N HSQC of full-length, native BAD-1 (Fig. 3 A–C). Thus the tandem repeats in the TR4 recombinant protein successfully replicate the conformation of the native repeats, and that these repeats predominantly adopt one uniform conformation. Alternative conformations, if present, must be minimally represented, rendering their NMR signature(s) undetectable.
Many residues of the four identical repeats in TR4 have very similar chemical shifts, resulting in peaks that closely overlap in their NMR spectra, but the spectra did vary slightly between repeats. Furthermore, no evidence of interaction between repeat domains was found in the NMR spectra and the hinge regions between the tandem repeats were not resolvable via NMR (probably attributable to localized flexibility). These results suggest that TR4 does not adopt a unique tertiary fold in solution and precluded resolution of the molecule as a single homogenous structure. As an alternative, we calculated the 3-D structure for a single, representative BAD-1 repeat. To this end, the NOESY data derived from the residues of one tandem repeat were identified, compiled separately and submitted for algorithmic derivation of distance constraints and structural calculation. To accommodate inconsistencies in NOE peak intensities derived from partially overlapping peaks, NOE-derived distance restraints (used for automatic calibration by CYANA) were relaxed, marginally increasing average distance limits. Nevertheless, this approach yielded a consistent, tightly-folded 17 amino acid loop structure constrained at the base by a disulfide bond between the two conserved cysteines (Fig. 3D). The only identifiable secondary structure within this loop was in the WxxWxxW motif, which forms a short α-helix. Constraints are reported in Table 1.
The 3-D structure determined for the tandem repeat is inconsistent with the calcium-binding EF-hand structure previously hypothesized [10]. Every acidic residue is found on the external surface of the tandem repeat loop structure and not proximal to one another (Fig. 3E). Furthermore, the interior of the loop is occupied by aromatic residue side chains leaving little room for the pentagonal-bipyramidal calcium-coordination structure typical of an EF-hand. Thus, the high-capacity, low-affinity calcium-binding function of BAD-1 is not derived from an EF-hand-like structure in the tandem repeats. Alternatively, individual repeats could offer bi-dentate interactions with calcium, such that coordination of ions between repeat domains remains a possibility.
Figure 3F is a composite of the twenty best predictions available from CYANA and illustrates the consistency of this model with regard to the structure of the tandem repeat loop. Variability is predominantly seen in the orientation of surface-exposed side-chains interacting with the solvent environment.
Dynamic light scattering (DLS) data from two independent samples of BAD-1 showed polydispersity, but a relatively homogeneous population with a hydrodynamic radius of 7.2±0.5 nm accounted for ∼70% of the scattering intensity. This hydrodynamic radius contains contributions from the protein, bound water of hydration and shape factors (frictional coefficients) [13]. From the partial specific volume of BAD-1, the unhydrated molecule would have a spherical radius of 3.4 nm, which increases to 4.1 nm upon hydration. These values are much smaller than those measured by DLS. It would require a sphere comprised of 6 hydrated BAD-1 polypeptides as the diffusing complex to achieve the measured hydrodynamic radius. Sedimentation equilibrium studies showed BAD-1 to be monomeric. BAD-1 did not tend to oligomerize (up to at least 1.5 µM), suggesting that the large hydrodynamic radius is not due to formation of oligomers. We thus attribute the large measured radius to asymmetry in the shape of the molecule. The ratio of the measured radius to that of the sphere composed of a single hydrated BAD-1 molecule provides a quantitative estimate for this asymmetry, 1.8. The simplest models used to interpret hydrodynamic shapes are based on prolate and oblate ellipsoids of revolution. For an asymmetry of 1.8, a prolate ellipsoid would have semiaxes in the ratio of 15∶1∶1; an oblate ellipsoid would have axes in a ratio of 20∶20∶1. Our measurements suggest that the native conformation of BAD-1 occupies an expanded space - either a semi-flexible chain adopting multiple configurations or an extended rod-like structure.
NMR data suggests a 5 amino-acid flexible “hinge” between each tandem repeat “loop”, like beads on a string. Given the results of DLS, it is likely that BAD-1 is elongated with little in the way of tertiary structure. The hinge regions would afford limited flexibility. Energy minimization of the hinge regions of this model, constrained by the steric limitations of the tandem repeat “beads”, supports a (theoretical) extended, helical conformation (Fig. 3G).
Each tandem repeat loop of BAD-1 contains a significant number of tryptophans (4 of 24 amino acids - 17%). Because of this, BAD-1 absorbs UV wavelengths exceptionally well and may be detected readily by its characteristic OD280. A 1 mg/ml preparation of BAD-1 has an OD280 of over 6.6 [14]. The conserved arrangement of the tryptophans likewise stands out. With rare exception, three of the four tryptophans are arranged in a WxxWxxW motif. This motif is common to a number of glycosaminoglycan (GAG)-binding proteins, and is highly conserved within the type 1 heparin-binding repeats of TSP-1 [15].
We initially tested whether BAD-1 might bind GAGs in the extracellular matrix (ECM) by studying the adherence of yeast to Matrigel, which contains heparan sulfate and other ECM proteins such as laminin, collagen IV and nidogen. Blastomyces yeast bound Matrigel (Fig. 4A). Anti-BAD-1 antiserum blocked yeast binding to Matrigel (Fig. 4B), suggesting that the binding is BAD-1 dependent. Yeast did not reproducibly bind to the highly purified ECM components laminin, collagen IV or nidogen alone, implying that other Matrigel constituents such as heparan sulfate proteoglycan might be a target of BAD-1.
We investigated whether native BAD-1 protein could bind heparin. Initially, purified BAD-1 was incubated with a heparin-coated agarose resin. Due to the strong UV absorbance of BAD-1, the percentage of BAD-1 that bound to resin could be assayed spectrophotometrically by measuring the OD280 of the aqueous phase before and after incubation. Heparin-agarose resin pulled BAD-1 (100 µl of 0.1 mg/ml) out of solution, with a 20 µl volume of resin absorbing nearly 100% of the soluble BAD-1. Binding was dependent upon resin volume (0.5–1.5 µg BAD-1/µl resin) and saturable (Fig. 4C). While binding was maximal at lower ionic strengths (≤50 mM NaCl), it was still appreciable at the established ionic strength of alveolar fluid (100 mM NaCl) and plasma (150 mM NaCl) [16], [17]. In subsequent assays, we fluorescently labeled BAD-1 (with eFluor605) to quantify binding via fluorescence spectroscopy. Labeling did not alter its binding characteristics since assays with unlabeled protein gave similar results (Fig. S3). Fluorescent BAD-1 bound avidly to heparin-agarose, but not to controls of unmodified agarose resin or resins coated with mannan, BSA or hemoglobin (Fig. 4D). We included the two latter controls since BAD-1 [10] and heparin [18] both coordinate polyvalent cations and have the potential for non-specific association via polyvalent cation bridging [19]. BSA [20] and hemoglobin [21] also coordinate multiple polyvalent cations. Nevertheless, BAD-1 showed little affinity for these control resins or for the carbohydrate-rich mannan-agarose resin.
We assessed the specificity of the interaction between BAD-1 and immobilized heparin using soluble competitors (Fig. 4E). Pre-incubation of BAD-1 with soluble heparin diminished its binding to agarose-immobilized heparin in a concentration-dependent manner. Closely related GAGs, including chondroitin sulfate A and hyaluronan, did not significantly inhibit the binding of BAD-1 to immobilized heparin (Fig. 4F). Dermatan sulfate, also called chondroitin sulfate B, inhibited only at the highest concentrations.
To measure the affinity of BAD-1 for heparin, biotinylated heparin was immobilized to several densities on a neutravidin NLC chip. Dose-response curves were generated with BAD-1 in concentrations ranging from 1.5 µM to 94 nM. Figure 5A shows the results for two low-density heparin surfaces. A striking feature is the slow dissociation of BAD-1 once it has bound to the surface. This likely reflects the multivalency of BAD-1. Figure 5B shows binding of 0.375 µM BAD-1 in the absence of heparin or mixed with 3.75 µM heparin. Binding to immobilized heparin is completely blocked by a 10-fold excess of free heparin. These two figures establish the specific binding of BAD-1 to the heparin surface.
The kinetic constants for the dose-response data above were calculated using the Langmuir 1∶1 binding model in the instrument software. For the curves shown in figure 5A, the on and off rates were fit to each sensorgram separately, but the maximum response level was fit as a single value for each heparin density. The on-rate was 5800±2000/M/s with an off-rate of 1.7±0.3 e-4/s. Each pair of rate constants was used to compute an equilibrium dissociation constant, or affinity, of 33±14 nM, which closely approximates that reported for TSP-1 (80 nM) [22]. Thus, BAD-1 binds to immobilized heparin in a concentration-dependent manner, which is inhibitable by free heparin. Once bound, BAD-1 dissociates from the heparin surface slowly, demonstrating the kind of high avidity/moderate affinity interaction typical of GAG-binding proteins with multiple binding sites [23].
SPR analysis of the binding of truncated Trepeat20 BAD-1 to heparin showed little variation in affinity compared to full length BAD-1 (Fig. S4). The reduced virulence observed for strains producing this truncated adhesin is thus not explained by a diminished affinity or avidity for heparin, but may hinge on some other factor(s). Nevertheless, reduced length is a feature known to impact the function of other adhesins [24]. Alternatively, if individual (free) BAD-1 repeats were to mediate in vivo effects on interaction with the immune system (below), the Trepeat20 strain has only half the molar equivalent of repeats of the parental strain.
The interaction between heparin and TSP-1 involves an ionic component and may be inhibited by concentrations of NaCl above 350 mM [22]. We observed that NaCl could similarly inhibit the binding of BAD-1 to heparin-agarose (Fig. S5A). Binding of BAD-1 to heparin diminished sharply at 250 mM NaCl and above, suggesting that this interaction involves an ionic component. The binding is maximal at low pH, but falls off substantially above pH 8 (Fig. S5B), indicating that protonated histidine residues may be integral to the binding site
The interaction between TSP-1 and heparin is inhibited by peptides containing a WxxW motif [15]; for example a peptide with the sequence SHWSPWSS. We used this peptide (and a control, mutant peptide with tryptophan residues replaced by glutamine) to test whether BAD-1 binds to the same site on heparin. The WxxW peptide failed to inhibit, but instead augmented BAD-1 binding to heparin agarose (Fig. 6A), and control peptide (SHQSPQSS) had no effect on binding. Soluble heparin blocked 65% of BAD-1 binding to heparin agarose. Thus, the initial binding of BAD-1 to heparin could be facilitated by a site outside the tandem repeat (see below).
We also tested whether TR4 could be used to block BAD-1 binding to heparin agarose, as each repeat bears a WxxWxxW motif. Native TR4, however, bound poorly to immobilized heparin (Fig. 6B). BAD-1 lacking a C-terminal EGF domain (ΔC-term) did bind heparin, which implied that this domain does not mediate heparin binding. TR4 could only be induced to bind heparin by reducing the peptide's disulfide bonds with DTT and allowing re-oxidation in the presence of heparin. This maneuver also enhanced heparin binding by BAD-1 and ΔC-term protein, but when these proteins were maintained in a reduced state (DTT≥1 mM) the capacity to bind heparin was lost (data not shown), suggesting that primary sequence alone is insufficient to drive the heparin interaction and that a novel secondary structure must form to bind heparin.
Once we established that DTT-unfolded TR4 binds heparin agarose upon oxidation, we investigated whether this domain – harboring a WxxWxxW motif – could interfere with BAD-1 binding to heparin. As with the synthetic WxxW peptide, reduced TR4 failed to inhibit binding and instead augmented BAD-1 binding to heparin agarose in a concentration-dependent manner (Fig. 6C). This result suggests that the WxxWxxW motif in the tandem repeats is not enough, by itself, to foster initial interaction with heparin. An alternate possibility is that it may be half of the two-component heparin-binding cleft described by Cardin-Weintraub [25]. Thus, instead of inhibiting binding, tryptophan-containing peptides might pair with stretches of basic residues to stimulate binding. In fact, examination of the N-terminus of BAD-1 reveals an xBBxBx Cardin-Weintraub heparin-binding motif [26] (B = basic residue, x = any residue) within this short stretch of amino acids (Fig. 1A). This known heparin-binding motif could be involved in the initial engagement of heparin (or similar polysulfated GAG) by BAD-1.
Molecules with heparin-binding motifs can modulate the activation of immune cells such as T cells via interaction with GAG-modified cell surface proteins. The heparin-binding protein, TSP-1, inhibits the activation of T cells via interaction with the surface protein CD47 through a GAG-modified serine at position 64 [27]. While the C-terminal TSP-1 domain is sufficient for this interaction [28], we hypothesized that BAD-1 might also block T cell activation in a CD47 dependent manner. Anti-CD3 antibody activation of Jurkat T cells or CD47-deficient JinB8 T cells resulted in a ∼25-fold increase in CD69 expression at 2 hours. As reported previously [27], TSP-1 decreased CD69 expression in activated Jurkat T cells by 45% (Fig. 7A), while failing to inhibit the activation of JinB8 T cells (Fig. 7B). BAD-1 also sharply reduced CD69 expression in activated Jurkat T cells by 65%, but failed to block activation of CD47-deficient JinB8 T cells.
We next tested the role of GAG modification of CD47 in BAD-1 suppression of T cell activation. We transfected CD47-deficient JinB8 T cells with a plasmid encoding wild-type CD47 or a mutant CD47 containing a serine to alanine substitution that precludes GAG-modification (CD47-S64A). Transfections led to re-expression of CD47 on JinB8 cells. Re-expression of wild-type CD47, but not CD47-S64A, in JinB8 cells enabled TSP-1 to significantly inhibit CD69 induction by 35.1% (Fig. 7C). BAD-1 also significantly inhibited CD69 induction in JinB8 cells transfected with wild-type CD47 (34.9%), but not in JinB8 cell transfected with CD47-S64A. BAD-1, like TSP-1, thus modulates T cells activation via surface protein CD47 and the inhibitory action of BAD-1 requires CD47 GAG-modification at Serine 64.
We next tested the impact of BAD-1 on the function of primary CD4+ T cells that respond to Blastomcyes in an antigen-specific manner and mediate immunity during infection [29]. 1807 TCR transgenic mice produce such Blastomyces reactive CD4+ T cells [30]. We studied naïve CD4+ T cells from these mice, analyzing their ability to become activated and differentiate into cytokine producing cells in vitro in response to co-culture with BAD-1 null Blastomyces yeast and dendritic cells (DC). Upon culture with yeast, 1807 cells became activated as measured by CD69 expression in ∼30% of the cells (Fig. 7D). 1807 cells also responded to the fungus by producing IL-17 and IFN-γ (Fig. 7F). The addition of exogenous BAD-1 to 1807 cells curtailed their activation and expression of CD69 in response to co-culture with yeast and DC (Fig. 7D), and was associated with binding to the T cell surface (Fig. 7E). In addition to CD69, several other markers of T cell activation, including CD25 (IL-2R), CD44 and CD62L, showed that 1807 responses to antigen were similarly blunted by BAD-1 (Fig. 7E). BAD-1 also suppressed 1807 cell production of IL-17 and IFN-γ in a concentration dependent manner (Fig. 7F). Thus, BAD-1 mimics TSP-1 and is capable of suppressing activation and effector functions of T cells, in this case CD4+ T cells that confer resistance to infection.
Herein, we describe novel structural and functional properties of BAD-1, an essential virulence factor of B. dermatitidis. We establish that the tandem repeats are indispensible for the role of BAD-1 in virulence, provide NMR based 3-D structural characterization of the BAD-1 tandem repeats, and identify a tryptophan-rich motif in the tandem repeats involved in heparin binding. We further establish that BAD-1 suppresses T cell function via interaction with CD47, thus mimicking TSP-1. Though initial analyses of the primary structure of BAD-1 reported 30 highly conserved tandem repeats [6], a more in-depth assessment of stringently conserved elements adds 11 more repeats to this total. This new model characterizes BAD-1 as essentially an extended series of tandem repeats (comprising >80% of the protein's length) with a C-terminal, chitin-binding domain to anchor it to the surface of B. dermatitidis yeast and a predicted N-terminal Cardin-Weintraub heparin-binding site [26]. Because the C-terminus is dispensable for virulence in a murine model of pulmonary infection [12], we postulated that virulence must depend upon the tandem repeats. Indeed, we found that even partial deletion of the repeats attenuates pathogenicity mediated by BAD-1.
The tandem repeats of BAD-1 contribute to its adhesive functions [1], [6]. Pathogen surface adhesins often contain tandem repeat domains. In some cases, the tandem repeats themselves mediate host-cell surface adhesion (e.g. Staphylococcus aureus MSCRAMM) [31]. In other cases, the repeats act as “spacer arms” that orient and present a binding domain to host receptors. Upon engagement of host ligands, certain adhesins are “dynamic”. For example, Als5p of Candida albicans reacts to mechanochemical pressure as it is stretched, evolving new conformations that propagate Als5p adhesive nanodomains on the fungal surface, while also drawing the host and pathogen together [32], [33]. By structural analysis, the repeats of BAD-1 are linked together by non-rigid “hinge” regions, like beads on a string. DLS analysis predicts an elongated conformation for BAD-1, arguing for a degree of flexibility in these hinge regions. If so, then a structure similar to the rod-like adhesins Als5p of C. albicans [34] and invasin of Yersinia sp. [35] is plausible as proposed in figure 3G.
Because the tandem repeats are necessary for virulence, we sought further insight into their structure and function. One notable aspect of their sequence is an exceptionally high tryptophan content. In an average protein perhaps one residue in a hundred will be tryptophan, but BAD-1 surpasses this ratio 17-fold. In nearly every repeat, three tryptophans are arranged in a highly conserved WxxWxxW pattern, a motif proven to mediate heparin binding by TSP-1. Short peptides bearing this motif are capable of binding to GAGs [22], and the presence of adjacent basic sequences enhances both affinity and specificity for heparin. In TSP-1, tryptophans are arrayed along a short, α-helical domain, coordinating with basic residues on an adjacent anti-parallel strand to create a surface-exposed recognition groove. The basic residues intercalate between the tryptophans, thus sharing their pi-orbital electrons in a configuration that is key for heparin association [36]. NMR analysis of the BAD-1 tandem repeat model - TR4 - demonstrated that the disulfide bond present within each repeat constrains the 17 residues between them into a consistent, tightly folded loop. While this loop localizes basic residues adjacent to the tryptophans of the WxxWxxW motif, one fold of this loop stretches transversely across these putative active residues in a manner that appears to interdict surface-exposure. In this way, the 3-D structure of TR4 contrasts with the heparin-binding motif of TSP-1. Yet, strong parallels between BAD-1 and TSP-1 prompted us to explore heparin-binding activity.
Our work discloses a previously unrecognized capacity of BAD-1 to bind heparin. The activity is saturable, specific and high-affinity, and constitutes an advance in our understanding of BAD-1 and B. dermatitidis pathogenesis. BAD-1 is known to alter host innate immune responses, suppressing TNF-α [2] and inducing TGF-β production [5]. We now show that BAD-1 suppresses T lymphocyte receptor signaling in a manner similar to that reported for TSP-1 [27]. TSP-1 binds to a heparan sulfate-modified serine residue of CD47, suppressing T cell activation. We do not provide direct biochemical evidence that BAD-1 binds CD-47, but our data show that BAD-1 similarly suppresses T cell function in a CD-47 dependent manner. This BAD-1 mimicry of TSP-1 resulted in impaired T cell activation and differentiation, with reduced production of effector cytokines including IL-17 and IFN-γ. Since T cell activation is vital to the host's ability to clear yeast via adaptive immunity [29], inhibition of T cell function could foster pathogen survival and immune evasion. TSP-1 also regulates immune-tolerance by phagocytic cells [37], nitric oxide signaling [38], activation of TGF-β [39], [40] and binding and clearance of matrix metalloproteinases involved in the healthy egress of lung inflammatory cells [41]. The ability of BAD-1 to mimic TSP-1 and modulate host immunity to its advantage could hinge on any or all of these activities.
Despite the evolutionary similarity between fungi and mammals, molecular mimicry is infrequently reported in pathogenic fungi. Candida albicans expresses an integrin-like protein Int1p [42] that binds yeast to vascular endothelium and promotes filamentation and virulence [43], [44]. Histoplasma capsulatum CBP1, a virulence factor, is similar in its 3-D NMR structure to mammalian saposin B [45], but evidence of saposin-like interactions with host glycolipid has not yet been reported. We describe a striking example of molecular mimicry involving sequences in the BAD-1 tandem repeat that mimic those in mammalian TSP-1 structurally and functionally, and confer pathogen survival.
There are additional implications stemming from the finding that BAD-1 binds heparin. Binding of mammalian cell surface GAGs is a mechanism used by many pathogens. Viruses, bacteria, and parasites exploit host cell-surface GAGs, mediating attachment with adhesins to impede clearance [46], [47]. The BAD-1 adhesin similarly binds yeast to lung tissue [48], macrophages [1], [6], [12] and ECM. While we do not provide evidence here that the adhesive features of BAD-1 are due to its affinity for heparin, parallels with other pathogens make this idea plausible.
Our observation that the TR4 repeats did not share the heparin-binding activity of BAD-1 was at first surprising because NMR analysis established structural identity between TR4 and the native BAD-1 repeats. Perhaps the need for an activation trigger in a critical pathogenicity factor like BAD-1 makes sense. Forty-one tandem repeats constitute enormous potential avidity, and non-specific adherence may disadvantage the pathogen. Given that a segment of the repeat loop lies across the putative heparin-binding motif, it was perhaps to be expected that our “model” tandem repeat would bind heparin poorly. Our observation that relaxing TR4's structure via disulfide reduction brings its heparin-binding up to parity with BAD-1 suggests that a reconfiguration of the loop structure may be requisite for heparin binding. Enzymatic reduction of disulfide bonds in the extra-cellular environment is one means of achieving this [49]. Plasmin is activated this way [50], and gp120 of HIV-1 requires the reduction of two disulfide bonds by cell-surface protein disulfide isomerase (PDI) to ligate lymphocyte receptors [51]. Importantly, reductases like PDI are apt to collect on cell surfaces enriched with GAGs [51].
There are other means of exchanging disulfide bonds. Cell-surface adhesins exposed to significant mechanochemical stress (stretching) are subject to accelerated disulfide cleavage and reorganization after initial ligation of target molecules [33], [52], [53], [54]. We hypothesize that BAD-1 associates initially with host cell-surface GAGs through its N-terminal Cardin-Weintraub domain [26], and that subsequent reorganization of its disulfide structure, either catalyzed or spontaneous, permits the tandem repeats to participate in heparin binding. This model would be expected to magnify the strength of the interaction (consistent with the slow dissociation rate of BAD-1 and heparin observed via SPR) and effectively draw the host and pathogen together as additional repeats are engaged, essentially “zippering” BAD-1 and yeast down onto heparin.
Our effort to solve the NMR structure of TR4 complexed with heparin was unsuccessful. The 3-D structure of the tandem repeat motif that binds heparin therefore remains uncertain and is the subject of ongoing work. Figure 8 depicts a theoretical configuration of the tandem repeat structure (8B), showing the alignment of its residues in accordance with the heparin-binding motif of the thrombospondin-related anonymous protein (TRAP) of the malaria parasite (8A). In the “proximal” model (Fig. 8B and Fig. 8C, left image), the repeats would form a regular, anti-parallel β-sheet conformation. Alternative structures are possible including a so-called “distal” model also forming a β-sheet (Fig. 8C, middle) and a “hairpin” model (Fig. 8C); the latter two are rendered in 3-D in Figure S6.
In conclusion, we describe here structural features of BAD-1 tandem repeats, as determined by NMR, and a novel heparin-binding function associated with this essential virulence domain. This activity may govern B. dermatitidis yeast adherence to host cells and ECM via binding to heparan sulfate. In binding heparin, BAD-1 mimics TSP-1 in similarly down-regulating the activation of T cells through its interaction with a GAG-modified serine of CD47. These findings shed new light on the structure of BAD-1, its diverse functions, and novel mechanisms through which it may promote fungal pathogenicity. BAD-1 is a rare, but stunning example of molecular mimicry among fungal pathogens.
All animal procedures were performed in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. Care was taken to minimize animal suffering. The work was done with the approval of the IACUC of the University of Wisconsin-Madison.
Complete Mini protease inhibitor tablets (EDTA free) were from Roche (Indianapolis, IN). Heparin was from Sagant pharmaceuticals (Schaumberg, IL). Unless noted otherwise, chemicals were from Sigma.
American Type culture Collection (ATCC) strain 26199 of B. dermatitidis, a wild-type, virulent isolate originally obtained from a human patient, was used in this study, together with the isogenic attenuated, BAD-1 knockout strain #55 [48]. Truncated chimeras of the BAD-1 gene were used to transform strain #55 to secrete a full length BAD-1 with a 6-his tag (BAD1-6H), BAD-1 lacking the C-terminal region (ΔCterm as previously described) [12], and BAD-1 lacking 20 of the tandem repeats (Trepeat20)(described below). All isolates of B. dermatitidis were maintained in the yeast form on Middlebrook 7H10 agar slants with oleic acid-albumin complex, grown at 39°C. Liquid cultures of yeast were grown in Histoplasma macrophage media (HMM) [14].
An expression cassette for a truncated form of BAD-1 in which 20 of the tandem repeats were deleted was created by digesting the deletion construct pBAD1-6H [12] with BamH1 restriction enzyme (NEB Biolabs, Ipswich, MA) and then re-ligating, removing 1355 bp of the original cDNA. The deletion construct was excised from the pUC18 vector in an EcoR1/Xba1 digest, and then inserted into the EcoR1/Xba1 sites in the polylinker of plasmid pCB1532 (a vector carrying the Sulphonyl Urea resistance [SUR] gene of Magneportha grisea, generously provided by Dr. James Sweigard [Dupont, Wilmington, DE]) [55]. This plasmid was used to transform BAD-1 knockout strain #55 and producing strains were identified by Western blotting nitrocellulose overlays placed on replica plates of picked transformants, as previously described [48]. Strains were selected that most closely reconstituted the BAD-1 production seen in the 26199 parental strain. Production levels were estimated by Western blot of surface extracted protein probed with anti-BAD-1 mAb DD5-CB4 [1], [6] followed by goat anti-mouse (GAM) IgG-alkaline phosphatase (Promega). Production levels were further quantified using a FACscan flow cytometer (Becton Dickenson) using DD5-CB4 and GAM-FITC (Sigma) (Fig. S1).
Male BALB/c mice ∼5–6 wk of age (Harlan Sprague Dawley) were infected intra-nasally with B. dermatitidis yeasts as previously described [56]. In brief, mice were anesthetized with inhaled Metafane (Mallinckrodt Veterinary Inc.). A 25-µl suspension of yeast cells in PBS was then applied drop-wise into their nares. To insure a lethal infection was established, 104 yeast were thus administered. All animal procedures were performed in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. Care was taken to minimize animal suffering. The work was done with the approval of the IACUC of the University of Wisconsin-Madison.
Complementary oligos RS33 and RS34 were annealed and ligated into pUC18 digested with BamH1/EcoR1 to make pUC18/33-34.
(RS33-GATCCGAAGACGACCCTACAACTGTGACTGGGACAAGTCCCATGAGAAGTATGATTGGGAGCTCTGGGATAAGTGGTGCAAGGACG,
RS34-AATTCGTCCTTGCACCACTTATCCCAGAGCTCCCAATCATACTTCTCATGGGACTTGTCCCAGTCACAGTTGTAGGGTCGTCTTCG)
pUC18/33-34 contains DNA coding for one tandem repeat, with a Bbs I site and a BamHI site just upstream. The annealed complementary oligos RS35 and RS36 were ligated into these sites, creating a new set of BbsI/BamHI sites for the next digestion/ligation cycle. After each cycle, the newly cloned plasmid was opened by Bbs I/BamHI and purified from an agarose gel (Freeze and Squeeze DNA extraction spin columns, Bio-Rad).
(RS35- GATCCGAAGACGACCCTACAACTGTGACTGGGACAGTCCCATGAGAAGTATGATTGGGAACTCTGGGATAAGTGGTGCAAGGAC,
RS36- AGGGGTCCTTGCACCACTTATCCCAGAGTTCCCAATCATACTTCTCATGGGACTTGTCCCAGTCACAGTTGTAGGGTCGTCTTCG)
This process was repeated until the plasmid contained four repeats. The final construct was cut out of pUC18 with BamH1/EcoRI (EcoR1 site blunted) and ligated into pQE32 cut with BamH1/HindIII (HindIII site blunted) creating plasmid pQETR4 for expression of the 4-repeat model protein in E. coli with a 6-histidine tag (TR4).
TR4 (Fig. 1C)
MRGSHHHHHHGIRRRPYNCDWDKSHEKYDWELWDKWCKDPYNCDWDKSHEKYDWELWDKWCKDPYNCDWDKSHEKYDWELWDKWCKDPYNCDWDKSHEKYDWELWDKWCKDELA
TR4 was expressed in E.coli (grown at 30°C in LB medium) by induction with IPTG and isolated from cell lysates using an NiNTA column (Qiagen, Valencia, CA). Protein was refolded while immobilized on NiNTA resin by subjecting it to a gradient from 100% buffer A (6M urea, 10 mM hepes, pH 8, 300 mM NaCl, 10% glycerol, 2 mM mercaptoethanol, 1 mM CaCl2) to 100% buffer B (10 mM hepes, pH 8, 300 mM NaCl, 10% glycerol, 5 mM 4∶1 GSH∶GSSG, 1 mM CaCl2) over the course of three hours. Refolded protein was then eluted with 250 mM Imidizole, which was removed by dialysis.
“In Liquid” digestion and mass spectrometric analysis was done at the Mass Spectrometry Facility (Biotechnology Center, University of Wisconsin-Madison). In short, 5 µg of purified protein in 125 mM NH4HCO3 (pH 8.5) was reduced with DTT (62.5 mM final) for 30 minutes at 55°C. Another 5 µg sample was left untreated as a control. Samples were spun through Pierce detergent removal columns (Thermo Scientific) to remove DTT and subsequently digested with trypsin solution (Trypsin Gold from Promega Corp.). Peptides were loaded on LC/MSD TOF (Agilent Technologies) and analyzed by MALDI TOF/TOF (AB SCIEX). (Additional detail available in supplementary data- Materials and Methods)
The hydrodynamic radius of BAD-1 was measured by dynamic light scattering (DLS) collected at the 90 degree angle using a Beckman-Coulter N4 Plus instrument with the sampling time and prescaling as optimized by the instrument. Two samples of BAD-1 at 7 µM in 70 mM NaCl, 40 mM Tricine pH 7.0 were measured. For each sample ten autocorrelation functions were recorded and analyzed using both the unimodal and the size distribution software supplied with the instrument. The averages reported exclude repetitions with >1% baseline error in unimodal analysis or >5% dust fraction in the size distribution analysis. The refractive index of water at 15°C, 1.333, from instrument's database was used. The contribution of 40 mM tricine to the solvent viscosity was approximated by linear interpolation between increments reported for 20 and 100 mM tricine [57]. This contribution was added to the viscosity computed for 70 mM NaCl at 25°C, and then corrected to 15°C assuming its behavior paralleled that of water [58] for a value of 1.41 cP. The partial specific volume of BAD-1 was computed based on the sequence to be 0.704 mL/g. Hydration of 0.469 g water/g polypeptide at pH 7 was calculated based on the amino acid composition [59]. Additional detail is available in Supplementary Methods.
Three samples of BAD-1 prepared by serial dilution were analyzed by sedimentation equilibrium to ascertain the association state. The buffer was 10 mM sodium phosphate, 100 mM NaCl at pH 7.6 with a computed density of 1.005 g/mL [47] and a partial specific volume of 0.704 mL/g. Equilibrium data at 4°C were collected using a Beckman analytical ultracentrifuge. Gradients were monitored at 280 nm and equilibrium data was recorded at speeds of 5600, 7800, 9200 and 12000 rpm. Analysis utilized software written in IGOR Pro (Wavemetrics, Inc.) by D. R. McCaslin. Additional detail is available in Supplementary Methods.
BAD-1 was purified as described [10] with a modification. Yeast was grown in liquid HMM in a gyratory shaker at 37°C for 5 days. Yeast was pelleted and washed once in PBS, and BAD-1 was released from cell surfaces with three 1-hour washes in dH2O and then purified on a metal-chelate resin (NiNTA). Due to its divalent cation-binding property, BAD-1 protein could be purified on NiNTA resin regardless of whether it included a 6-histidine tag. The stringency of the wash buffer was reduced by eliminating imidazole and reducing NaCl to 150 mM. Homogeneity of purified BAD-1 was analyzed by SDS-PAGE, Sypro Ruby stain (Invitrogen), and Western blot using anti-BAD-1 antibody (DD5-CB4) [1], [6], [60].
Matrigel (Collaborative Biomedical Products, Bedford, MA) was diluted to 5 mg/ml, 0.5 mg/ml and 0.05 mg/ml in RPMI. 30 µl was put into wells of a 96-well plate and allowed to gel at 37°C overnight. Blastomyces yeast were labeled with Na51CrO4 for 105 minutes at 37°C, washed and diluted to 2×107/ml in HBSS+0.1%BSA. 50 µl was added to each well and incubated for 60 minutes, then washed with HBSS. Well contents were subjected to scintillation counting (compared to controls with known numbers of labeled yeast) to quantify bound yeast. To block BAD-1-mediated binding, rabbit anti-BAD-1 immune serum [8] (or control pre-immune serum) was applied to labeled yeast, incubated for 1 hour, and washed with HBSS before binding assays.
Heparin-agarose resin was obtained from Sigma as were control agarose beads and agarose beads coated with BSA, hemoglobin, and mannan. 100 µl of 0.1 mg/ml BAD-1 was incubated with agarose resins (5 µl bed volume) in 20 mM tricine, pH 7, 50 mM NaCl for 30 min at 25°C with agitation. Resin beads were pelleted by centrifugation in a microfuge at 7000xG. Concentration of BAD-1 before and after incubation with resin was monitored by A280 via Nanodrop Spectrophotometer (ND1000, Thermo Scientific). Binding was calculated by comparing the A280 of supernates to that of starting material. Binding inhibition studies were done with soluble medical grade heparin purchased from Elkins-Sinn Inc (Cherry Hill, NJ), dermatan sulfate (chondroitin sulfate B)(Sigma), chondroitin sulfate A (chondroitin-4-sulfate, fraction A)(Sigma) and hyaluronan (Sigma). Baseline absorbance was corrected to account for absorbance of added GAG inhibitors. During optimization studies, heparin resin bed volume was varied from 1 µl to 20 µl, NaCl concentrations were varied from 40 mM to 2000 mM and alternative buffers were tested (20 mM Na-acetate, pH 5, 20 mM Tricine pH 7 and pH 8, 20 mM Na-carbonate, pH 9). Reduction of BAD-1 and TR4 was accomplished with 10 mM DTT and heating at 100°C for three minutes. Re-oxidation of thiols was accomplished by diluting or dialyzing away DTT followed by exposure to air at room temperature.
BAD-1 binding to heparin and control resins, and binding inhibition studies also were performed using BAD-1 labeled with the eFluor605NC kit from eBioscience (San Diego, CA). BAD-1 was fluorescently labeled following the manufacturer's instructions. BAD-1 (eFluor605) was incubated with heparin-coated agarose beads or control beads in 20 mM tricine buffer, pH 7, 50 mM NaCl and washed three times with the same buffer before quantification. Association of BAD-1 (eFluor605) with resins was verified visually using an Olympus BX60 fluorescent microscope. For binding and inhibition studies, BAD-1 (eFluor605) was quantitated using a FilterMax F5 multi-mode microplate reader (Molecular Devices, Sunnyvale, CA) and opaque 96 well plates (Costar, Cambridge, MA). BAD-1 binding to heparin also was reproduced at higher ionic strengths of 100 mM NaCl and 150 mM NaCl, as observed in alveolar mucus and plasma, respectively [16], [17].
E. coli strain XL-1 Blue transformed with pQE-TR4 was grown in defined M9 medium supplemented with 15N ammonium chloride (1 g/L) and 13C dextrose (4 g/L) (Cambridge Isotope Laboratories Inc., Andover, MA) for 20 hrs at 30°C, under ampicillin selection (100 µg/ml) followed by induction with IPTG and 4 hours of further incubation. [13C,15N]-labeled TR4 was purified and refolded as described above. 15N labeled native BAD-1 was produced by growing 26199 B. dermatitidis yeast at 37°C for five days in M9 medium supplemented with 15N ammonium chloride (1 g/L) and purifying as described above. Purity of isolated proteins was verified by PAGE prior to NMR analysis.
All NMR spectra were recorded at the National Magnetic Resonance Facility at Madison (NMRFAM) on Varian VNMRS (600 MHz and 900 MHz) spectrometers equipped with triple-resonance cryogenic probes. The temperature of the sample was regulated at 25°C for experiments. For sequence specific backbone resonance assignments, a series of two-dimensional (2D) and three-dimensional (3D) heteronuclear NMR spectra were collected on a sample containing 0.1 mM of the [13C,15N]-labeled TR4 dissolved in NMR buffer with 10 mM phosphate, pH 8.0, 95% H2O, 5% D2O [61]. Raw NMR data was processed with NMRPipe [62] and analyzed using the program Sparky [63]. 2D 1H-15N HSQC and 3D HNCO data sets were used to identify the number of spin systems, and these identifications plus 3D HNCACB and 3D CBCA(CO)NH data sets were used as input to the PINE server to determine sequence specific backbone resonance assignments [64]. Due to the complexity of the system, the automated backbone resonance assignments needed to be refined manually with the help of a 3D 15N-edited 1H-1H NOESY spectrum. To assign side chain and HB and HA resonances 2D 1H-13C aliphatic HSQC, 3D HBHA(CO)NH, 3D HC(CO)NH, 3D C(CO)NH and 3D H(C)CH TOCSY experiments were used. Furthermore, a 2D 1H-13C aromatic HSQC spectrum together with a 3D 13C aromatic-edited 1H-1H NOESY were used to assign resonances from aromatic side chains. Finally, a 3D 15N-edited 1H-1H NOESY (100 ms mixing time) spectrum, a 3D 13C aliphatic-edited 1H-1H NOESY (100 ms) spectrum and a 3D 13C aromatic-edited 1H-1H NOESY (100 ms) spectrum were acquired and used to derive the distance constraints to determine the three dimensional structure of the protein.
15N resolved 1H-1H 3D NOESY and 13C resolved 1H-1H 3D NOESY spectra were used to derive 1H-1H distance restraints. Backbone dihedral angle restraints φ and ψ were obtained from 1H, 15N, 13CA, 13CB, 13C′ using TALOS+ software [65]. CYANA software version 3.0 was used for automated NOESY peaks assignments and structure calculation following the standard simulated annealing protocol [66]. The program PYMOL (Schrödinger sales) was used to calculate the root mean square deviation (rmsd) and for graphical analysis. The PSVS server was used to check the quality of the structures [67].
A Biorad Proteon XPR36 instrument was used for surface plasmon resonance studies. To prepare a heparin surface, medical grade heparin was biotinylated with EZ-link Biocytin Hydrazide (Thermo Scientific) and 1-Ethyl-3-(3-dimethylaminopropyl) carbodiimide (EDAC) following manufacturer's instructions. Unreacted biotin was removed by precipitating biotinylated heparin with 66% methanol, which was then applied in binding buffer (50 mM Hepes, pH 7, 70 mM NaCl, 0.1% Tween-20) (0.1 mg/ml) to a Biorad neutravidin (Proteon Sensor Chip NLC) according to the manufacturer's protocol. Subsequently, single and multiple injections of varying lengths of time and with varied concentrations of biotinylated-heparin were made onto chip in its vertical orientation. After immobilization, the surface was washed with several injections of 1 M NaCl in 50 mM NaOH. Final immobilization levels of biotinylated heparin were: 0, 5, 30, 59 and 96 RU. Analysis of BAD-1 binding utilized the horizontal orientation of the chip; length of injection and dissociation time varied, but the flow rate was a constant 30 µL/min. BAD-1 remaining bound at the end of the dissociation period was removed by a 1-minute pulse at 30 µL/min of 1 M NaCl, 50 mM NaOH. The buffer used in the binding studies was 20 mM Hepes pH 7,100 mM NaCl, and 0.1% Tween20. To demonstrate specificity of the binding, a mixture of BAD-1 and heparin were co-injected across the surfaces.
Baseline and injection alignments used software supplied with the instrument. BAD-1 did bind to surfaces lacking heparin, but to a much lesser degree than to modified surfaces; therefore, the inter-spot data was used as a reference to correct for this non-specific binding. The kinetic analysis of the corrected sensorgrams utilized the 1∶1 Langmuir binding model in the software. The fitted curves shown in Figure 5A allowed the on and off rates for each trace to vary but were constrained to a common value for each heparin density.
The UW-biotechnology center synthesized a peptide of SHWSPWSS based on the published sequence known to bind to heparin and inhibit binding by TSP-1 [15]. This peptide represented a minimal version of the WxxWxxW heparin-binding motif. A control peptide SHQSPQSS was synthesized in which the tryptophans were replaced with glutamine residues. These peptides were analyzed for purity by HPLC and mass spec and purified/desalted chromatographically.
Heparin resin was washed with binding buffer (20 mM tricine buffer, pH 7, 50 mM NaCl) three times. Binding competitors (WxxW peptide, control peptide, TR4 or reduced TR4) were added to 4 cubic mm of resin and incubated with agitation at room temp for 20 min. Resin was washed once with binding buffer before addition of fluorescent BAD-1 (eFluor605) and incubation continued for another 20 min. Resin was washed three times with binding buffer and binding was quantified by fluorescence on a FilterMax F5 multi-mode microplate reader as above.
T cell inhibition was assessed as described with minor modifications [27]. Anti-CD3 antibody (5 µg/mL) was immobilized on Nunc Maxisorp 96-well round bottom plates in carbonate buffer (pH 9.5) for 1 hr. Parental Jurkat or related JinB8 (CD47-deficient) T cells were pre-incubated for 10 min at 37°C with 10 µg/mL of recombinant TSP-1 (R&D Systems) or native BAD-1 prior to activation with immobilized anti-CD3 antibody for 2 hr. Total RNA was isolated using an RNeasy kit (Qiagen) and RNA was reverse transcribed using iScript cDNA synthesis kit following manufacturer's instructions (Bio-Rad). Real-time PCR primers for human CD69 and human HPRT1 were generated as described [27]. Real-time PCR was performed using SsoFast EvaGreen Supermix (Bio-Rad) on a MyIQ real-time PCR detection system (Bio-Rad). Fold change in CD69 mRNA expression was normalized to HRPT1 mRNA levels.
Primary T cells were obtained from Blastomyces-reactive 1807 TCR transgenic mice [30]. CD4+ T cells were purified with magnetic beads (BD Biosciences, Franklin Lakes, NJ) according to the manufacturer's instructions. Purified 1807 cells (3×105/well) were added to co-cultures of B. dermatitidis yeast strain #55 (3×105/well) and bone-marrow derived DCs (3×105/well). After 96 hours of co-culture, supernate was harvested and tested for levels of IL-17A or IFN-γ
according to manufacturer's instructions (R&D Systems, Minneapolis, MN), and T cells were analyzed by FACscan flow cytometry (BD Biosciences) for activation as measured by surface display of CD69, CD25, CD44, and CD62L (eBioscience, San Diego, CA; and BD Biosciences). 1807 cells were detected with antibody against the surface Thy1.1 marker specific for the T cells. In some experiments of T cell suppression, 1807 cells were pre-incubated with BAD-1 in varied amounts in PBS/0.5% BSA for 90 minutes at 37°C, and the cells were washed to remove free BAD-1 before addition into the assay. BAD-1 was also tested for suppression of T cell function by adding the protein directly into the co-culture of yeast, DC and T cells.
CD47-deficient JinB8 T cells were transiently transfected with plasmids encoding either CD47 or CD47 with a serine-to-alanine mutation at position 64 (CD47-S64A)(a generous gift of Dr. David Roberts) using Lipofectamine Plus (Invitrogen). Transfections were performed overnight prior to initiation of experiments. To verify re-expression, untransfected and transfected JinB8 cells were incubated with PE-conjugated anti-CD47 antibody (BD Bioscience) and analyzed by flow cytometry. T cell inhibition studies were done, as above, using these transfected cells.
The disulfide loop was created using the Rosetta suite of protein structure prediction software (www.rosettacommons.org). The sequence WCKDPYNCD produced several models of the disulfide region, with the best-fit model used to link the two chains. All the models were energy minimized using the Sybyl suite of Tripos software and the Tripos force field until convergence. All modeling was performed in Sybyl.
Kaplan Meier [68] survival curves were generated for mice that received a lethal infection. Survival times of mice that were alive by the end of the study were regarded as censored. Time data were analyzed by the log rank statistic and exact P values were computed using the statistical package Stat Xact-3 by CYTEL Software Corporation. Survival of different groups are considered significantly different if the two-sided P value is <0.05. When multiple comparisons were made simultaneously, P values were adjusted according to Bonferroni's correction to protect the overall significance level of 0.05. All binding data was analyzed by Prism (Graphpad Corp.) with error bars representing simple SEM.
The coordinates and structure factors have been deposited at the Protein Data Bank (PDB) with the following accession codes. PDB: 2LWP; BMRB: 18618.
Supplementary Methods and Figures are available and appended.
|
10.1371/journal.ppat.1003565 | Coronaviruses Lacking Exoribonuclease Activity Are Susceptible to Lethal Mutagenesis: Evidence for Proofreading and Potential Therapeutics | No therapeutics or vaccines currently exist for human coronaviruses (HCoVs). The Severe Acute Respiratory Syndrome-associated coronavirus (SARS-CoV) epidemic in 2002–2003, and the recent emergence of Middle East Respiratory Syndrome coronavirus (MERS-CoV) in April 2012, emphasize the high probability of future zoonotic HCoV emergence causing severe and lethal human disease. Additionally, the resistance of SARS-CoV to ribavirin (RBV) demonstrates the need to define new targets for inhibition of CoV replication. CoVs express a 3′-to-5′ exoribonuclease in nonstructural protein 14 (nsp14-ExoN) that is required for high-fidelity replication and is conserved across the CoV family. All genetic and biochemical data support the hypothesis that nsp14-ExoN has an RNA proofreading function. Thus, we hypothesized that ExoN is responsible for CoV resistance to RNA mutagens. We demonstrate that while wild-type (ExoN+) CoVs were resistant to RBV and 5-fluorouracil (5-FU), CoVs lacking ExoN activity (ExoN−) were up to 300-fold more sensitive. While the primary antiviral activity of RBV against CoVs was not mutagenesis, ExoN− CoVs treated with 5-FU demonstrated both enhanced sensitivity during multi-cycle replication, as well as decreased specific infectivity, consistent with 5-FU functioning as a mutagen. Comparison of full-genome next-generation sequencing of 5-FU treated SARS-CoV populations revealed a 16-fold increase in the number of mutations within the ExoN− population as compared to ExoN+. Ninety percent of these mutations represented A:G and U:C transitions, consistent with 5-FU incorporation during RNA synthesis. Together our results constitute direct evidence that CoV ExoN activity provides a critical proofreading function during virus replication. Furthermore, these studies identify ExoN as the first viral protein distinct from the RdRp that determines the sensitivity of RNA viruses to mutagens. Finally, our results show the importance of ExoN as a target for inhibition, and suggest that small-molecule inhibitors of ExoN activity could be potential pan-CoV therapeutics in combination with RBV or RNA mutagens.
| RNA viruses have high mutation rates (10−3 to 10−5 mutations/nucleotide/round of replication), allowing for rapid viral adaptation in response to selective pressure. While RNA viruses have long been considered unable to correct mistakes during replication, CoVs such as SARS-CoV and the recently emerged MERS-CoV are important exceptions to this paradigm. All CoVs encode an exoribonuclease activity in nonstructural protein 14 (nsp14-ExoN) that is proposed to prevent and/or remove misincorporated nucleotides. Because of the demonstrated resistance of SARS-CoV to the antiviral drug ribavirin (RBV), we hypothesized that ExoN is responsible for CoV resistance to RNA mutagens. Using RBV and the RNA mutagen 5-fluorouracil (5-FU), we show that CoVs lacking ExoN activity (ExoN−) are highly susceptible to RBV and 5-FU, in contrast to wild-type (ExoN+) CoVs. The inhibitory activity of 5-FU against ExoN− viruses resulted specifically from 5-FU incorporation during viral RNA synthesis that lead to extensive mutagenesis within the viral population, and was associated with a profound decrease in virus specific infectivity. These results demonstrate the proofreading activity of ExoN during virus replication and suggest that inhibitors of ExoN activity could be broadly useful inhibitors of CoV replication in combination with RBV or RNA mutagens.
| The potential for CoVs to cause significant human disease is well demonstrated, with six known HCoVs—HKU1, OC43, NL63, 229E, SARS-CoV and MERS-CoV—causing colds, pneumonia, systemic infection, and severe or lethal disease [1]–[5]. Four of these viruses have been identified in just the last 10 years, with two, SARS-CoV and MERS-CoV, causing lethal respiratory and systemic infection [1], [3]–[6]. Studies over the past 10 years have expanded the known phylogenetic, geographic, and species diversity of CoVs, and support multiple emergence events of CoVs into humans from bats and other zoonotic pools [7]–[10]. The most recent evidence for CoV trans-species movement comes from the emergence of the novel MERS-CoV [1], [11], [12]. From April 2012 to June 2013 MERS-CoV has caused 72 laboratory confirmed cases and up to 50% mortality from severe respiratory and systemic disease in at least 8 countries, with evidence for human-to-human transmission [13]. MERS-CoV is most closely related to the bat CoVs HKU4 and HKU5 [11], and the recently identified receptor dipeptidyl peptidase 4 (DPP4) is present on both human and bat cells [14], providing a compelling argument that zoonotic CoV infections resulting in severe human disease may be more frequent events than previously thought. Because of the lack of epidemiological data, it remains unknown whether multiple introductions from a zoonotic source or human transmission of a mild or asymptomatic disease is responsible for these continuing cases of sporadic severe infections. However, based on the high mortality rates associated with SARS-CoV and those reported for MERS-CoV [13], this novel virus potentially represents a serious threat to global health for which no vaccines or therapeutics currently exist.
CoVs contain the largest known RNA genomes (27–32 kb) and encode an array of 16 viral replicase proteins, including a 3′-to-5′ exoribonuclease (ExoN) domain within nonstructural protein 14 (nsp14) [2], [15]–[17]. Similar to the proofreading subunit (ε) of E. coli DNA polymerase III, CoV nsp14-ExoN is a member of the DEDD superfamily of DNA and RNA exonucleases [15], [18]. This superfamily contains four conserved D-E-D-D acidic residues that are required for enzymatic activity, and mutation of these critical residues within CoV ExoN ablates or significantly reduces ExoN activity [15]. Studies from our group have demonstrated that ExoN activity is essential for high-fidelity replication in both the model CoV murine hepatitis virus (MHV) and SARS-CoV [19], [20]. Inactivation of ExoN activity due to alanine substitution of the first two active site residues results in 15- to 20-fold reduced replication fidelity in cell culture [19], [20] and a 12-fold reduction during SARS-CoV infection in vivo [21], associated with profound and stable attenuation of SARS-CoV virulence and replication. A recent study has shown that bacterially-expressed SARS-CoV nsp14-ExoN can remove mismatched nucleotides in vitro, and that ExoN activity is stimulated in vitro through interactions with the non-enzymatic CoV protein nsp10 [22]. Thus all bioinformatic, genetic and biochemical studies to date support the hypothesis that nsp14-ExoN is the first identified proofreading enzyme for an RNA virus and functions together with other CoV replicase proteins to perform the crucial role of maintaining CoV replication fidelity.
Retrospective clinical studies during the SARS epidemic ultimately concluded that treatment with ribavirin (RBV), an antiviral drug shown to be mutagenic for some RNA viruses [23], [24], was ineffective against SARS-CoV [25]–[28]. Because ExoN activity is required for CoV high-fidelity replication [19]–[21], we sought to determine if ExoN was responsible for CoV resistance to RNA mutagens. Using the nucleoside analog RBV and the base analog 5-fluorouracil (5-FU; [29]) we show that CoVs lacking ExoN activity (ExoN−) are up to 300-fold more sensitive to inhibition than wild-type CoVs (ExoN+). Additionally, using full-genome next-generation sequencing we show that ExoN− viruses accumulate 15- to 20-fold more A:G and U:C transitions, consistent with 5-FU incorporation during RNA synthesis. Ultimately our results suggest the exciting possibility that small-molecule inhibitors of ExoN activity could be potential pan-CoV therapeutics, especially when used in combination with RBV or RNA mutagens.
Murine astrocytoma delayed brain tumor cells (DBT cells) were grown at 37°C and maintained in DMEM (Invitrogen) containing 10% FBS, supplemented with penicillin, streptomycin, HEPES, and amphotericin B. VeroE6 (Vero) cells were grown at 37°C and maintained in MEM (Invitrogen) containing 10% FBS supplemented with penicillin, streptomycin, and amphotericin B. All work with MHV was performed using the reverse genetics infectious clone based on strain MHV-A59 [30], and work with SARS-CoV was performed using the reverse genetics infectious clone based on the Urbani strain [31]. Viral studies using SARS-CoV were performed in Select Agent certified BSL-3 laboratories using protocols reviewed and approved by the Institutional Biosafety Committee of Vanderbilt University and the Centers for Disease Control for the safe study and maintenance of SARS-CoV.
5-fluorouracil (5-FU), ribavirin (RBV), guanosine (GUA) and mycophenolic acid (MPA) were obtained from Sigma. 5-FU and RBV were made as 200 mM stock solutions, and were prepared in DMSO and sterile water, respectively. GUA and MPA were prepared in DMSO as 40 mM or 100 mM stocks, respectively. Low concentration (µM) working stocks were prepared as needed in sterile water prior to dilution in DMEM. Viability of DBT and Vero cells was assessed using CellTiter-Glo (Promega) in 96-well plate format according to manufacturer's instructions. DBT and Vero cells were seeded into opaque tissue culture grade 96-well plates, and DMEM containing RBV or 5-FU was added to each well to achieve the concentrations indicated. Water or DMSO vehicle controls were performed, in addition to a 20% ethanol control for cell death. The cells were then incubated at 37°C for either 12 or 24 h, and cell viability was determined using a Veritas Microplate Luminometer (Promega). The resultant values were then normalized to untreated cells.
Subconfluent monolayers of DBT cells in 6-well plates were pretreated for 30 min at 37°C with 1 mL of DMEM containing vehicle or the indicated concentration of RBV, 5-FU, MPA, or GUA. The drug was then removed and cells were infected with MHV-ExoN+ or ExoN− viruses at an MOI of 1 plaque forming units (PFU)/cell (single-cycle) or 0.01 (multi-cycle) for 30 min at 37°C. Virus was then removed and 1 mL of DMEM containing vehicle, RBV, 5-FU, MPA, or GUA was added to each well. Cells were then incubated at 37°C for either 12 (single-cycle) or 24 (multi-cycle) h. The supernatant was harvested and virus titer was determined by plaque assay on DBT cells. For SARS-CoV studies, subconfluent monolayers of Vero cells in T25 flasks were pretreated for 30 min at 37°C with DMEM containing vehicle, RBV, or 5-FU. The drug was removed and cells were infected with either SARS-ExoN+ or ExoN− viruses at an MOI of 0.1 PFU/cell (single-cycle) for 30 min. The virus was removed and DMEM containing vehicle, RBV, or 5-FU was added back. Cells were then incubated for 24 h, at which point the supernatant was harvested and virus titer was determined by plaque assay on Vero cells. All treated samples were normalized to the untreated vehicle control, and values were expressed as fold change from untreated virus titers.
Viral RNA was harvested from infected cell monolayers using TRIzol reagent (Invitrogen), and was reverse transcribed (RT) using SuperScript III (Invitrogen). Random hexamers (1 µL of 50 µM stock) and 1 µg of total RNA were incubated for 5 min at 70°C. The remaining reagents were then added according to the manufacturer's protocol, and the mixture was incubated at 50°C for 1 h and then at 85°C for 5 min. All RT reactions were performed in a final volume of 20 µL. Real-time qRT-PCR was performed on the RT product using the Applied Biosciences 7500 Real-Time PCR System with Power SYBR Green PCR Master Mix (Life Technologies). Each reaction was performed in a total volume of 25 µL containing 12.5 µL of the Power SYBR Green PCR Master Mix, 125 ng each of the forward and reverse primers and 1 µL of the RT product which was diluted 1∶1000. Viral genomic RNA was detected using primers (forward: ACAGGGTGGAGTTCCCGTTA and reverse: ACGGAAGCACCACCATAAGA) optimized to generate a ∼120 nt portion of ORF1a. These values were normalized using the 2−ΔΔCt method [32] to endogenous expression of the housekeeping gene glyceraldehyde-3- phosphate dehydrogenase (GAPDH) using primers (forward: GGGTGTGAACCACGAGAAAT and reverse: CCTTCCACAATGCCAAAGTT) optimized to yield a ∼120 nt portion of GAPDH [33], [34]. Triplicate wells of each sample were analyzed, and averaged into one value representing a single replicate to minimize well-to-well variation. The cycle parameters were as follows: Stage 1, (1 rep) at 50°C for 2 min; Stage 2, (1 rep) 95°C for 10 min; Stage 3, (40 reps) at 95°C for 15 sec and 57°C for 1 min. One representative product from each treatment was verified by melting curve analysis and agarose gel electrophoresis.
Viral RNA from SARS-ExoN+ or ExoN− infected Vero monolayers was harvested using TRIzol reagent, and was reverse transcribed (RT) using SuperScript III as described above except with 5 µL of random hexamers (50 µM stock), 5 µg of total RNA, and in a final volume of 100 µL for each reaction. Four microliters of RT product was then used to generate 12 overlapping ∼3 kb amplicons for each virus treated with either 0 or 400 µM 5-FU by PCR. The high-fidelity polymerase Easy A (Agilent) was used to ensure that errors were minimal during PCR. All primer sets generated single bands which were then purified using the Wizard SV Gel and PCR Clean-Up System (Promega).
Prior to sequencing, cDNA amplicons were fragmented (Fragmentase, NEB), clustered, and sequenced with Illumina cBot and GAIIX technology as previously described [35]. Between 1.4×108 and 4.5×108 bases, comprised of ∼69-nt reads, were obtained per virus, and CASAVA 1.8.2 was used to demultiplex and create the fastq files. Low quality bases from the ends of each sequence read were then trimmed, using Phred scores as the guiding metric (error probabilities higher than 0.001), and sequences with less than 16 bases after trimming were discarded to reduce false alignment and subsequent false variant calls. The program fastq-clipper (http://hannonlab.cshl.edu/fastx_toolkit/index. html) was used for this quality filtering. The Burrows-Wheeler Alignment tool was then used to align reads to the SARS-CoV ExoN+ or ExoN− reference genomes with a maximum of two mismatches per read [36]. Base calling at each position was determined using SAMTOOLS [37]. After the pileup, an in-house script collected the data per-position. For each position throughout the viral genome, the bases and their qualities were gathered, each variant allele's rate was initially modified according to its covering read qualities based on a maximum likelihood estimation and test for significance using Wilks' theorem. Additionally, an allele confidence interval was calculated and output for each allele. Only alleles with statistically significant p<0.05 values were retained and considered to be true variants. Above 0.01% all variants were found to be statistically significant, while below 0.01% many variants could not be distinguished from background error. Thus, the background noise caused by sequencing error was determined to be 0.01% or less.
Statistical tests were applied where noted within the figure legends and were determined using GraphPad Prism (La Jolla, CA) software. Statistical significance is denoted (*P<0.05, **P<0.01, ***P<0.0001) and was determined using an unpaired, two-tailed Student's t test compared to either untreated samples or to the corresponding ExoN+ sample. For the cell viability studies, treated samples were compared to the DMEM sample containing DMSO.
Because RBV has been shown to be incorporated as ribavirin monophosphate (RMP) into viral RNA during replication [23], [24], [38]–[42], the presence of a proofreading enzyme would be predicted to exclude and/or remove nucleotide misincorporation [43]–[47]. If ExoN is responsible for the resistance phenotype, viruses lacking ExoN activity (ExoN−) should demonstrate increased titer reduction following RBV treatment as compared to wild-type viruses containing ExoN activity (ExoN+). To test this hypothesis, we examined the sensitivity of MHV-ExoN+ and ExoN− viruses to RBV during single-cycle (MOI = 1 PFU/cell) replication in murine astrocytoma delayed brain tumor cells (DBT cells). No toxicity was observed in DBT cells following treatment with up to 400 µM RBV (Figure 1A). MHV-ExoN+ viruses were resistant to 10 µM RBV (Figure 1B), while MHV-ExoN− virus titers decreased by ∼200-fold following treatment with 10 µM RBV. The capacity of 10 µM RBV to inhibit MHV-ExoN− replication is surprising because at least 10-fold higher concentrations of RBV are required to inhibit poliovirus and chikungunya viruses [48]–[50]. This observation could be due to the longer genomes of CoVs or to the mechanism(s) by which RBV inhibits CoV replication.
If RBV is exerting antiviral activity primarily through mutagenesis following incorporation of RMP, MHV-ExoN− viruses should exhibit increased sensitivity during multi-cycle replication. To test this, we determined the sensitivity of MHV-ExoN+ and ExoN− viruses to RBV at a low multiplicity of infection (MOI = 0.01 PFU/cell). Unexpectedly, multi-cycle replication of MHV-ExoN− viruses in the presence of RBV (Figure 1B) was indistinguishable from single-cycle replication.
RBV has been reported to exert antiviral activity through numerous mechanisms [38] including disruption of viral RNA synthesis and inhibition of the cellular enzyme inosine monophosphate dehydrogenase (IMPDH). To determine if RBV treatment was affecting CoV RNA synthesis, we performed two-step real-time quantitative reverse transcription PCR (real-time qRT-PCR) to determine viral genomic RNA levels in the presence or absence of RBV. Similar to Figure 1B, MHV-ExoN+ titers were unaffected, whereas there was a dose-dependent reduction in MHV-ExoN− titers following RBV treatment (Figure 1C, filled bars). Corresponding dose-dependent reductions in MHV-ExoN− genomic RNA were observed (Figure 1C, hatched bars) following RBV treatment, demonstrating that treatment with 10 µM RBV decreased MHV-ExoN− RNA synthesis by nearly 100-fold during replication. Because RBV caused decreased RNA synthesis in MHV-ExoN− viruses, we calculated the relative specific infectivities of both viruses at each RBV concentration (Table 1). The relative specific infectivity of MHV-ExoN− viruses was decreased by 6- to 9-fold following treatment with RBV, while MHV-ExoN+ viruses were unaffected.
In addition to decreasing viral RNA synthesis, RBV could be exerting antiviral activity against MHV-ExoN− through competitive inhibition of IMPDH by RMP [51]. To test this possible mechanism, we treated MHV-ExoN+ and MHV-ExoN− viruses with the specific IMPDH inhibitor mycophenolic acid (MPA; [52]–[54]) during both single- and multi-cycle replication. A concentration-dependent decrease in MHV-ExoN− virus titer was observed following MPA treatment during single-cycle replication (Figure 1D). MHV-ExoN+ titers were reduced by less than 10-fold, consistent with what was observed following RBV treatment (Figure 1B). Similar to RBV, increased sensitivity of MHV-ExoN− viruses to MPA was not observed during multi-cycle replication. If RBV is acting via IMDPH inhibition, addition of extracellular guanosine (GUA) should restore virus titers, as has been demonstrated previously for Dengue virus [55]. Addition of 100 µM GUA following RBV or MPA pretreatment and viral infection had no effect on MHV-ExoN+ viruses (Figure 1E), but completely restored MHV-ExoN− titer even in the continued presence of 10 µM RBV or 1 µM MPA (Figure 1F). These data indicate that the antiviral activity of RBV against MHV-ExoN− viruses is occurring at least in part through decreasing viral RNA synthesis and inhibition of IMPDH. Because our primary goal was to test the role of nsp14-ExoN in the prevention and/or removal of nucleotide misincorporation we did not further investigate how RBV was specifically inhibiting ExoN− viruses. However, these results do show that the presence of ExoN activity is capable of preventing RBV inhibition of CoV replication.
We next examined the sensitivity of MHV-ExoN+ and ExoN− viruses to the pyrimidine base analog 5-FU, which has been shown to be mutagenic for many RNA viruses [29], [56]. Treatment of DBT cells with up to 400 µM 5-FU did not result in any detectable cellular toxicity (Figure 2A). Following treatment with up to 200 µM 5-FU (Figure 2B) during single-cycle infections, MHV-ExoN+ titers were inhibited less than 3-fold, while titers of MHV-ExoN− decreased ∼900 fold, representing a ∼300-fold increase in sensitivity as compared to MHV-ExoN+. During multi-cycle replication, MHV-ExoN+ virus titers were reduced by less than 10-fold following 5-FU treatment, while MHV-ExoN− showed a ∼50,000-fold reduction in titer (Figure 2B). Virus was undetectable by plaque assay at 5-FU concentrations above 80 µM. Analysis of viral RNA synthesis by two-step real-time qRT-PCR demonstrated that MHV-ExoN+ RNA levels were not reduced following 5-FU treatment, while 5-FU treatment resulted in minimal two-to-five fold decreases in MHV-ExoN− RNA (Figure 2C). The specific infectivity of MHV-ExoN− was decreased by 14- and 128-fold following treatment with 100 µM and 200 µM 5-FU, respectively (Table 1). These results demonstrate that ExoN activity confers resistance to 5-FU, and support the hypothesis that 5-FU is driving increased genomic mutagenesis in MHV-ExoN− virus populations, leading to lethal mutagenesis and extinction.
To determine whether SARS-CoV viruses lacking ExoN activity (SARS-ExoN−) also were inhibited by RBV and 5-FU, we infected Vero cells with either SARS-ExoN+ or ExoN− viruses in the presence or absence of RBV or 5-FU. Treatment of Vero cells with up to 400 µM RBV or 5-FU did not decrease cell viability by more than 20% (Figure 3A). Recent reports have described the lack of RBV uptake by Vero cells due to the absence of specific equilibrative nucleoside transporters [57], [58]. Additionally, previous studies have shown that RBV failed to inhibit SARS-CoV replication in Vero cells [59]. Consistent with those reports, in our experiments both SARS-ExoN+ and ExoN− viruses were unaffected by treatment with up to 400 µM RBV (Figure 3B). We therefore performed subsequent experiments with 5-FU. SARS-ExoN+ titers were reduced 3- and 10-fold following treatment with 200 or 400 µM 5-FU, respectively (Figure 3C). In contrast, SARS-ExoN− titers were reduced ∼300-fold by 200 µM 5-FU (Figure 3C), similar to MHV-ExoN− viruses. At 400 µM 5-FU, SARS-ExoN− virus was inhibited 2,000-fold during a single replication cycle, representing a ∼160-fold increase in 5-FU sensitivity compared to SARS-ExoN+ viruses. Thus, our data indicate that increased sensitivity of CoVs to RNA mutagens in the absence of ExoN activity is conserved across diverse members of the CoV family. Of interest, our studies with SARS-ExoN+ also indicate that ExoN-mediated protection from nucleotide misincorporation can be overcome at higher concentrations of mutagen.
Studies with the RNA viruses lymphocytic choriomeningitis virus (LCMV), foot-and-mouth disease virus (FMDV) and vesicular stomatitis virus (VSV) have demonstrated that 5-FU is incorporated as 5-fluorouridine monophosphate (FUMP) into replicating viral RNA, thus increasing genomic mutations [60]–[62]. To determine whether 5-FU was causing increased mutagenesis in SARS-CoV populations, we performed full-genome NGS analysis of both virus populations replicating in the presence or absence of 5-FU. To analyze the entire spectrum of mutations arising during replication, we extracted total intracellular RNA from Vero cells infected with either SARS-ExoN+ or ExoN− viruses following treatment with either 0 µM or 400 µM 5-FU. We then generated 12 overlapping cDNA amplicons of approximately 3 kb in length for each sample. For each of the four samples, 1.4×108 to 4.5×108 bases were sequenced, corresponding to an average coverage depth of between 4,600 and 15,000 at each nucleotide position. We compared the statistically significant minority variants, defined as having a p-value of ≤0.05 following a multiple-testing correction (Benjamini-Hochberg), between the untreated and 5-FU-treated SARS-ExoN+ and ExoN− populations. Following treatment with 400 µM 5-FU (Figure 3D), there was an increase in mutations within the SARS-ExoN+ population from 11 to 259 (24-fold). In contrast, for SARS-ExoN− there were 3648 mutations present within the 5-FU-treated SARS-ExoN− population compared to the 99 mutations in the untreated population (40-fold increase). Most remarkably, this represented a 16-fold increase in the number of statistically significant minority variants between 5-FU treated ExoN+ and ExoN− SARS-CoV. Thus, these data support our hypothesis that 5-FU was increasing genomic mutations through incorporation of FUMP into viral genomes in the absence of ExoN activity.
Incorporation of FUMP instead of uracil into replicating RNA allows FUMP to base pair with both guanosine and adenine [61], [63]. This decreased specificity in base pairing has been shown in studies with LCMV and primarily results in A-to-G (A:G) and U-to-C (U:C) transitions [29], [61], [63]. To determine if FUMP was being incorporated at higher levels in the absence of ExoN-mediated proofreading, we analyzed the numbers and types of transitions and transversions occurring in each virus population (Figure 4). Transitions are indicated in grey boxes and transversions in white boxes, with the number for each shown. Transversions comprised the majority of variants for both untreated ExoN− and ExoN+ viruses. Treatment with 5-FU caused the number of U:C and A:G transitions to increase in both ExoN+ and ExoN− populations, from 2 to 197 for SARS-ExoN+ and from 16 to 3304 for SARS-ExoN− (Figures 4A and B). This increase and bias toward U:C and A:G transitions is consistent with FUMP being incorporated into both minus- and plus-strand RNA [63] during both ExoN+ and ExoN− replication; however the absolute numbers were dramatically increased (16-fold) during ExoN− replication compared to ExoN+. In untreated cells, A:G and U:C transitions accounted for less than 25% of the total minority variants within each population (Figure 4C). Following 5-FU treatment, A:G and U:C transitions accounted for 70–95% of the total minority variants within each population.
To further examine the genomic distribution of these two transitions, we plotted the total number of A:G and U:C transitions occurring at a frequency of between 0.1% and 1% (Figure 5). Approximately 75% and 90% of the total minority variants occurring at a frequency between 0.1 and 1% following 5-FU treatment were due to A:G or U:C transitions (Figure 5), for the SARS-ExoN+ and ExoN− populations, respectively. In both populations, these mutations were distributed across the entire genome following treatment with 400 µM 5-FU. Thus our data provide direct evidence indicating that 5-FU drives increased genomic mutations within SARS-CoV in the absence of ExoN proofreading activity.
Viral sensitivity to RNA mutagens is determined by several factors including polymerase selectivity [39], [40], [64]–[67], mutational robustness [68], and the acquisition of mutations that increase or decrease replication fidelity. Increased and decreased fidelity mutants have been described for picornaviruses and arboviruses [35], [48], [50], [69], all of which have occurred in the viral RdRp. The CoV nsp14-ExoN is the first identified RNA virus protein distinct from the RdRp that affects replication fidelity [19]–[21], [70]. While the G641D mutation within the chikungunya (CHIKV) nonstructural protein 2 (nsP2) has been implicated in CHIKV resistance to RBV, a direct role for this protein in fidelity regulation has not been described [48]. A Sindbis virus variant containing mutations within nsP1, a viral guanylyl- and methyltransferase [71], has been shown to be resistant to both RBV and MPA [72]. However, this phenotype is related to viral RNA capping and not replication fidelity [72]. In this report, we identify CoV ExoN activity as a critical determinant of viral sensitivity to RNA mutagens. Using two phylogenetically distant β-CoVs we demonstrate that this phenotype is well conserved across CoVs. Clearly, there is a profound increase both in overall mutations and in specific 5-FU-associated mutations within the ExoN− population as compared to the ExoN+ wild-type population. Furthermore, the vast majority of statistically significant mutations were distributed genome-wide at frequencies between 0.2 and 1%, providing strong evidence supporting ExoN-mediated proofreading during CoV replication. Of interest, our experiments also revealed that ExoN-mediated prevention and/or removal of misincorporated nucleotides is not absolute, especially in the setting of higher concentrations of mutagen. ExoN+ SARS-CoV populations demonstrated 24-fold more mutations following 5-FU treatment, suggesting that ExoN proofreading can be overwhelmed by higher concentrations of mutagens and likely by other nucleoside or base analogs. This raises the further possibility that ExoN may be less efficient at recognizing or removing some types of nucleoside or base analogs than others, and that such approaches to virus inhibition might be viable, particularly in combination with inhibitors that target ExoN activity.
The antiviral nucleoside analog RBV is currently used to treat hepatitis C virus (HCV; [73]–[75]), Lassa virus [76] and respiratory syncytial virus (RSV) infections [77], [78]. The potential clinical use of RBV for CoV infections is complicated by the multiple mechanisms of action that have been reported [38], and by the potential for disease exacerbation, as reported during the SARS-CoV epidemic [25]–[28]. Our data suggest that RBV primarily inhibits MHV-ExoN− virus replication through decreasing viral RNA synthesis and inhibition of IMPDH (Figure 1). Inhibition of IMPDH by RMP has been shown to decrease intracellular GTP pools [51], thus altering the balance of nucleoside triphosphates (NTPs) within the cell. Decreased GTP levels could result in forced misincorporations due to NTP imbalances in the absence of ExoN activity [72]. However, the moderate 6- to 9-fold decreases in relative specific infectivity observed for MHV-ExoN− following RBV treatment (Table 1) suggests that mutagenesis is not the primary mechanism by which RBV is exerting an antiviral effect. An additional possibility is that the antiviral activity of RBV against ExoN− viruses is unrelated to the putative proofreading function of this enzyme. Both biochemical and cell culture studies have demonstrated that loss of ExoN activity leads to impaired RNA synthesis [15], [19], [20]. Furthermore, in addition to ExoN activity, nsp14 contains N7-methyltransferase (N7-MTase) activity, a critical step in RNA capping [79], [80]. A recent report has demonstrated that the ExoN and N7-MTase domains are structurally inseparable, and that residues within the ExoN domain are important for N7-MTase activity [81]. Thus, the increased sensitivity of MHV-ExoN− to RBV could result from the impairment of undefined functions of ExoN during CoV replication, particularly during RNA synthesis. The parallel use of ExoN+ and ExoN− viruses with RBV may allow us to define how RBV is exerting an antiviral effect against CoVs and the potentially novel mechanisms by which ExoN may act to counter that inhibition.
Since the identification of nsp14-ExoN activity [15] and studies demonstrating the requirement for ExoN in high-fidelity replication [19]–[21], mounting evidence points to a role for nsp14-ExoN in proofreading activity during RNA virus replication [22]. Here we used NGS to determine the number of mutations present in SARS-ExoN+ and ExoN− populations. The characteristic 5-FU-mediated transitions U:C and A:G comprised 90% of the total statistically significant minority variants within SARS-ExoN− population, and were present at levels 15- and 20-fold higher than those same transitions within the ExoN+ population (Figure 4). Overall, our data represent the first direct test of ExoN proofreading during SARS-CoV replication in the absence of ExoN. Furthermore, the sequencing depth attained using NGS shows that ExoN inactivation likely skews the spectrum of spontaneous mutations present within the untreated population (Figure 4). Such overrepresentation of specific mutations in the context of ExoN inactivation is similar to studies of S. cerevisiae DNA polymerases ε and δ containing mutations within their respective 3′-to-5′ DEDD exonucleases [82]–[86]. This altered distribution due to ExoN inactivation could have profound implications for CoV adaptation and evolution.
Lethal mutagenesis occurs through the accumulation of mutations within the viral genome during replication, and ultimately results in virus extinction (reviewed in [56], [87]). While lethal mutagenesis has been studied extensively [87], our work is the first to identify an RNA virus protein distinct from the RdRp that directly regulates the sensitivity of RNA viruses to genomic mutations resulting from mutagen incorporation. Currently, RBV is the only FDA-approved antiviral with demonstrated mutagenic activity. The first demonstration of RBV acting as a mutagen was performed using poliovirus [23], [24] almost 30 years after the antiviral activity of RBV was described [88]. The nucleoside analog T-705 (Favipiravir; [89]) is currently in clinical development, and has been shown recently to drive lethal mutagenesis of influenza virus [90]. We have shown that ExoN+ viruses replicate well in the presence of RBV or 5-FU. However, we also have shown that ExoN− mutants of SARS-CoV and MHV have 15-to-20-fold decreased fidelity [19], [20], are attenuated, are subject to rapid loss of replication and clearance in vivo [21], and are highly susceptible to low concentrations of RNA mutagens. An exciting possibility is that this conserved CoV proofreading enzyme could be targeted for inhibition, thus leading to the development of broadly useful CoV therapeutics. While ExoN inhibitors alone might be efficacious, combining an inhibitor of CoV fidelity with an RNA mutagen would magnify the intrinsic fidelity defect of ExoN inhibition and drive high-level mutagenesis. A potential advantage of such an approach would be to rapidly drive the virus to extinction, while limiting or blocking the capacity of the virus to overcome inhibition by reversion. ExoN− mutants of both MHV and SARS-CoV have shown no reversion over multiple passages in culture or during persistent infections in vivo [19]–[21]. Furthermore, we did not observe any primary reversions within the ExoN DEDD motif following 5-FU treatment. While mutations within the CoV RdRp could emerge during acute treatment, mutations within other RNA virus RdRps have demonstrated that the maximum tolerance for increased or decreased fidelity without loss of virus viability is between ∼3- to 6-fold [35], [48], [69], [91]. In addition, our data demonstrate that ExoN− viruses are profoundly sensitive to inhibition by lower concentrations of mutagen, providing a possible improved therapeutic index and margin of safety for use.
In summary, this study provides the most direct evidence to date that CoV ExoN provides a proofreading function during virus replication, and identifies ExoN as the critical determinant of CoV sensitivity to RNA mutagens. Because CoV replication fidelity is likely determined by the concerted effort of multiple virus proteins [19], [20], [22], our data suggest the exciting possibility that significant attenuation of CoV fitness and pathogenesis could be achieved by targeting the conserved process of CoV replication fidelity. Ultimately, uncovering the mechanism of fidelity regulation and methodologies to disrupt this critical process will be vital to responding to both endemic and future emerging CoVs such as SARS-CoV and MERS-CoV.
|
10.1371/journal.ppat.1005284 | Influenza Virus Induces Cholesterol-Enriched Endocytic Recycling Compartments for Budozone Formation via Cell Cycle-Independent Centrosome Maturation | Influenza virus particles are assembled at the plasma membrane in concert with incorporation of the virus genome, but the details of its spatio-temporal regulation are not understood. Here we showed that influenza virus infection induces the assembly of pericentrosomal endocytic recycling compartment (ERC) through the activation of Rab11a GTPase and cell cycle-independent maturation of centrosome by YB-1, a multifunctional protein that is involved in mitotic division, RNA metabolism and tumorigenesis. YB-1 is recruited to the centrosome in infected cells and is required for anchoring microtubules to the centrosome. We also found that viral infection accumulates cholesterol in ERC and is dependent on YB-1. Depletion of YB-1 shows reduced cholesterol-enriched ERC and prevented budozone formation at the plasma membrane. These results suggest that cholesterol in recycling endosomes, which are emanated from ERC, may trigger the virus assembly concomitantly with the packaging of the virus genome. We propose that the virus genome is transported to the plasma membrane by cholesterol-enriched recycling endosomes through cell cycle-independent activation of the centrosome by YB-1.
| Influenza virus particles are assembled at the plasma membrane in concert with incorporation of the virus genome, but the details of its spatiotemporal regulation are unknown. We found that the virus genome is transported to the plasma membrane using cholesterol-enriched recycling endosomes through cell cycle-independent activation of the centrosome by recruiting YB-1, which is a mitotic centrosomal protein. We also revealed that the cholesterol-enriched endosomes are important for clustering of viral structural proteins at lipid rafts to assemble the virus particles. These results suggest that local accumulation of cholesterol, via fusion of endosomes to the plasma membrane, is one of the triggers for the virus assembly concomitantly with arrival of the virus genome beneath the plasma membrane.
| Endocytic transport pathways are important to arrange the plasma membrane components for diversified cellular processes at the plasma membrane including virus budding. Endocytosed proteins are first delivered to the early/sorting endosomes, from where proteins are either recycled back to the plasma membrane or transported to late endosomes and lysosomes. Rab small GTPase family members show distinct intracellular localization and function as molecular switches to regulate vesicle carrier formation and fusion with target membranes. Rab11a-positive recycling endosomes are crucial for recycling and delivery of plasma membrane components to the cell surface [1–3]. The Rab11a-positive transport vesicles emerge from specific organelles called endocytic recycling compartments (ERC). ERCs constitute a collection of tubular organelles that are close to the nucleus and associated with the microtubule organizing centre (MTOC). However, the functional significance of ERCs is not fully understood.
MTOC is a highly dynamic structure that achieves precise control of the microtubule array for the spatial and temporal regulation of several fundamental processes. Microtubule dynamics is controlled through continuous switching between phases of growth and shrinkage, as well as the level and timing of nucleation from the centrosome, which is the major MTOC in animal cells. The centrosome is composed of a pair of centrioles surrounded by pericentriolar material (PCM), a matrix of more than a hundred different proteins. PCM proteins are organized radially around the centriole in a toroid-like arrangement [4–7] and PCM serves as a platform for microtubule nucleation. During mitosis, in a process known as centrosome maturation, PCM increases in size to promote the microtubule nucleation for mitotic spindle formation [8,9].
The influenza viral genome forms a viral ribonucleoprotein complex (vRNP) with viral RNA polymerases and nucleoprotein (NP). After viral genome replication in the nucleus, the progeny vRNP is nuclear-exported and then accumulates around the centrosome [10]. vRNP is then transported to the budding site beneath the cell surface along microtubules through Rab11a-dependent recycling endosomes [11–13]. Recently, Y-box binding protein-1 (YB-1) was reported to function as a porter to facilitate vRNP accumulation at the centrosome [14]. YB-1 is a major component of cellular mRNA ribonucleoprotein complexes and it regulates mRNA translation and degradation [15]. It is also reported that YB-1 accumulates in the centrosome during G2/M phases [16] and is required for the centrosome maturation [17].
Cholesterol is a major constituent of the plasma membrane in eukaryotic cells. It regulates the physical state of the plasma membrane and is involved in the formation of membrane microdomains, called lipid rafts. Lipid rafts are defined as small (10–200 nm), heterogeneous, highly dynamic, sterol- and sphingolipid-enriched domains that compartmentalize cellular processes [18]. Small rafts can sometimes coalesce to form larger platforms through protein-protein, protein-lipid, and lipid-lipid interactions. Three viral membrane proteins, HA, NA, and M2, are embedded in the influenza virus envelope. M1 covers the inner viral membrane leaflet and binds to the cytoplasmic tails of HA and NA [19]. The assembly and budding of viral particles are coupled with the formation of functionalized raft domains, called budozone [20]. In the budozone, HA, possibly together with NA, is enriched by clustering several small rafts [21,22]. M2 possesses cholesterol-binding motifs [23,24], but a relatively short transmembrane domain of M2 prevents complete immersion of the protein in the more ordered raft domains. Thus, M2 is thought to localize to the edge of the budozone to mediate the pinching off of virus particles from the plasma membrane [25]. Finally, vRNP is recruited to the budozone through the interaction of vRNP with M1 to initiate budding and release of virus particles.
Here we showed that influenza virus infection induces the assembly of pericentrosomal ERCs through the activation of Rab11a and microtubule dynamics. Using three-dimensional structured illumination microscopy (3D-SIM), we found that YB-1 forms a toroid-like structure with a beads-on-a-string distribution pattern around the centriole. Knockdown (KD) analyses indicated that influenza virus stimulates the spontaneous centrosome maturation in interphase by recruiting YB-1 to anchor newly synthesized microtubules onto the centrosome. We also found that cholesterol accumulates in the pericentrosomal ERC with vRNP in an YB-1-dependent manner. Disruption of the cholesterol-enriched ERC formation by YB-1 KD results in defective viral budozone formation at the plasma membrane. Collectively, these results suggest that the recycling endosomes containing cholesterol and vRNP emanate from ERC, and cholesterol in recycling endosomes is a trigger for the viral budozone formation concomitantly with vRNP trafficking to the plasma membrane.
Transferrin is a typical marker to monitor the organization of active recycling endosomes during endocytosis and its return to the cell surface. To examine the dynamics of the recycling pathway in influenza virus-infected cells, cells were pulse-labeled for 30 min with transferrin Alexa fluor 568, followed by a chase for 30 min without fluorescent transferrin. At 3 h post infection, transferrin-positive recycling endosomes were accumulated in ERC at a juxta-nuclear region, possibly near the centrosome (Fig 1A and 1B, white arrowheads). Transferrin recycling proceeds with a t1/2 of approximately 20 min [26], therefore the transferrin uptake should correspond to a steady-state distribution of the labeled ligand (Fig 1A). We next performed an indirect immunofluorescence assay using anti-Rab11a antibody and FISH assay using a probe that hybridizes with the segment 1 virus genome (Fig 1C, arrowheads). As is the case for transferrin, Rab11a was also present in the juxta-nuclear region and colocalized with the virus genome in approximately 40% of infected cells at 6 h post infection (P<0.001), suggesting that the virus genome is recruited to the pericentrosomal ERC after nuclear-export, as previously reported [10–14].
It has been shown that active Rab11a shows a marked accumulation of ERC at the centrosome [27]. To evaluate whether Rab11a is activated by influenza virus infection, we purified active Rab11a (Rab11-GTP) by GST pull-down assays using Rab11-binding domain of Rab11-FIP2. Rab11-FIP2 acts as an effector molecule for Rab11-GTP through a highly conserved Rab11-binding domain (RBD) among Rab11-FIP family proteins [28]. Therefore, we can purify Rab11-GTP (constitutive active mutant Q70L, lane 8), but not the GDP form (dominant negative mutant S25N, lane 9), using GST-fused 41 amino acid peptide derived from RBD of Rab11-FIP2 (GST-RBD) (Fig 2A). Next, we performed GST pull-down assays with lysates prepared from infected cells using GST-RBD at 8 h post infection (at which the virus genome is actively transported) and the co-purified Rab11a was analyzed by western blotting with anti-Rab11a antibody (Fig 2B). The amount of Rab11a co-purified with GST-RBD from infected lysates was 4.5 ± 0.6 times more than that from mock-treated lysates (Fig 2B; representative results from three independent experiments are shown), suggesting that a guanine nucleotide exchange factor (GEF) for Rab11a may be activated in response to infection.
By interacting with a number of Rab11-FIPs, Rab11a associates with distinct motor proteins, enabling bidirectional transport along microtubules. Thus, recycling endosomes closely associate with microtubules, and their intracellular transport is fully dependent on the microtubule dynamics, which undergo cycles of nucleation, growing, and shrinking. The precise spatial and temporal regulation of the cycles is essential for the numerous cellular functions in which microtubules are involved.
Previously, we reported that YB-1 accumulates in the centrosome with vRNP during interphase [as shown in Fig 1C, and [14]]. At 48 h post transfection of YB-1 siRNA, the expression level of YB-1 in KD cells decreased to 25% of that in control cells (S1 Fig). The virus titer in YB-1 KD cells decreased to approximately 30% of that in control cells (Fig 3A). We also found that Rab11a does not accumulate in the centrosome of infected YB-1 KD cells (Fig 3B), suggesting that YB-1 is required for pericentrosomal ERC formation. Note that YB-1 is responsible for centrosome maturation in order to establish the polarity-dependent dynamic instability in the mitotic phase [17]. Thus, we hypothesized that YB-1 may stimulate pericentrosomal ERC formation through spontaneous centrosome maturation in infected interphase cells as it does in the mitotic phase. To test this hypothesis, we examined the centrosomal localization of YB-1 using 3D-SIM super-resolution microscopy (Fig 3C, 3D, 3E and 3F). Note that only centrosomes showing a cross-sectional view of PCM during interphase were selected for this analysis. YB-1 formed a toroidal structure with a beads-on-a-string distribution pattern around GFP-centrin-2, a marker protein of the centriole (Fig 3C and 3E). The mean diameter of the YB-1 toroid at the peak intensity (545 ± 48 nm; n = 8) was similar to that of pericentrin toroid (a marker for PCM; 581 ± 42 nm; n = 8), suggesting that YB-1 localizes in PCM (Fig 3D and 3F). However, YB-1 did not co-localize with pericentrin (Fig 3E). It has been proposed that pericentrin exists as elongated fibrils that extend radially from the centriole [5,6]. The spatial domains separated by pericentrin are filled with a number of PCM proteins required for microtubule nucleation and anchoring, suggesting that YB-1 also regulates the microtubule nucleation and/or anchoring at PCM in response to infection at interphases.
Next, we observed the dynamics of microtubule nucleation to examine the centrosome function in infected cells using EB1-GFP [8], which interacts specifically with growing microtubule ends (Fig 4 and S1, S2, S3 and S4 Videos). The time series of EB1-GFP were acquired at 1.56-sec intervals for 1 min. In image sequences, EB1-GFP comets continually emerged from the centrosome. In the control, the mean growth rate of nucleated microtubules in the infected cells was increased compared to that of the uninfected mock cells (Fig 4B, P<0.001). In contrast, EB1-GFP in infected cells treated with YB-1 siRNA mostly did not move in a straight line, but rather in a Brownian motion (Fig 4A and S4 Video). Because growing microtubule ends decorated with EB1-GFP accumulated primarily in the centrosome of infected YB-1 KD cells (Fig 4A, arrow head), it is likely that the microtubules nucleated from the centrosome even in infected YB-1 KD cells. Therefore, it is possible that the newly synthesized microtubules are released from the centrosome in infected YB-1 KD cells. Further, although most microtubules were still elongated radially from the centrosome (Fig 4A), some of the EB1-GFP signals showed a faster migration rate in uninfected YB-1 KD cells (Fig 4A and 4B). It has been reported that short microtubules released from the centrosome migrate faster than the centrosomal microtubules [29], therefore YB-1 appears to be required, at least in part, for anchoring microtubules to the centrosome in uninfected interphase cells.
To address whether YB-1 is involved in the anchoring of microtubules to the centrosome in response to infection, we carried out microtubule regrowth assays using nocodazole, a potent inhibitor of microtubule polymerization (Fig 5). After nocodazole treatment for 1 h, microtubules were depolymerized, and α-tubulin was dispersed throughout the cytoplasm (Fig 5B, 5G, 5L and 5Q). After washing out the drug, cells were incubated at 37°C to allow the regrowth of the microtubules for 3, 5, and 15 min. As expected, the nucleation of microtubules from the centrosome was stimulated by infection in control cells at 5 min post release (Fig 5I). In contrast, noncentrosomal microtubules were sporadically found at peripheral regions of the cytoplasm in infected YB-1 KD cells (Fig 5R and 5S, arrowheads). These results suggest that YB-1 is required for anchoring newly polymerized microtubules to PCM when the microtubule nucleation is stimulated by infection.
ERC is reported to be involved in intracellular sorting and polarized trafficking of apical plasma membrane components [26]. However, details regarding the roles of ERC remain to be clarified. Therefore, we next examined the loading of vRNP onto the recycling endosomes by using YB-1 siRNA to disrupt ERC formation. Cells constitutively expressing FLAG-Rab11a were subjected to immunoprecipitation assays with anti-FLAG antibody (Fig 6A). We found that the amount of PB1 subunit of viral polymerase and NP coimmunoprecipitated with FLAG-Rab11a from YB-1 KD lysates were decreased to approximately 30% of those from control lysates (Fig 6A, lane 6). This result is supported by the fact that vRNP hardly colocalized with Rab11a in YB-1 KD cells as shown in the enlarged panel of Fig 3B. Furthermore, we examined the activation of Rab11a in YB-1 KD cells by GST pull-down assays using GST-RBD. The amount of Rab11-GTP was not changed between the control and YB-1 KD cells (Fig 6B), suggesting that YB-1 KD does not influence the amount of active recycling endosomes. Thus, it is likely that the formation of pericentrosomal ERC is important to load vRNP onto the endosomal vesicles.
Cholesterol is not uniformly distributed in the membrane, and 80–90% of total cellular cholesterol is enriched in the plasma membrane [30]. Although recycling endosomes contain considerably less cholesterol than the plasma membrane, it is known that the endocytic transport pathway through recycling endosomes is important for cholesterol trafficking and homeostasis in cells [31,32]. Therefore, we hypothesized that vRNP is transported to the plasma membrane via recycling endosomes with cholesterol. To test this hypothesis, we observed the intracellular localization of cholesterol in infected cells using the fluorescent cholesterol-binding polyene antibiotic, filipin. Some recycling endosomes were partially colocalized with cholesterol in uninfected cells (Fig 6C). However, along with the formation of pericentrosomal ERC by infection, we found that cholesterol is highly enriched in ERC in an YB-1-dependent manner. Similar results were obtained in A549 cells infected with A/Panama/2007/99, which is one of the representative strains of seasonal influenza A virus (H3N2) (S2 Fig). These findings suggest that vRNP is transported to the plasma membrane via recycling endosome vesicles that contain a higher concentration of cholesterol.
Some viruses, including influenza virus, are known to utilize lipid rafts for budding from the plasma membrane [33]. Viral budozone formation is thought to be dependent on the spatial assembly of eight-segmented vRNP complexes and viral membrane proteins via clustering of lipid rafts. Although it has been reported that reorganization of cortical actin is required for the control of viral budozone formation [25,34,35], the trigger to initiate the coalescence of lipid rafts is unclear. Thus, we examined whether the pericentrosomal ERC is required for budozone formation by using in situ proximity ligation assay (PLA) to detect the proximity between M2 and HA. In the in situ PLA system, the theoretical maximum distance between two target proteins is around 40 nm to yield amplified signals. At 8 h post infection, cells were subjected to in situ PLA using anti-HA and either anti-M2 or anti-M1 antibodies (Fig 7). Strong punctate PLA signals (red) between HA and M2 or between HA and M1 were observed at the plasma membrane in the infected control cells (Fig 7A and 7B). Although HA and M2 were successfully transported to the plasma membrane in YB-1 KD cells (Fig 7C), the intensity of PLA signals between HA and M2 was significantly decreased by YB-1 KD (P<0.001; Fig 7B, left panel). In contrast, the signal intensity between HA and M1 was not decreased in YB-1 KD cells (Fig 7B, right panel). This could be due to the direct binding of M1 with the cytoplasmic tail of HA [19]. Next, we examined whether cholesterol is required for the YB-1-dependent viral budozone formation using nonraft HA mutant virus, which has alanine substitutions at I533, Y534, and S535 in the transmembrane domain of HA. It is reported that this mutant HA rarely associates with lipid rafts and that the apical transport is delayed, but not blocked [36]. At 12 h post infection, a significant amount of HA was observed at the plasma membrane in nonraft virus infected cells (Fig 7D, green). However, the intensity of PLA signals between HA and M2 was dramatically reduced in nonraft virus-infected cells compared with that in wild-type infected cells (Fig 7D and 7E). Thus, as expected, it is likely that most of the signals observed in the in situ PLA system were mediated by lipid rafts. Furthermore, in contrast to wild type virus (Fig 7B), the PLA signals between nonraft HA and M2 were nearly unaffected by YB-1 KD (Fig 7E, compare lane 2 with lane 3), suggesting that the interaction of HA with cholesterol is important for YB-1-mediated viral budozone formation.
The lipid-lipid, lipid-protein, and protein-protein interactions facilitate the formation of small raft domains into functional platforms for signal transduction, membrane trafficking, and cell adhesion [37–39]. Sphingolipids that have been enriched in these assemblies have saturated and longer acyl chains with larger polar headgroups, so cholesterol functions as spacers between sphingolipids through their acyl chains [40]. This cholesterol-sphingolipids interaction results in the packing and condensing of lipid rafts for their clustering. Fig 7 shows that YB-1 is important for clustering of viral membrane proteins at the plasma membrane through the interaction of viral raft protein with cholesterol. It is noteworthy that the amount of cholesterol at the plasma membrane was unchanged between the control and YB-1 KD cells (S3 Fig), suggesting that small raft domains should be intact at the plasma membrane in YB-1 KD cells. This is possibly due to the fact that the recycling endosomes and TGN contain much less cholesterol than the plasma membrane [41]. However, it is known that moderate changes in the level of cholesterol transported through these compartments appear to have drastic effects on cellular homeostasis [41]. Taking these findings together, we propose that the fusion of cholesterol-enriched recycling endosomes with the plasma membrane induces the accumulation of sphingolipids that contain viral raft proteins which form viral budozone concomitantly with the arrival of vRNP beneath the plasma membrane (Fig 8).
In general, cells acquire cholesterol mainly through receptor-mediated endocytosis of low-density lipoprotein (LDL) [30]. After LDL internalization, LDL-cholesterol is delivered to late endosomes and lysosomes to release the cholesterol molecules from LDL. The majority of cholesterol in late endosomes is then delivered to the plasma membrane. Although the itinerary of cholesterol from late endosomes to the plasma membrane is not clear, it is thought that cholesterol is transported through ER, TGN, and recycling endosomes. We found that influenza virus infection stimulates cholesterol accumulation in ERC (Fig 6C). This could be due to a possibility that the accumulation of recycling endosomes in ERC (Fig 1A) may slow down the delivery of cholesterol to the plasma membrane.
YB-1 is required for centrosome maturation during mitosis [17], but little is known about the function of YB-1 in the centrosome. In infected cells, YB-1 was localized in PCM and formed a radial and toroidal structure around the centriole (Fig 3). It is proposed that the PCM proteins might be assembled based on the nine-fold radial symmetry of the centriole [5,6]. In which case, it is assumed that YB-1 is also a structural component of the PCM matrix for microtubule assembly. It is also reported that YB-1 interacts with microtubules and coats the outer surface of the microtubule wall in vitro [42]. Thus, YB-1 may connect microtubules to the PCM matrix by decorating the microtubules’ minus ends.
The spatiotemporal regulation of Rab GTPase activity is of particular importance. Among the several GEFs known to regulate Rab GTPases, no GEF that activates Rab11a has been identified in mammalian cells despite a systematic characterization of the DENN domain subfamily of Rab GEFs [43]. It is necessary to identify the GEFs responsible for virus infection.
Rab11a plays a role in the transport of M2 to the apical membrane [25], although M2 is directly transported through TGN to the plasma membrane [44]. This is due to the fact that Rab11a also functions in constitutive exocytosis from TGN in addition to the recycling processes via ERC [45,46]. In YB-1 KD cells, HA and M2 were successfully transported to the plasma membrane (Fig 7D), suggesting that centrosome maturation by YB-1 is required for their transport through ERC but not through TGN. It has been reported that the minus end of microtubules, which is released from the centrosome, could subsequently be captured by the Golgi membrane and then elongated into linear arrays [47]. Thus, even in the absence of YB-1, the exocytic transport from TGN might be achieved along microtubules that are elongated from Golgi stacks.
The majority of membrane proteins are sorted at TGN before their delivery to the appropriate cell surface domain. In addition to TGN, some other cellular lipid raft proteins, such as TLR4 and EGF receptor, are transported to the plasma membrane through the recycling endosomes [48,49]. Additionally, the transport rates of recycling endosomes are controlled in response to signaling pathways to increase or decrease the surface expression of molecules, such as insulin-regulated glucose transporter GLUT4 [50,51]. In this study, we propose that the recycling endosomes deliver cholesterol to the plasma membrane for not only cholesterol homeostasis, but also lipid raft clustering. Our findings contribute to the understanding of the molecular mechanism of lipid raft clustering in response to several signals that utilize lipid rafts as a platform.
Influenza virus A/Puerto Rico/8/34 strain and rabbit polyclonal antibodies against PB1, NP, M1, and YB-1 were prepared as previously described [14]. Mouse antibodies against HA (TaKaRa; C179), Rab11a (BD; 47/Rab11), Pericentrin (Abcam; ab28144), α-tubulin (Sigma; DM1A), and a rabbit antibody against M2 (Abcam; ab56086) were purchased. HeLa cells (a gift from Dr. Masa-atsu Yamada of University of Tokyo) were grown in minimal essential medium (MEM) containing 10% fetal bovine serum. Plasmids expressing GFP-centrin-2 and EB1-GFP were prepared as previously described [14]. To establish HeLa cell lines constitutively expressing either GFP-centrin-2 or EB1-GFP, cells were transfected with pSV2-Neo and either pCAGGS-GFP-centrin-2 or pCAGGS-EB1-GFP. The transfected cells were cultured in the presence of 1 mg/ml of G418 for 2 weeks, and then the G418-resistant colonies were isolated. For the construction of plasmid expressing GST-Rab-binding domain (RBD) of FIP2, cDNA was amplified from pCAGGS-FIP2 (provided by Dr. F. Momose, Kitasato University) with primers 5ʹ-CCGGAATTCGAGCTGGTGAAACAC-3ʹ and 5ʹ-ACGCGTCGACTCACGGCACTCTGAG-3ʹ. The cDNA was cloned into pGEX-6P-1. Nonraft HA virus was generously provided by Drs. Y. Morikawa and F. Momose (Kitasato University) [36] and amplified using MDCK cells constitutively expressing HA (provided by Dr. N. Takizawa, Institute of Microbial Chemistry).
Transferrin conjugated with Alexa 568 was purchased (Life Technologies). Cells were incubated with 100 μg/ml of Transferrin for 30 min at 37°C. After washing with medium, cells were further incubated for 30 min at 37°C and then fixed in 4% paraformaldehyde (PFA).
Indirect immunofluorescence assays and fluorescence in situ hybridization (FISH) assays were carried out as previously described [14]. Briefly, cells infected with influenza virus at multiplicity of infection (MOI) of 10 were fixed with 1% PFA for 10 min and then pre-permeabilized on ice with 0.01% digitonin in PBS for 5 min on ice. After being washed with PBS, cells were fixed in 4% PFA for 10 min and permeabilized on ice with 0.5% Triton X-100 in PBS for 5 min. After incubation in PBS containing 1% bovine serum albumin for 1 h, coverslips were incubated with each antibody for 1 h and then with Alexa Fluor 488-, 568-, and 633-conjugated secondary antibodies, respectively (Life Technologies). After indirect immunofluorescence assays, FISH assays were performed using an RNA probe complementary to the segment 1 virus genome. Images were acquired using confocal laser scanning microscopy (LSM700; Carl Zeiss) or super-resolution microscopy (3D-SIM ELYRA; Carl Zeiss).
Cells were fixed in 4% PFA for 10 min and then incubated with 200 μg/ml of filipin (Sigma). After washing with PBS, images were acquired by Axio Observer Z1 microscope using 63x Apochromat objective (NA = 1.4) with AxioCam MRm camera (Carl Zeiss).
Observations were made with Axio Observer Z1 microscope using 63x Apochromat objective. Images were acquired at 1.56-sec intervals for 1 min with confocal laser scanning microscopy (LSM700; Carl Zeiss). All experiments were carried out at 37°C under 5% CO2 in a temperature-controlled stage (Carl Zeiss). Sequential images were processed using Image J digital image processing software (National Institutes of Health, Bethesda). The average velocity of the punctate fluorescent signals of EB1-GFP was measured using a manual object tracking plugin, MTrackJ, for Image J.
Cells were fixed with 4% PFA, followed by blocking with 1% milk for 30 min. The cells were incubated with mouse anti-HA antibody for 1 h and fixed again in 4% PFA. Cells were then permeabilized with 0.5% Triton X-100 for 5 min and incubated with either rabbit anti-M1 or anti-M2 antibody for 1 h. PLA was carried out using Duolink In Situ PLA kit (Olink Bioscience) according to the manufacturer’s protocol. The mean intensity of the PLA signals was measured using IMARIS software (Carl Zeiss).
Knockdown of YB-1 was examined as previously described [14]. Briefly, cells (5 x105) were transfected with 30 pmol of siRNA using Lipofectamine RNAi Max (Life Technologies) according to the manufacturer’s protocol.
|
10.1371/journal.pntd.0003198 | Leishmania Specific CD4 T Cells Release IFNγ That Limits Parasite Replication in Patients with Visceral Leishmaniasis | Visceral leishmaniasis (VL) is associated with increased circulating levels of multiple pro-inflammatory cytokines and chemokines, including IL-12, IFNγ, and TNFα, and elevated expression of IFNγ mRNA in lesional tissue such as the spleen and bone marrow. However, an immunological feature of VL patients is that their peripheral blood mononuclear cells (PBMCs) typically fail to respond to stimulation with leishmanial antigen. Unexpectedly, it was recently shown that Leishmania specific IFNγ, can readily be detected when a whole blood stimulation assay (WBA) is used. We sought to define the conditions that permit whole blood cells to respond to antigen stimulation, and clarify the biological role of the IFNγ found to be released by cells from VL patients. CD4+ T cells were found to be crucial for and the main source of the IFNγ production in Leishmania stimulated whole blood (WB) cultures. Complement, antibodies and red blood cells present in whole blood do not play a significant role in the IFNγ response. The IFNγ production was reduced by blockade of human leukocyte antigen (HLA)-DR, indicating that the response to leishmanial antigens observed in WB of active VL patients is a classical HLA- T cell receptor (TCR) driven reaction. Most importantly, blockade of IFNγ in ex-vivo splenic aspirate cultures demonstrated that despite the progressive nature of their disease, the endogenous IFNγ produced in patients with active VL serves to limit parasite growth.
| Our research aims to understand the immune failure underlying progression of human visceral leishmaniasis (VL). A key immunological feature of VL patients is that their peripheral blood mononuclear cells (PBMCs) do not respond to stimulation with leishmanial antigen. Surprisingly, when employing a whole blood assay we discovered significant levels of IFNγ in response to soluble Leishmania donovani antigen (WBA) in VL patients. We were interested to understand the relevance of the IFNγ to the anti-parasitic response. Animal models and in vitro studies have shown that IFNγ is a key effector cytokine required for control of the infection, however, the role of endogenous IFNγ in control of parasites in VL patients, has not been demonstrated. Our results show that CD4 cells were required for and were the source of Leishmania specific IFNγ in WBA of VL patients. Optimal IFNγ response required interaction with HLA-DR, supporting that VL is not due to an intrinsic Th1 response defect per se. The Leishmania driven IFNγ appears to limit parasite growth in patients with active VL, since blockade of IFNγ ex-vivo in splenic aspirate cultures enhanced parasite survival. This suggests that IFNγ may have been prematurely dismissed as an adjunct therapy in treatment of VL.
| Visceral leishmaniasis is a chronic disease caused by the protozoan parasites Leishmania donovani and Leishmania infantum/chagasi. Leishmania are transmitted by the bite of phlebotomine sand flies, and replicate within macrophages of their mammalian hosts. In VL, the target organs are chiefly the liver and the spleen. The disease is characterized by prolonged fever, spleno-hepatomegaly, wasting, hypergammaglobulinemia, pancytopenia and almost always leads to death if left untreated.
Based on experimental models, acquired resistance against Leishmania infection requires the development of a Th1 type immune response, characterized by IL-12 production by antigen presenting cells (APC) and IFNγ production by T cells [1], [2]. IFNγ is a key effector cytokine required for activation of infected macrophages for killing (reviewed by Kima and Soong [3]). Patients with active VL have depressed cell-mediated immune responses, reflected by the failure of their peripheral blood mononuclear cells (PBMCs) to proliferate and/or to produce IFNγ in response to stimulation with Leishmania antigens, while their ability to respond to polyclonal stimulation or other antigens, such as the purified protein derivative of Mycobacterium tuberculosis (PPD), remains relatively intact [4], [5]. The absence of antigen specific responses is thought to underlie the disease progression. Paradoxically, the acute phase of VL is associated with elevated expression of IFNγ mRNA in lesional tissue, such as the spleen and bone marrow, as well as increased circulating levels of multiple pro-inflammatory cytokines and chemokines, including IL-12, IFNγ and TNFα [4], [6]. These results imply that the failure to respond to Leishmania antigen stimulation observed in VL patients is not due to a defect in the ability to mount protective Th1 responses per se, but rather to induction of suppressive factors, e.g. IL-10, resulting in unresponsiveness of infected macrophages to activation signals [7].
Studying immunological aspects of human VL has been severely hampered by the inability to measure antigen specific responses, including IL-10, using PBMC. The discovery of antigen specific cytokine responses following stimulation of whole blood (WB) [8] showed that VL patients are not void of Leishmania specific IFNγ responses, findings that could be reconciled with the elevated levels of IFNγ mRNA and circulating cytokines detected in active VL patients. Subsequent studies reported that the whole blood assay (WBA) could also be used to detect antigen-specific IL-10 responses [9], [10]. Thus, the WBA has opened up new possibilities for research aimed at understanding immunological determinants of the disease [8], [9], [10], [11].
We sought to define the requirements for IFNγ production seen using the WBA, and determine if the IFNγ had a biological function in patients with active VL. We show that CD4+ T cells produce Leishmania specific IFNγ in WB cultures. The responses to stimulation with Leishmania antigen observed in WB cultures of active VL patients occurred in the absence of complement, antibodies or cytokines present in serum of VL patients. Employing a splenic aspirate (SA) culture technique, as previously described [11], we show that IFNγ neutralization promotes parasite growth in active VL cases ex-vivo. These findings demonstrate that the elevated levels of IFNγ in patients with active VL serve to limit parasite replication and suggest that therapeutic administration of IFNγ may still hold potential.
All VL patients presented with clinical symptoms of kala-azar at the Kala-azar Medical Research Center (KAMRC), Muzaffarpur, Bihar, India, and were confirmed to be VL positive by detection of amastigotes in splenic aspirates and/or by detection of antibodies against the recombinant antigen, K39. Venous blood and/or splenic aspirates (SA) samples collected from 84 (33 female and 51 male) patients with active VL were included in this study. All patients were treated with Amphotericin B and eventually cured disease. Aggregate clinical data of active VL patients are presented in Table 1.
The use of human subjects followed recommendations outlined in the Helsinki declaration. Informed written consent was obtained from all participants and/or their legal guardian when under 18 years of age. All human samples were coded an analysed anonymously. Ethical approval (Dean/2008-09/314, Dean/2012-2013/89) was obtained from the ethical review board of Banaras Hindu University (BHU), Varanasi, India.
Whole blood (WB) was cultured using a volume of 0.5–1 ml blood per culture condition in round bottom 5 ml polypropylene tubes. For stimulation the samples were treated with SLA (10 µg/ml). Control samples were treated with PBS. In some assays PHA (10 µg/ml) or Staphylococcus enterotoxin B, SEB, (5 µg/ml) was used as positive controls (not shown). Samples were incubated for 37°C in the presence of 5% CO2 for 24 hours if not otherwise indicated in figure text.
To block HLA-TCR interaction 20 µg/ml anti-HLA-DR, clone 243, or isotype control IgG2a, clone MOPC-173, both ultra-LEAF purified (BioLegend, US) were added to the cultures simultaneously with antigen.
To test if complement, antibody and/or other proteins present in plasma, but removed during purification of PBMC, affected SLA induced IFNγ production we replaced the plasma in the WB samples. In brief, total blood cells were pelleted by centrifugation (500 g, 10 minutes, 18°C), the plasma was removed and blood was washed twice with PBS. To determine if complement affected the response, the plasma was heat inactivated [12] at 56°C for 30 minutes and added back to the autologous sample to restore the original blood volume. Alternatively, the plasma was replaced with heat inactivated fetal calf serum (HI-FCS).
To determine the effects of different cell populations on SLA induced IFNγ production we used magnetic beads and columns designed for the isolation/depletion of CD4 and CD8 cell subsets from whole blood as per manufacturer protocol (Whole Blood Column kit, Milteny Biotec). To control for the effect and spontaneous uptake of magnetic beads [13] we used anti-FITC beads (Milteny Biotech) as control. Whole blood and whole blood depleted of the cell subsets of interest were subsequently stimulated as described above. In these assays the patient plasma was replaced with HI-FCS prior to incubation with whole blood beads.
The influence of RBC on SLA induced IFNγ was tested by lysis of RBC. Briefly, the total blood cells were centrifuged (500 g, 10 minutes, 18°C), followed by removal of plasma (see above). The cell pellet was resuspended in 5 ml hypotonic saline (0.6% NaCl) solution/1 ml blood for 20–30 seconds to lyse RBC. To stop the lysis an equal volume of hypertonic solution (1.6% NaCl) was added. The tube was filled with PBS and the non-lysed cells were pelleted and resuspended in autologous plasma to reconstitute the original volume, and stimulated as described above.
Splenic needle aspirates were collected for diagnostic purposes before treatment of VL. Approximately 100 µl SA was obtained by fine needle biopsy, following preparation of smears for diagnostic purpose, the residual cells were placed directly in 1 ml RPMI supplemented with 10% heat-inactivated fetal calf serum (HI-FCS) 200 mM Streptomycin and 100 U/ml penicillin (C-RPMI) and 5 U/ml heparin. Samples were transported to the laboratory at BHU maintaining a temperature of 4–8°C. All samples were processed within 24 h of collection. For stimulation, the SA divided into two equal parts and treated with SLA (10 µg/ml) as done for the in the WBA (described above).
For baseline quantification of amastigotes by limiting dilution, 150 µl SA suspension was directly plated in a 96-well and serially diluted by transfer of 50 µl SA onto biphasic medium of 50 µl blood agar overlaid by 100 µl of M199/C, as previously described [14]. The remaining SA suspension was seeded into 96 well-culture plates (250 µl/well). Monoclonal antibody against human IFN-γ, clone 25723 (R&D Systems) or control IgG2b clone 20116 (R&D Systems) were each added to a final concentration of 20 µg/ml. The SA was incubated for 3 days at 37°C in 5% CO2, the supernatants were collected for cytokine assessment and the removed volume replaced by C-M199 medium, prepared as previously described [15]. From the SA culture 150 µl was transferred into a 96-well plate for estimation of parasite load by limiting dilution as described above. The number of viable parasites was determined from the highest dilution at which promastigotes could be grown out after 7 to14 days of incubation at 25°C.
For comparison with the WB, SA suspension was divided in two parts, stimulated with SLA (10 µg/ml) or with PBS and incubated for 24 hours at 37°C in 5% CO2, where after the supernatant was collected for cytokine assessment.
Following 24 hours of stimulation (if not otherwise indicated) IFNγ and IL-10 were measured in culture supernatants by ELISA. ELISA was performed as per manufacture instruction. For detection of IFNγ the ELISA Max Deluxe set (BioLegend) or the QuantiFeron kit (Cellestis, Australia) were used. IL-10 was measured using matched antibody pair kits from BD Pharmingen. All values calculated from standard curve over or equal to zero were considered in statistical analysis. Negative values were assigned the value zero.
To determine the cellular source/s of cytokines in the WBA, the cultures were stimulated for 16–24 hours. To block cytokine secretion cultures were for the last 6–8 hours of stimulation treated with GolgiStop (BD Biosciences) according to manufactures instructions. Following lysis of RBC using BD RBC lysis buffer (BD Biosciences), cells were surface stained using combinations of FITC, PE and PerCP/PE-Cy5 conjugated antibodies directed to CD3 (Clone UCHT1), CD4 or CD8 (all from BD Biosciences). Surface stained cells were fixed and permeabilized using BD Cytofix/Cytoperm, as per manufactures instruction, washed in permeabilization buffer (BD) and stained for presence of intracellular IFNγ and IL-10 using APC and PE conjugated antibodies (both from Pharmingen) respectively. Following intra cellular staining (ICS), samples were acquired on FACSort (BD Biosciences) and analyzed using CellQuest Pro (BD) or FlowJo (Treestar) software. Analysis was done on cells gated as viable lymphocytes based on their forward–side scatter. SEB (10 µg/ml) stimulated samples were used as positive control for ICS (not shown).
Statistical analyses were done using PRISM5 (GraphPad Software). Different treatments using the same donor samples were compared by the Wilcoxon signed rank test for paired samples. Correlation between results was determined using Spearman-test for non-parametric correlations. Differences with P-values<0.05 were considered as significant. Outliers (donors with extreme values in one or more of the test conditions) were removed from data sets after being defined as outlier using GraphPad on-line Grubb's test for outliers.
The whole blood Quantiferon assay (WBA) was originally designed as a tool for diagnosis of tuberculosis, and detects cytokine (IFNγ) concentrations in plasma supernatants after 16–24 hours of incubation with antigen. To determine the kinetics of the WB responses in VL patients we measured secreted cytokines in supernatants after 6 hours to five days of stimulation with soluble Leishmania antigen (SLA). The induction of IFNγ was rapid and observed in supernatants already 6 hours after stimulation, reaching a plateau at 18–24 hours (Figure 1a, b). Antigen-induced IFNγ was not detected in WB cultures following 72 hours culture or more (figure 1b). We conclude that the IFNγ response seen in the WB cultures is rapid and short lived. For practical reasons stimulation times of 24 hours were used in subsequent assays if not otherwise indicated.
We further tested if antigens specific responses could be detected in short-term (24 hr) splenic aspirate (SA) cultures. In line with the observations made using the WBA, an increase in IFNγ was observed in supernatants of 73% of SA cultures following stimulation with SLA, indicating that antigen specific cells are present at the site of infection (figure 1c). In contrast to the SLA stimulated WB cultures where IL-10 tended to be induced [9], [10], IL-10 levels dropped in SA cultures following SLA stimulation (figure 1d).
The immune system of patients with VL is highly activated. We considered the possibility that other blood cell or serum components that are removed in the process of PBMC purification could be required for the Leishmania specific WB response. To address the effect of plasma components we replaced the plasma with i) autologous heat-inactivated plasma, to determine the role of complement, or ii) heat inactivated fetal calf serum (HI-FCS), to remove antibodies, complement, or other serum factors such as cytokines that may be elevated in VL. To address if RBC were important, we lysed the RBC using hypotonic treatment. None of these treatments affected the net production of IFNγ measured using the WBA (figure 2), indicating that complement, antibodies, cytokines, or RBC are not important for the observed SLA induced IFNγ production in WB. Indeed, removal of autologous plasma with HI-FCS potentiated the SLA induced response (figure 2). The replacement of plasma with FCS was subsequently employed in some of the assays that followed.
Understanding the cellular source/s of IFNγ in the WB is critical to our reinterpretation of the immunologic defects in kala-azar. To determine the cellular requirements for IFNγ production we removed various subsets from whole blood of VL patients prior to stimulation with SLA. Removal of CD4 cells caused a substantial loss of SLA induced IFNγ in WB cultures, while removal of CD8 cells had no effect (figure 3a). Blockade of HLA using a pan-HLA-DR antibody caused a significant loss of SLA induced IFNγ (figure 3b). This suggests that the IFNγ response induced by SLA stimulation depends on HLA-TCR interaction. Three out of the 12 patient samples in which the effect of HLA-DR blockade was evaluated had low IFNγ responses to SLA (<100 pg/ml). To confirm CD4 T cells as the source of IFNγ in WB, we assessed intracellular IFNγ by FACS. SLA induced IFNγ was only observed in the CD3+ population (all events considered). Figure 3c shows that the IFNγ is produced by CD3+CD4+ cells, while figure 3d shows that there is a strong correlation between the frequency of IFNγ positive T cells (CD3+) and the IFNγ measured in WB culture supernatants by ELISA. IFNγ was not detected in the CD3+CD8+ population following SLA stimulation and almost all cells producing IFNγ following SLA stimulation were CD3+CD8− (not shown).
To test if neutrophils contributed to the IFNγ responses CD15+ cells were removed using depletion beads, this caused a partial though significant loss of SLA induced IFNγ (figure S1), which may indicate an involvement of neutrophils in the observed SLA response, but since CD15 can be expressed on other cells, i.e. monocyte, we cannot exclude that the effect seen is due to removal of these cells.
IL-10 can be induced in stimulated WB from VL patients, albeit at low levels. Removal of CD4 cells caused a small but significant reduction of the amount of detectable IL-10 in SLA stimulated WB (figure 3e), indicating that CD4+ and other cells are sources of antigen-specific IL-10 in VL patients. CD8 cells do not appear to contribute to SLA induced IL-10 response, and their removal caused a slight enhancement of this response (Figure 3e). The source of SLA induced IL-10 could not be confirmed by intracellular staining as the number IL-10 positive cells was below the limit of reliable detection.
In experimental models it is well established that IFNγ mediates control of parasite replication [16] and that lack of IFNγ signalling causes disease progression [17], [18]. The same protective function is assumed in humans, but the direct proof that IFNγ controls parasite replication in human VL is lacking. To test if the endogenous IFNγ, which we now know to be elevated during active disease, plays a role in parasite control, we treated ex-vivo SA cultures with neutralizing antibodies against human IFNγ followed by assessment of parasite growth, as previously described in assays designed to test the function of endogenous IL-10 [11]. Following neutralization of IFNγ, the parasite load in SA increased in 19/31 (61%), was unchanged in 8/31 (26%) and decreased in 4/31 (13%) samples (figure 4a). The IL-10 levels in the SA supernatants were not affected by neutralization of IFNγ (figure 4b), suggesting that the inhibitory effect of IL-10 on parasite killing does not completely abolish the parasite-controlling effects of endogenous IFNγ. The background levels of IFNγ detectable in ex vivo SA cultures were significantly reduced when CD4 cells were removed (figure 4c), indicating that CD4 cells are needed for the splenic IFNγ production.
In the search for markers of L. donovani infection, epidemiological studies utilising a WBA revealed Leishmania specific IFNγ responses, long considered absent, in patients with active VL [8]. The goals of the current study were to validate the prior WBA results, to reveal the conditions required for SLA induced IFNγ secretion by WB and to determine if the IFNγ seen in patients with active disease functions to limit the infection.
Whole blood contains cell populations, proteins, lipids and sugars that are largely removed when PBMC are purified. To test if such components were required for the antigen specific response we deprived WB cultures of RBC, plasma and complement. We found that replacement of autologous plasma and RBC lysis had no effect on the SLA induced IFNγ response. By contrast, removal of CD4+ cells revealed these cells to be the main source of antigen specific IFNγ secretion in the WB cultures, a finding that was substantiated by direct intracellular staining. In line with previous observation CD8 T cells were not found to contribute to SLA responses in patients with active VL [19].
Removal of CD15+ cells also reduced the IFNγ levels detectable in the SLA stimulated WB. CD15 (Lewis X) is a carbohydrate adhesion molecule primarily expressed on mature neutrophils in blood, but is also present on a subset of monocytes [20]. The decline in IFNγ levels following CD15 depletion may thus be explained by a reduction of APCs required for the T cell response, but could also imply that neutrophils contribute to the response. By contrast, Abebe et al. have proposed, based on the observation that VL patients have more CD15+ and higher content of arginase expressing CD15+ cells pre compared to post treatment patients or endemic controls, that neutrophils contribute to the unresponsiveness of VL PBMC [12]. Neutrophil inhibition of the antigen-specific IFNγ response in VL patients is not supported by the data presented here, where a reduction in IFNγ secretion by WB cells was observed following CD15 depletion.
The detection of IFNγ responses in stimulated splenic aspirate cells (figure 1c) indicates that antigen specific and responsive cells are present at the site of infection. Depletion of CD4 cells from ex vivo SA cultures support these cells as the source of IFNγ at the site of infection. In contrast to the WB, where IL-10 was also induced following SLA stimulation, IL-10 levels decreased in SA following SLA stimulation (figure 1d). More critically, we show that the endogenous IFNγ produced by splenic cells is biologically active and served to limit parasite growth in the SA cultures from the majority of VL patients, as shown by the increase in parasite numbers after IFNγ neutralization ex-vivo. The lack of effect of the IFNγ neutralization on parasite growth observed in some samples can be attributed to the nature of the SA. The sampling is done blind and the aspirates may vary in red and white blood cell content as well as the extent of disruption of infected cells, resulting in extracellular amastigotes that will be unaffected by the level of IFNγ released. The treatment with anti-IFNγ-antibodies did not affect the IL-10 levels detected in the SA supernatants (figure 4), suggesting that the inhibitory effects of IFNγ on parasite survival and growth occurs even in the presence of high levels of IL-10. The IFNγ response we detect in active cases, while functional, is clearly not a sufficient condition for cure, as the patients would succumb to the disease without treatment. We propose that fragility and/or short life span of these cells may limit their ability to mediate a fully curative response, although other factors, in particular IL-10, are clearly involved [11].
Our data suggest that even in untreated patients, their disease progression would be far worse in the absence of the endogenous IFNγ that they produce. Notably, there are patients whose cellular responses cannot be detected even when using the WBA. While not directly reflected in the clinical parameters (i.e. blood chemistry), these patients may have progressed further in the disease and lost the responding population. It may be noted that there was a negative correlation between SLA induced IFNγ response in WBA and parasite load in blood (Spearman r = −0.66; p = 0.004, n = 17), which indicates that the WB SLA response to a degree may reflect the severity of disease. Genetic or acquired defects in their ability to mount Th1 responses to Leishmania may also underlie the lack of response in some patients. We found that the SLA induced IFNγ response involved HLA-DR interaction as treatment with HLA-DR blocking antibody reduced the IFNγ levels in all donors tested (figure 3b), with an average decrease of 70% compared to control antibody treatment. The partial effect observed may be explained by utilization of HLA-DQ in the presentation of leishmanial antigens to T cells. While HLA-DR together with it's peptide is the classical ligand for T cells recognizing foreign antigens, HLA-DQ may also present peptides from pathogens and initiate T cells responses. The role of HLA molecules on WB SLA responses are of interest since risk alleles for development of VL were recently identified within in the MHC class II region [21]. The influence of allelic differences and role of different MHC molecules in the ability to drive Leishmania specific responses in the WB culture are under current investigation.
The functional Th1 response in active VL patients may also be highly relevant to their response to treatment. L. donovani infection in T cell deficient mice revealed a clear role for antigen specific T cells in the curative response to pentavalent antimony [22]. Our findings reinforce the rationale for the prior VL treatment trials carried out in the 1990s involving recombinant IFNγ, indicating that monotherapy could be beneficial [23], [24]. The lack of response to monotherapy in some patients and the absence of a long-lasting therapeutic effect, as well as the limited success as adjunct therapy with sodium stibogluconate [25], discouraged further trials. Our present and more recent studies suggest that antigen-specific IFNγ production may in some patients not be the limiting factor in their non-curative response.
In summary, our data support the notion that disease progression in VL is not due to a complete failure in Th1 development. Our findings make clear that WB cultures may allow detection of functionally relevant immune responses not seen using PBMC. Most patients with VL have antigen specific CD4 T cells capable of secreting IFNγ both in the blood and at the site of infection - the spleen. We further show that the IFNγ produced by VL patients play a role in limiting parasite growth.
|
10.1371/journal.pbio.2000731 | Lung Basal Stem Cells Rapidly Repair DNA Damage Using the Error-Prone Nonhomologous End-Joining Pathway | Lung squamous cell carcinoma (SqCC), the second most common subtype of lung cancer, is strongly associated with tobacco smoking and exhibits genomic instability. The cellular origins and molecular processes that contribute to SqCC formation are largely unexplored. Here we show that human basal stem cells (BSCs) isolated from heavy smokers proliferate extensively, whereas their alveolar progenitor cell counterparts have limited colony-forming capacity. We demonstrate that this difference arises in part because of the ability of BSCs to repair their DNA more efficiently than alveolar cells following ionizing radiation or chemical-induced DNA damage. Analysis of mice harbouring a mutation in the DNA-dependent protein kinase catalytic subunit (DNA-PKcs), a key enzyme in DNA damage repair by nonhomologous end joining (NHEJ), indicated that BSCs preferentially repair their DNA by this error-prone process. Interestingly, polyploidy, a phenomenon associated with genetically unstable cells, was only observed in the human BSC subset. Expression signature analysis indicated that BSCs are the likely cells of origin of human SqCC and that high levels of NHEJ genes in SqCC are correlated with increasing genomic instability. Hence, our results favour a model in which heavy smoking promotes proliferation of BSCs, and their predilection for error-prone NHEJ could lead to the high mutagenic burden that culminates in SqCC. Targeting DNA repair processes may therefore have a role in the prevention and therapy of SqCC.
| Human lungs are constantly exposed to inhaled chemicals that have the potential to damage cellular DNA. Lung stem cells must therefore have the ability to repair DNA damage to survive and achieve tissue homeostasis. Lung airways are composed of different types of cells, including basal cells, which have been proposed to be the stem cells of the lung. Here, we show that lung basal stem cells have a superior ability to resolve DNA damage compared to alveolar progenitor cells, thus allowing these cells to survive and proliferate after injury. Accordingly, basal stem cells isolated from patients with a long history of tobacco smoking had remarkable proliferative potential compared to those extracted from never smokers. However, we demonstrate that basal stem cells predominately use nonhomologous end joining to repair DNA double-strand breaks, a notoriously error-prone pathway. A subset of polyploid basal stem cells was observed in cigarette-smoking patients, pointing to the accumulation of genetic instability in these long-lived cells. Gene expression analyses revealed that lung squamous cell carcinoma, a subtype of lung cancer that almost exclusively occurs in smokers, carries a transcriptional fingerprint of basal cells, suggesting that lung basal stem cells could be the cells of origin of this subtype of lung cancer. We postulate that further unravelling of DNA repair in lung cells may lead to potential therapeutic targets in the prevention or treatment of lung diseases.
| Human lungs are constantly exposed to inhaled environmental and chemical insults that have the potential to damage cellular DNA. Lung stem and progenitor cells must be capable of repairing their DNA to maintain healthy survival. The failure of stem cells to repair DNA damage can contribute to tissue loss through damage-induced cell death, whereas unfaithful DNA repair in stem cells may invoke carcinogenesis through the accumulation of genetic aberrations [1]. Lung squamous cell carcinoma (SqCC), the second most common histological subtype of lung cancer, exhibits strong genomic instability and occurs almost exclusively in smokers, with 96% of patients having a history of tobacco use [2–4]. The carcinogens present in cigarette smoke are likely responsible for the extraordinarily high mutational rate observed in SqCC compared to other cancers [4].
The early molecular events caused by tobacco exposure and the cell types in which these genetic aberrations occur to induce SqCC formation are not well known. Stem/progenitor cells are putative tumour-initiating cells because of their capacity for renewal and their longevity, allowing for accumulation of genetic lesions. Susceptibility of different lung epithelial progenitor cells to DNA damage has not been explored and could further inform the mechanisms involved in smoking-induced carcinogenesis. DNA damage encompasses alterations to bases, strand cross-links, single-strand breaks (SSBs), and double-strand breaks (DSBs). DSBs, which have been shown to arise after cigarette smoke exposure [5–7], are the most dangerous type of DNA lesion, as they can result in loss or gain of genetic information through insertions, deletions, or chromosomal translocations. DSB repair occurs through either homologous recombination (HR), a high-fidelity DNA repair mechanism, or nonhomologous end joining (NHEJ), an unfaithful mechanism that is implicated in genomic instability and tumour formation [1,8].
Different types of lung progenitor cells have been proposed in distinct anatomical regions of the lung [9]. Lung airways are composed of basal, secretory, ciliated, and neuroendocrine cells. Basal stem cells (BSCs), present only in the human cartilaginous airways or the mouse trachea [10], are located between the basement membrane and the luminal airway cells and have been proposed as stem cells of the lung [9,11–13]. The alveolar compartment is composed of alveolar type 1 and type 2 (AT1 and AT2) cells. AT2 cells have progenitor activity and can replenish both AT1 and AT2 cells following lung injury [14,15], although recent studies suggest that AT1 cells could also serve as progenitors in the mouse lung after pneumonectomy [16,17]. Different cell surface markers have been used to isolate human lung BSCs [12,18,19], but few markers allowing separation of other lung epithelial cell types have been identified [20,21].
Here we used flow cytometry to isolate BSCs, luminal (club, goblet, and ciliated) cells, and AT2 cells from fresh human proximal and distal lung tissue and showed that BSCs and AT2 cells behave as progenitor cells in an in vitro colony-forming assay. BSCs from heavy smokers had an increased proliferative potential compared to those of never smokers, whereas AT2 progenitor activity was diminished in patients with long smoking histories. To investigate this striking difference in lung stem/progenitor cell response to cigarette smoke exposure, we asked if the DNA repair mechanisms differed between the two cell types. DNA damage studies following ionizing radiation or exposure to a chemical agent demonstrated that human and mouse BSCs repair their DNA more efficiently than alveolar progenitor cells using the unfaithful NHEJ pathway, leading to cell survival and proliferation. In addition, polyploidy, a phenomenon occurring during oncogenesis [22], was only observed in the BSC subset, indicating that these cells may be more prone to transformation. Bioinformatics analyses revealed that lung SqCCs carry a transcriptional fingerprint of human lung BSCs, suggesting that BSCs may behave as the cells of origin of this subtype of lung cancer. In addition, high expression levels of key NHEJ genes in lung SqCCs are associated with increased genomic instability. Collectively, our data indicate that error-prone DNA repair is a hallmark of lung SqCC and suggest that targeting NHEJ may play a role in SqCC prevention and/or treatment.
Fresh human lung samples were obtained from patients undergoing lung cancer surgery and held intact in media until processing. The tissues were collected distally from the tumour sites and subdivided into proximal (containing cartilaginous airways and surrounding parenchyma) and distal (containing distal noncartilaginous airways and surrounding parenchyma) regions. A novel fluorescence-activated cell sorting (FACS) strategy was developed to deplete pre-erythrocytes, fibroblasts, and haematopoietic and endothelial cells from lung cell suspensions. Epithelial cells (EpCAM+, epithelial cell adhesion molecule) were then subdivided based on their expression level of CD166 (encoded by ALCAM), CD49f (α6 integrin), and T1α (also known as podoplanin). Three populations were defined in proximal samples: CD49fhiT1α+CD166mid (termed P5), CD49fmidT1α-CD166hi (termed P6), and CD49fmidT1α-CD166mid (termed P10), and two populations in distal samples: CD49fmidT1α-CD166hi (P6) and CD49fmidT1α-CD166mid (termed P10) (Fig 1A and S1A Fig). These populations were consistently observed in the 121 lung samples analysed.
Quantitative PCR analysis (Fig 1B) and intracellular FACS staining (Fig 1C) of known intracellular markers of lung epithelial cells revealed that each population contained distinct cell types. P5 cells expressed high levels of the basal cell markers TP63 and keratin 5 (KRT5). In contrast, both proximal and distal P6 populations contained cells that exhibited strong expression of markers of club (secretoglobin 1A1, SCGB1A1), goblet (mucin 5AC, MUC5AC), and ciliated cells (forkhead box J1, FOXJ1; acetylated-tubulin) indicating this subpopulation is enriched in luminal airway cells (Fig 1B, 1C and 1D). The P10 subsets expressed high levels of the AT2 lineage marker surfactant protein C (SFTPC). Transmission electron microscopy further showed that the P5 population (referred to as BSC) contained cells with numerous mitochondria and keratin filaments, consistent with a basal cell phenotype [23] and their exclusive location in the proximal lung (Fig 1A and 1E). P10 populations (referred to as AT2) contained a homogenous population of AT2 cells, as evidenced by the presence of microvilli and multiple lamellar bodies (Fig 1E). These data establish that the expression of EpCAM, CD49f, CD166, and T1α is sufficient to delineate cellular compartments enriched in BSCs, luminal cells, and AT2 cells in human lung samples.
We then assessed the colony-forming capacity of the five cellular subsets described above in a three-dimensional assay. Only BSCs and AT2 cells generated colonies that were phenotypically distinct (Fig 2A). BSCs formed clonal, hollow, spherical colonies, whilst AT2 colonies were saccular and less uniform (Fig 2A, S1B and S1C Fig). Immunostaining showed that BSC colonies maintain expression of KRT5, whilst AT2 colonies expressed SFTPC, indicating that the cells retain their lineage commitment in this culture system (Fig 2B). Distal AT2 cells had a significantly higher number of colony-forming units (CFUs) than proximal AT2 cells (Fig 2C), suggesting heterogeneity in the AT2 population between proximal and distal lung. We focused on the distal AT2 compartment (named AT2 from now on) because of its increased progenitor activity. Diversity in the basal cell compartment has also been suggested from studies in the mouse trachea, with BSCs and basal progenitor cells possessing different colony-forming capacities [24]. Heterogeneity in the BSC population may also exist in the human lung and could explain the variability in BSC CFUs observed in our study (Fig 2C). To further interrogate the diversity in the colony-forming capacity of BSCs, we investigated the association between patient tobacco-smoking history and the proliferative potential of human lung progenitor cells. Strikingly, BSCs isolated from an exsmoker patient formed numerous large colonies compared to BSCs isolated from a never smoker, which only formed a limited number of small colonies (Fig 2D). Conversely, AT2 cells isolated from an exsmoker had reduced colony-forming capacity compared to AT2 cells from a never smoker that formed multiple large saccular colonies (Fig 2D). Linear correlation analysis demonstrated a positive correlation between years of smoking and the number of BSC CFUs, whereas tobacco exposure was inversely correlated with the number of AT2 cell CFUs (Fig 2E). The number of years since a patient had quit smoking, patient age, and patient sex did not correlate with colony-forming capacity (S2A, S2B and S2C Fig). These data demonstrate that exposure to cigarette smoke activates lung BSCs yet impairs AT2 cells and that this effect is maintained after smoking cessation.
To investigate the molecular mechanisms driving the differential response of AT2 cells and BSCs to cigarette smoking, we performed RNA sequencing on freshly isolated cells from current and exsmokers. Unsupervised clustering showed that each population was molecularly distinct (Fig 3A and S3A Fig). Gene ontology analyses revealed that cell cycle and DNA repair genes were up-regulated in BSCs compared to AT2 cells (Fig 3B). Human BSCs also expressed high levels of telomere maintenance genes, including TERT (S3B Fig), and were found to have longer telomeres than AT2 cells (S3C Fig). Given that both active DNA repair and telomere maintenance are properties of stem cells [25,26], these findings align with mouse studies and confirm that BSCs have greater stem cell-like characteristics than the AT2 progenitor cells [11,27]. BSCs exhibited up-regulation of key genes that control the activation of DNA repair pathways, including ATM (ataxia telangiectasia mutated) [28] (Fig 3C), suggesting that BSCs may have an enhanced ability to respond to DNA damage. To evaluate the sensitivity of human lung stem and progenitor cells to DNA damage, we subjected fresh human lung tissue to ionizing radiation (IR) and assessed the presence of DSBs over time by immunofluorescence analysis of phosphorylated histone 2AX (γH2AX), an early marker of DSBs. Strong γH2AX staining was observed in both the alveolar and BSC compartment 1 h after IR exposure (Fig 3D). However, DSBs were resolved in the basal cell compartment 24 h post IR, whereas γH2AX staining was still detected in the alveolar region at this time point (Fig 3D). These results suggest that human BSCs exhibit an increased capacity to repair their DNA following IR compared to alveolar cells.
Given that the transcriptomic analysis and the study of the response to IR were performed on human samples from patients with different smoking histories, we sought to determine whether this striking difference in gene expression profile and DNA damage response of human BSCs and AT2 cells was acquired as a result of chronic cigarette smoke exposure or if it was an inherent property of the cells. Healthy lungs from never smokers are difficult to obtain; hence, we performed RNA-seq transcriptional profiling on mouse tracheal BSCs and lung alveolar cells (S4A Fig). Gene expression profiles of mouse BSCs and alveolar cells significantly correlated with their human counterparts, indicating that the transcriptome of lung stem/progenitor cells is highly conserved between species (Fig 4A). Consistent with the human data, we observed that DNA repair genes and cell cycle genes were up-regulated in mouse BSCs compared to alveolar cells (S4B Fig), suggesting that both mouse and human lung BSCs may be intrinsically positioned to repair DNA damage. γH2AX immunofluorescence staining of mice subjected to IR revealed that both BSCs and alveolar cells had DSBs 1 h post IR, and this was detected in a dose-dependent manner (Fig 4B, S4C Fig). However, γH2AX expression was resolved in BSCs 4 h post IR, whilst it was still strongly detected in the alveolar compartment 24 h after IR (Fig 4B). Similar results were observed when the mice were injected with bleomycin, a DNA-damaging agent known to induce DSBs [29] (Fig 4B). The sensitivity of mouse alveolar cells to IR was reflected in increased apoptosis that was not observed in BSCs (Fig 4C, S4D and S4E Fig). Whilst it is possible that BSCs undergo cell senescence, BSCs proliferated 4 h and 8 h after IR (Fig 4D), suggesting that senescence in these cells is unlikely. Therefore, lung BSCs have superior DNA repair capabilities leading to cell survival and proliferation that is conserved across species, while alveolar cells exhibit limited DNA repair capacity resulting in DNA damage-induced cell death.
Our results show that BSCs have a prompt and enhanced ability to repair DSBs compared to alveolar cells in vivo. NHEJ is a rapid DSB repair process that occurs in all phases of the cell cycle, while HR functions only in actively cycling cells [30]. Given that the majority of BSCs and alveolar cells reside in the G0/G1 phase of the cell cycle (S4D Fig) and that RAD51, an early marker of HR, was not detected in mouse BSCs after IR (Fig 5A), we interrogated whether BSCs preferentially used NHEJ to repair DSBs. Genes regulating NHEJ were found to be up-regulated in human and mouse BSCs compared to alveolar cells (Fig 5B and S5A Fig), including PRKDC that encodes for the DNA-dependent protein kinase catalytic subunit (DNA-PKcs), a necessary enzyme for the initiation of NHEJ [31]. Activation of DNA-PKcs was observed in lung samples from cigarette-smoking patients and in irradiated mouse lung and trachea as detected by immunofluorescence staining of phosphorylated DNA-PKcs (Fig 5C and 5D).
To assess whether BSCs use NHEJ to repair their DNA, we analysed severe combined immune deficiency (SCIDPrkdc) mice, that have a mutation in Prkdc, leading to a 50% reduction in DNA-PKcs activity and impaired NHEJ [32]. Identical levels of γH2AX were induced in the respiratory system of wild-type (WT) and SCIDprkdc mice 1 h after IR (Fig 6A). Strikingly, DSBs were still detected in tracheas of SCIDprkdc mice 8 h post IR, whereas DSBs were completely resolved in WT tracheas at this time point, indicating that BSCs use NHEJ to rapidly repair their DNA following damage (Figs 4B and 6A). Quantification by flow cytometry confirmed that BSCs from SCIDprkdc mice exhibited delayed DSB repair after IR compared to WT mice, whilst the levels of γH2AX in the alveolar compartment were not affected by reduced DNA-PKcs activity (Fig 6B and 6C). Treatment with bleomycin similarly resulted in delayed DSB repair in BSCs isolated from SCIDprkdc mice compared to WT mice, whereas knock-down of DNA-PKcs activity did not alter γH2AX expression in AT2 cells (S5B and S5C Fig). The impaired ability of BSCs from SCIDprkdc mice to repair their DNA was associated with an increase in cell death following IR that was not observed in WT BSCs (Fig 6D and 6E). In addition, alveolar cells isolated from WT and SCIDprkdc mice had similar levels of apoptosis (S5D Fig). These data establish that BSCs predominantly use the NHEJ pathway to repair DSBs to maintain cell survival and proliferation, whilst lung alveolar progenitor cells have reduced NHEJ activity.
Use of the error-prone NHEJ repair pathway has been associated with increased genetic alterations and genomic instability [1]. Interestingly, polyploid cells were detected in the human BSC subset isolated from exsmokers, whereas no such population was present in AT2 cells (Fig 7A and 7B), suggesting that a proportion of BSCs have increased genetic instability compared to AT2 cell progenitors. These results led us to investigate the involvement of BSCs in lung carcinogenesis. We used expression signature analysis to relate the transcriptome of human lung cell subsets to expression profiles of tumours available from The Clinical Lung Cancer Project (S6A Fig) [33]. We found that the expression profile of human lung BSCs was strongly associated with that of SqCC (Fig 7C and 7D). No correlation was observed between BSCs and other lung cancer subtypes, nor did any other human epithelial cell subset associate with SqCC (S6B–S6E Fig), suggesting BSCs as the putative cells of origin of SqCC. We noticed that human BSCs express higher levels of genes known to be frequently altered in SqCC, such as NFE2L2, SOX2, and PTEN [3,34], compared to the other lung epithelial subsets (Fig 7E). Additionally, BSCs expressed high levels of APOBEC cytidine deaminase genes (Fig 7E), which could explain the APOBEC signature observed in lung SqCC [4]. Interestingly, lung SqCCs express higher levels of DNA repair and cell proliferation genes compared to lung adenocarcinoma, a cancer that arises from lung AT2 cells (S7A Fig) [35,36]. Analysis of the RNA-seq data from The Cancer Genome Atlas data [3,37] showed that key NHEJ genes such as PRKDC and XRCC6 are expressed at higher levels in lung SqCC compared to lung adenocarcinomas and normal lung tissue (Fig 7F and S7B Fig). Strikingly, high expression of PRKDC or XRCC6 in lung SqCC was found to be associated with increased genomic instability (Fig 7G). These data suggest that DNA repair by NHEJ could prove to be a hallmark of lung SqCC. We propose that the use of NHEJ by BSCs could lead to the accumulation of genetic alterations that may culminate in SqCC formation in cigarette-smoking patients (Fig 8), although this hypothesis will need to be validated with functional studies in vivo.
In this study, we used a novel combination of cell surface markers to simultaneously isolate distinct human lung epithelial cell populations and observed that lung BSCs have more stem cell-like characteristics than AT2 progenitor cells. BSCs were found to have longer telomeres and a superior ability to repair DSBs compared to AT2 cells. Our study provides new evidence to indicate that adult lung stem cells have developed more efficient DNA repair mechanisms than differentiated cells to promote cell survival and tissue repair. Consistently, haematopoietic stem cells appear more resistant to IR-induced cell death than myeloid progenitors and were found to activate NHEJ to repair their DNA [38]. NHEJ has also been proposed as a mechanism for DSB repair in breast and hair follicle bulge stem cells [39,40] as a process to evade apoptosis and ensure stem cell longevity.
Cigarette smoke contains a complex mix of carcinogens and toxins that cause DNA damage, including oxidative base damage, the formation of DNA adducts, SSBs, and DSBs [7,41,42]. The observation that lung BSCs have a greater capacity to repair DNA damage compared to alveolar progenitor cells may explain the varied smoking-induced pathologies observed in specific anatomical regions of the lung. Loss of AT2 cells has been identified as a mechanism participating in the pathogenesis of idiopathic pulmonary fibrosis and emphysema-like diseases [43,44]. Different genetic mutations have been associated with the development of these diseases, including aberrations in telomere maintenance genes, SFTPC, MUC5B, and alpha-1 anti-trypsin [45–47]. Our data showing that AT2 cells are highly sensitive to DNA damage, leading to increased cell death and reduced colony-forming capacity, suggest a novel molecular mechanism that may participate in tobacco smoking-induced emphysema. They also provide further supporting evidence that epithelial cell dysfunction plays a role in the pathogenesis of degenerative lung diseases. Radiation therapy is frequently used in lung cancer patients yet is often associated with damage of surrounding normal tissue, resulting in reduced quality of life [48]. Our observation that AT2 progenitor cells have limited DSB repair capacity and increased cell death following IR may also provide insights into the adverse loss of alveolar cells and radiation-induced fibrosis following γ-irradiation.
We observed that BSCs isolated from heavy tobacco users are drastically more proliferative than those from never-smoker patients, which is consistent with the basal cell hyperplasia frequently observed in cigarette smokers [23,49]. Multiple mechanisms most likely account for the activation of BSC proliferation following exposure to tobacco smoke. Firstly, endogenous levels of reactive oxygen species (ROS) have been shown to influence the proliferative capacity of lung cells [18], and elevated ROS levels, like those induced by cigarette smoke exposure, could participate in the higher proliferative potential of BSCs observed in smoker patients. Secondly, cigarette smoking damages luminal airway cells [49,50], and BSCs could be activated to replenish differentiated airway cells. Consistently, studies in mice have shown that depletion of luminal airway cells results in the expansion of BSCs and their differentiation into secretory cells and ciliated cells [12]. We propose an additional mechanism by which the enhanced DNA repair capabilities of BSCs promote their proliferation after cigarette smoke exposure and could participate in smoking-induced basal cell hyperplasia.
Our findings provide evidence that BSCs are more proficient than alveolar cells in using NHEJ to repair their DNA. NHEJ has been implicated in the accumulation of genetic lesions [51] and plays a role in chromothripsis [52–54]—phenomena that participate in the initiation of tumour formation. Quiescent haematopoietic stem cells use NHEJ and displayed increased genomic instability after irradiation compared to progenitor cells, further implicating NHEJ in oncogenesis [38]. We propose the ability of BSCs to rapidly repair DNA through error-prone NHEJ allows the cells to survive longer and places them at greater risk than lung progenitor cells to accumulate mutations, which may ultimately lead to the induction of carcinogenesis (Fig 8).
In vivo studies in genetically modified mice have been used to demonstrate the cell of origin of cancer. AT2 cells were found to act as the tumour-initiating cells in K-RasG12D-driven lung adenocarcinoma [35,36], whilst inactivation of Tp53 and Rb specifically in lung neuroendocrine cells resulted in small cell lung cancer [55]. Based on its anatomical location in the upper airways and the expression of BSC markers, lung SqCC is thought to arise from BSCs. Surprisingly, a recent study in genetically modified mice showed that overexpression of Sox2 in a Cdkn2ab/Pten null background could drive SqCC formation from BSCs, Club cells, or AT2 cells [56]. It remains to be seen whether multiple cells of origin are observed in other mouse models of lung SqCC, including mice with genetic backgrounds such as Lkb1-/-, Lkb1-/-/Pten-/-, or kinase-dead Ikkα [57–59]. In addition, such results may not be directly translatable in humans, given that mouse and human cells may not have the same degree of plasticity. Human tumours also carry much more genetic diversity than mouse cancer models, which is particularly relevant in lung cancer given its high levels of genomic instability [4]. To take into account the complexity of human cancers, computational comparison of normal cellular subset gene expression signatures to cancer subtypes has been used to gain insights into the cell of origin of human cancers [60,61]. Here we show that BSC gene expression signature closely resembles the human SqCC gene signature, suggesting that human lung BSCs are the candidate cells of origin of lung SqCC. An important caveat of such comparisons is the genetic signature of end-stage tumours may not fully represent the origin of the cancer. Our hypothesis will therefore need to be validated by introducing multiple genetic alterations in primary human lung cell subsets and determining their propensity for SqCC formation.
Invasive SqCC develops from preinvasive lesions in tobacco-smoking patients [62]. One-third of patients with basal cell hyperplasia will progress to carcinoma [63]; however, there are currently no biologic biomarkers to predict disease progression. Such biomarkers would greatly inform follow-up monitoring and perhaps enable early detection of invasive lesions. The discovery that BSCs are NHEJ competent and proliferate in response to DNA damage suggests that high levels of DNA-PKcs activation in cigarette smoking-induced basal cell hyperplasia may be a predictor of progression towards malignant disease. Assessment of the correlation between NHEJ activity in basal cell hyperplasia and progression to malignant disease would be necessary to validate this hypothesis. Advanced SqCCs are notoriously resistant to DNA-damaging agents [64]. Our data suggest that patients with strong expression of DNA repair genes such as PRKDC may benefit from therapy combining inhibitors of DNA repair and DNA-damaging agents. Overall, our study emphasizes the importance of fine-tuned control of DNA repair in stem/progenitor cells exposed to DNA-damaging agents, in which both unfaithful repair and failure to repair contribute to disease pathogenesis.
Adjacent normal lung specimens (confirmed by histology) were obtained through the Victorian Cancer Biobank from surgically resected tissue of lung cancer patients. Written informed consent was obtained from all patients by the Victorian Cancer BioBank prior to inclusion in the study, according to protocols approved by the Human Research Ethics Committee of the Walter and Eliza Hall Institute of Medical Research (WEHI) (approval #10/04). Patients were classified as current smokers (quit <10 y prior to surgery), exsmokers (quit >10 y), or never smokers (smoked less than 100 lifetime cigarettes). C57/Bl6 mice (8–12-wk-old males) were bred at the Walter and Eliza Hall Institute breeding facility, and SCIDprkdc mice (8–12-wk-old males) were obtained from the Animal Resource Centre (Australia). All animal experiments were approved by the WEHI Animal Ethics Committee (Approval #2013.028). Mice were maintained in our animal facilities according to institutional guidelines.
Lung tissue was classified as either large airway (LA, containing bronchi, cartilaginous airways, and attached alveolar tissue) or small airway (SA, containing bronchioles and attached alveolar tissue) and was processed either immediately or held intact for a maximum of 48 h at 4°C in DMEM/F12 media (Gibco) supplemented with 1 mg/mL of penicillin and streptomycin (Invitrogen). Samples were minced and then digested for 1 h at 37°C with 2 mg/mL collagenase (Worthington) and 200 U/mL deoxyribonuclease (Worthington) in 0.2% D-glucose (Sigma) in DPBS (Gibco). The cell suspension was strained through a 100 μm cell strainer and washed with 2% FCS-PBS, followed by red blood cell lysis to obtain a single-cell suspension.
Cells were blocked with rat immunoglobulin and CD16/CD32 FCγ II and III antibody (WEHI Monoclonal Antibody Facility) for 10 min at 4°C, followed by incubation with CD45-PE (HI30, BD Pharmingen), CD235a-PE (GA-R2, BD Pharmingen), CD140b-PE (28D4, BD Pharmingen), CD31-PE (WM59, BD Pharmingen), EpCAM-FITC (VU-1D9, Stem Cell Technologies), CD49f-PE-CY7 (GoH3, eBioscience), CD166-biotin (105902, R&D systems), and Podoplanin-APC (NC-08, BioLegend, also known as T1α) for 25 min at 4°C. The cells were then stained with streptavidin-APC-Cy7 (BD Pharmingen) before being washed and resuspended in 0.5 μg/mL propidium iodide. The cells were sorted on an Aria cytometer (BD Biosciences) using a 100 μm nozzle and processed immediately after sorting.
Human cells were prepared and stained as for sorting using LIVE/DEAD Aqua (Life Technologies) as a viability marker. Cells were fixed and permeabilised using the BD Fix/Perm kit (BD Biosciences) and stained with keratin-5 (polyclonal, Covance), MUC5AC (45M1, Thermo Scientific), acetylated-tubulin (6-11B-1, Sigma Aldrich), SCGB1A1 (polyclonal, Millipore), or phospho-histone H2A.X Ser139-BV421 (γH2AX, N1-431, BD Horizon) antibodies. Cells were stained where appropriate with either anti-rabbit Alexa594 or anti-mouse Alexa594 (Molecular Probes) before analysis on a LSR Fortessa (BD Biosciences). For 4′,6-diamidino-2-phenylindole (DAPI) nuclear content analysis, cells were subjected to an additional fixation step in Permeabilise Plus buffer according to the manufacturer’s protocol (BD Biosciences) and deoxyribonuclease treatment (Worthington) before staining with DAPI. All analyses were performed using FlowJo software.
Sorted cells were immediately fixed in 2.5% glutaraldehyde and postfixed in osmium tetroxide according to standard electron microscopy protocols. The cells were subsequently embedded in EPON Araldite resin. Ultra-thin sections were cut on a Leica UCT ultramicrotome, stained with lead citrate and uranylacetate, and imaged using a Gatan Ultrascan camera on a Hitachi H-7500 transmission electron microscope.
Freshly sorted cells were resuspended in DMEM/F12 (Gibco) supplemented with 1 mg/mL penicillin/streptomycin (Invitrogen), B27 (Gibco), 4 μg/mL heparin (Sigma Aldrich), 100 ng/mL EGF (Sigma Aldrich), insulin-transferrin-selenium (Gibco), 50 ng/mL human fibroblast growth factor-10 (R&D Systems), and 25 ng/mL human hepatocyte growth factor (R&D Systems), hereafter referred to as base media. A 50:50 matrigel (BD Biosciences):base media mix in a 96-well plate was allowed to set for 15 min at 37°C, before 2,000 cells were plated in media on top of the matrigel. Cells were grown for 14 d at 37°C in 5% CO2 and 5% O2 before colonies were photographed, counted, and processed for immunofluorescence studies.
Matrigel colonies were fixed with 2% paraformaldehyde for 10 min at RT. Colonies were permeabilised with 0.3% TritonX in PBS for 10 min at 4°C then rinsed in 100 mM glycine in PBS. Blocking was performed with 10% goat serum in immunofluorescence buffer (0.1% BSA, 0.2% TritonX-100, and 0.05% Tween20 in PBS) followed by primary antibody staining with keratin-5 (polyclonal, Covance) or pro-SFTPC (polyclonal, Millipore) antibodies. Secondary antibodies were anti-rabbit Alexa594 or Alexa488 (Molecular Probes). Counterstaining of nuclei was performed using DAPI (Sigma Aldrich). The colonies were mounted on SuperFrost slides before 3-D imaging with a laser-scanning confocal microscope (Zeiss LSM 780). All imaging analyses were performed using Fiji software.
Whole fragments of human lungs were either nonirradiated (control) or γ-irradiated at 6 Gy and left to recover in DMEM/F12 media supplemented with 1 mg/mL of penicillin and streptomycin for 1 or 24 h post irradiation at 37°C in 5% CO2 and 5% O2. At each time point, tissue portions were harvested and fixed in 10% neutral buffered formalin (Sigma) overnight at room temperature before paraffin embedding and sectioning for immunostaining. For mouse immunofluorescence studies, mice (C57Bl/6, SCIDprkdc) were exposed to 6 Gy of irradiation. Tracheas and lungs were harvested at 1, 4, 8, 24, or 96 h post irradiation and fixed/inflated in 4% paraformaldehyde in PBS pH 7.4 overnight at 4°C. Tracheas and lungs were then embedded in paraffin and sectioned for immunostaining. For mouse FACS analysis of cell cycle and apoptosis 24 h post IR, mice were irradiated and immediately injected with bromodeoxyuridine (BrdU, 50 mg/kg, Amersham). For FACS analysis of γH2AX expression, mouse lungs and tracheas were harvested 1 h or 4 h post irradiation, and single-cell suspensions generated as described below. This timing corresponds to 4 h or 7 h post irradiation including the time taken to generate single-cell suspensions.
Trachea and lungs from mice (C57/Bl6, SCIDprkdc) were harvested at 1, 4, 8, 24, or 96 h after bleomycin (Hospira) administration (IV 40 mg/kg) before fixation, sectioning, and immunostaining as described above. For FACS analysis of γH2AX expression, mouse lungs and tracheas were harvested 1 h post injection, and single-cell suspensions generated as described below. This timing corresponds to 4 h post bleomycin injection including the time taken to generate single-cell suspensions.
Lungs were minced and then digested in 2 mg/mL collagenase in 0.2% glucose-PBS for 45 min at 37°C. Red blood cells were lysed (0.64% NH4Cl) and cells filtered through a 40 μm cell strainer to obtain a single-cell suspension. Mouse lung cells were blocked as described for human cells and stained with CD45-PE-Cy7 (30-F11, BioLegend), CD31-PE-Cy7 (390, BioLegend), EpCAM-APC-Cy7 (G8.8, BioLegend), and CD104-FITC (346-11A, BioLegend). Alveolar cells were identified as CD45-CD31-EpCAMhiCD104lo as described previously [15]. Cells were then fixed and subjected either to Brdu/7-AAD staining (BD BrdU Flow Kit) according to the manufacturer’s instructions to identify apoptotic and proliferating cells or to γH2AX (20E3, Cell Signalling) staining.
Tracheal epithelial cells were isolated according to the protocol from Rock et al. [12]. Briefly, tracheas were cut into four pieces and incubated in 16 U/mL dispase (Roche) for 40 min at room temperature. Digestion was stopped with the addition of 5% FCS-DMEM (Gibco), and the epithelium peeled from the trachea. Epithelial sheets were washed and incubated in 2X trypsin-EDTA (Gibco) for 20 min at 37°C. Cells were then washed with 5% FCS-DMEM. Cells were blocked and stained with anti-NGFR (Abcam) or anti-T1α (clone 8.1.1, DHSB) antibodies for 30 min at 4°C. Cells were incubated with anti-rabbit Alexa Fluor488 or anti-hamster Alexa Fluor647 (Molecular Probes) for 15 min at 4°C. BSCs were identified as NGFR+ or T1α+ cells. Cells were then stained with BrdU, 7-AAD, or γH2AX (γH2AX, 20E3, Cell Signaling Technology).
Antigen retrieval was performed using citrate buffer (10 mM, pH 6) or high-pH antigen retrieval solution (Vector). Sections were blocked in 10% goat serum and incubated with antibodies overnight at 4°C followed by fluorophore-conjugated antibody for immunofluorescence or HRP-conjugated secondary antibodies (Vector) for immunohistochemistry. Antibodies used were phospho DNA-PKcs-S2056 (Abcam) and phospho-histone H2AX Ser139 (γH2AX, 20E3, Cell Signaling Technology) on human and mouse tissue; RAD51 (14B4, Genetex), T1α (clone 8.1.1, DHSB), and cleaved caspase 3 Asp175 (5A1E, Cell Signaling Technology) on mouse tissue; and T1α (NC-08, Biolegend) on human tissue. Nuclei were counterstained with DAPI where appropriate. Images were acquired using a DeltaVision fluorescence microscope (Applied Precision).
Human and mouse RNA was extracted from snap-frozen sorted cell pellets using a RNeasy Micro Kit (Qiagen), and DNase treatment was performed using the TURBO DNA-free Kit (Ambion) according to the manufacturer’s instructions. RNA sequencing was performed on an Illumina HiSeq at the Australian Genome Research Facility. Per human sample, 16–26 million 100 bp single-end reads were generated, and 13–17 million 100 bp single-end reads were generated per mouse sample. For human qPCR analyses, cDNA was generated using the SuperScript III system (Life Technologies) and subject to qRT-PCR using the Sensimix SYBR Hi-Rox kit (Bioline) on the Rotorgene RG-6000 (Corbett Research) under standard conditions. Three technical replicates were performed for each sample. Taqman gene expression assays were used for MUC5AC (Hs0087365_mH) and FOXJ1 (HS00230964_m1) using 18S (HS99999901_s1) or GAPDH (HS99999905_m1) as reference genes (Life Technologies). The sequence of the primers is available in S1 Table.
Human RNA-seq reads were aligned to the hg19/GRCh37 genome, and mouse reads were aligned to the mm10 genome using Rsubread [65]. Reads were assigned to Entrez gene IDs using featureCounts [66] and Rsubread’s in-built RefSeq annotation. The raw sequence data and read counts are available from GEO series GSE83492 (human) and GSE83991 (mouse). Filtering and normalization used the edgeR package [67]. Genes were filtered if their counts per million (CPM) values were above 1 in fewer than three samples for the human data and above 0.2 in fewer than three samples for the mouse data. Library sizes were normalized by the trimmed mean of M-values (TMM) method [68]. Multidimensional scaling (MDS) plots were produced using edgeR’s plotMDS function with the default settings. Distances between points on the MDS plots represent leading fold change, the root-mean-square log2-fold change for the 500 genes that best distinguish each pair of RNA samples.
Differential expression analyses used the limma package [69]. Counts were transformed to log2-CPM values with associated precision weights using the voom function [70]. Gene set tests were performed using roast rotation gene set testing [71]. Signature genes were defined for each normal cell population to be those genes that were consistently either up- or down-regulated in that population versus every other cell population of the same species. Differential expression was assessed for this purpose using limma’s treat function with fold change thresholds varying from 1 to 1.2 and a false discovery rate of 0.05 [72]. Larger fold-change thresholds were used for populations with more signature genes. A log-fold change was associated with each signature gene, being the log2-fold change for that gene between the population for which it is a signature and the next closest population.
Normalized microarray gene expression profiles for 261 lung cancers were downloaded from The Clinical Lung Cancer Genome Project (CLCGP) [33]. Gene symbols were converted to current official symbols with limma’s alias2SymbolTable function. Probes were filtered if their average log expression was in the bottom 50% or if no official symbol could be assigned. When more than one probe associated with the same gene, the probe with the highest average expression was retained. A signature score was computed for each normal lung cell population in each tumour profile: the signature scores were defined as Sum(wg yg) / Sum(|wg|), where yg is the log expression of the gene in the tumour and wg is the log-fold change of the gene between the normal populations. The sums were taken over all signature genes for the normal population. Signature scores were scaled to be between 0 and 1 for each normal population. Barcode plots were created using limma’s barcodeplot function. Correlation between normal expression signatures and cancer subtypes was assessed using rotation gene set tests, with 9,999 rotations and with the normal cell log-fold changes as gene weights.
Genewise RNA-seq read counts for 125 lung adenocarcinomas, 224 SqCC, and 54 normal lung samples from The Cancer Genome Atlas (TCGA) project were obtained from GEO series GSE62944 [73]. Genes were filtered if they failed to achieve 0.1 CPM in at least 54 samples. TMM normalization was applied, and differential expression between the cancer subtypes and normal samples was assessed using limma-voom and moderated t tests [70]. The proportion of the genome altered for each of the lung SqCCs (derived from somatic copy number information) was downloaded from the TCGA data portal (http://cancergenome.nih.gov).
Telomere lengths were determined as described previously [74]. Briefly, genomic DNA was extracted from snap-frozen FACS sorted human lung epithelial pellets or from 293T (ATCC) snap-frozen cell pellets using an Illustra Tissue and Cells GenomicPrep kit (GE Healthcare). The mastermix contained Sybr green I (0.75X, Life Technologies), AmpliTaq Gold (0.625 U, Life Technologies), Buffer II (1X, Life Technologies), MgCl2 (3 mM, Ambion), DTT (1 mM, Ambion), betaine (1 M, Sigma Aldrich), dNTPs (0.2 mM, Life Technologies), albumin primers (900 mM, IDT), and telomere primers (900 mM, IDT). The master mix was combined with 20 ng of experimental DNA and analysed in triplicate alongside a four-point 293T standard curve (50 to 1.85 ng of DNA) using albumin as the single-copy gene. PCR was performed on a Rotorgene RG-6000 using the following conditions: (15 min, 95°C) x 1, (15 s, 94°C − 15 s, 49°C) x 2, (15 s, 94°C − 10 s, 62°C −15 s, 74°C − 10 s, 84°C − 15 s, 88°C) x 32. Telomere lengths were calculated by determining the ratio of telomere signal (T) to single-copy gene signal (S) as determined from the 293T standard curve.
p-Values less than 0.05 were considered significant when conducting univariate tests. Error bars on plots represent mean ± SEM, and stars indicate significant differences in two-group comparisons: *p < 0.05, **p < 0.01, ***p < 0.001.
|
10.1371/journal.pcbi.1005923 | Modeling the genetic relatedness of Plasmodium falciparum parasites following meiotic recombination and cotransmission | Unlike in most pathogens, multiple-strain (polygenomic) infections of P. falciparum are frequently composed of genetic siblings. These genetic siblings are the result of sexual reproduction and can coinfect the same host when cotransmitted by the same mosquito. The degree with which coinfecting strains are related varies among infections and populations. Because sexual recombination occurs within the mosquito, the relatedness of cotransmitted strains could depend on transmission dynamics, but little is actually known of the factors that influence the relatedness of cotransmitted strains. Part of the uncertainty stems from an incomplete understanding of how within-host and within-vector dynamics affect cotransmission. Cotransmission is difficult to examine experimentally but can be explored using a computational model. We developed a malaria transmission model that simulates sexual reproduction in order to understand what determines the relatedness of cotransmitted strains. This study highlights how the relatedness of cotransmitted strains depends on both within-host and within-vector dynamics including the complexity of infection. We also used our transmission model to analyze the genetic relatedness of polygenomic infections following a series of multiple transmission events and examined the effects of superinfection. Understanding the factors that influence the relatedness of cotransmitted strains could lead to a better understanding of the population-genetic correlates of transmission and therefore be important for public health.
| Genomic studies of P. falciparum reveal that multi-strain infections can include genetically related strains. P. falciparum must reproduce sexually in the mosquito vector. One consequence of sexual reproduction is that parasites cotransmitted by the same mosquito are related to one another. The degree of genetic relatedness of these parasites can be as great as that of full-siblings. However, our understanding of the cotransmission process is incomplete, and little is known of the role of cotransmission in influencing population genomic processes. To help bridge this gap, we developed a simulation model to determine which of the steps involved in transmission have the greatest impact on the relatedness of parasites cotransmitted by a mosquito vector. The primary goal of this study is to characterize the outcomes of cotransmission following single or multiple transmission events. Our model yields new insights into the cotransmission process, which we believe will be useful for understanding the results from more complicated population models and epidemiological conditions. Such an understanding is important for the use of population genomics to inform public health decisions as well as for understanding of parasite evolution.
| Unlike most bacterial and viral pathogens, the malaria parasite P. falciparum, while predominantly haploid, must sexually reproduce in a mosquito vector before infecting a new human host. Sexual recombination has a significant impact on the population genomics of the parasite, and its effects depend on epidemiological conditions such as transmission intensity [1–3]. One outcome of sexual recombination is that parasites transmitted by a mosquito vector can be genetically related, which can be measured as the proportion of the genome that is identical-by-descent (IBD). IBD segments are region of the genome that originate from a recent common parental strain. A number of studies have used IBD to study transmission [4–8], survey antimalarial resistance [9], and detect signals of selection [10].
The effects of sexual recombination are also apparent in polygenomic (multi-strain) infections. Polygenomic infections can be formed through a series of infectious mosquito bites (superinfection) or through the transmission of multiple strains from a single mosquito bite (cotransmission) [5,7,11]. Coinfecting strains resulting from superinfection are assumed to be unrelated while those resulting from cotransmission are assumed to be genetically related [5–7]. While superinfection is believed to be common in high transmission settings, owing to high entomological inoculation rates and complexity of infections (COI, the number of strains per infection) [12,13], the frequency with which cotransmission occurs is less clear. Studies of genetic relatedness in symptomatic polygenomic infections reporting to clinics in mid-to-low transmission settings show that cotransmission is prevalent in these regions [5–8], but little is known of the frequencies of cotransmission and superinfection across transmission settings. Genetic relatedness studies reveal a large amount of variation in the relatedness of polygenomic infections. The fact that sexual recombination occurs within the mosquito suggests that the relatedness in these polygenomic infections is associated with transmission. High relatedness in polygenomic infections could be indicative of serial cotransmission chains [14], but it is unclear what other factors may influence the relatedness of polygenomic infections.
Part of the uncertainty stems from an incomplete understanding of the cotransmission process. When a female Anopheline mosquito bites an individual infected with malaria, she ingests male and female gametocytes. The ingestion of these gametocytes activates them to form gametes that fuse to create a diploid zygote. Gametes can fuse with other gametes of the same genotype, resulting in self-fertilization (selfing), or can fuse with gametes from other genotypes resulting in outcrossing. The zygote undergoes meiosis and develops into a motile ookinete that traverses the midgut epithelial layer and forms an oocyst. Within the oocyst, the parasite undergoes many rounds of mitosis to create thousands of haploid sporozoites. These sporozoites travel to the mosquito salivary glands and are stored until deposited by the mosquito into the human host during a blood meal. Only those sporozoites that invade the liver will survive to continue the malaria life cycle. How then could variation in within-host and within-vector transmission dynamics, such as the number of oocysts formed and the number of sporozoites infecting the liver, affect the relatedness of cotransmitted strains, and how could these variables in turn affect the relatedness of polygenomic infections in natural populations?
To address the complexity of this transmission cascade and better understand the process of cotransmission, we devised a classification framework based on parasite pedigrees and kinships to develop an understanding of how the various sampling and mating events within the mosquito vector affects the relatedness of transmitted sporozoites. We then created a transmission model to quantify the relatedness of cotransmitted strains under a variety of within-host and within-vector dynamics and used this model to examine the relatedness of polygenomic infections in transmission chains. Our study reveals new insights into the cotransmission process, which we believe will be useful for the interpretation of population genomic signals obtained from more complicated population-level models or from natural populations.
To simulate sexual recombination, we developed a P. falciparum-specific meiosis model based on the whole genome sequences of 69 genetically distinct progeny derived from 3 previously generated P. falciparum crosses involving different laboratory-adapted strains (3D7, HB3, Dd2, 7G8, and GB4) [15–19]. The whole genome sequences generated from these crosses are one of best sources of data for designing a P. falciparum-specific meiosis model because the genotypes of the parental strains are known. Furthermore, we can be confident of the number of sexual reproduction cycles separating progeny and parental strains. While previous IBD analyses of parasites from natural parasite populations have identified putative F1 progeny [5,7,20,21], having complete knowledge of parental ancestry simplifies the identification of IBD segments and allows us to better identify recombination events throughout the genome. We calculated the number of crossover events and inter-crossover distances (S1 Fig & S1 Table) using a hidden Markov model (HMM) [4,22] to identify IBD segments shared between progeny and parental strains (Methods). We then used this data to test the fit of two different meiosis models, one with and one without obligate chiasma formation. Both were based off the gamma model of crossover formation, which has been used to characterize recombination events in a wide variety of taxa, including H. sapiens, D. melanogaster, and S. cerevisiae [23–26]. The gamma model is an improvement over simpler Poisson-based crossover models because it allows us to explore a wide range of crossover interferences.
Regardless of whether obligate chiasma formation was modeled, the number of crossover events and intercrossover distances in our simulated meiotic events resembled those of the laboratory-crossed progeny (Fig 1A and 1B). However, both meiosis models underestimated the frequency of short intercrossover distances (< 50 cM) (Fig 1A), which we suspect is because our HMM overestimated the frequency of short intercrossover distances in the laboratory-cross data (S2 Fig). We found that the obligate chiasma model generated crossover events that were more consistent with that of the laboratory-crossed progeny, but overestimated the number of chromosomes with two crossover events. Using a pseudo-likelihood function (Methods), we determined that an obligate chiasma model fit the data better than a non-obligate chiasma model (Fig 1C). However, we could not estimate the level of crossover interference. Because crossover interference is observed in a wide-variety of organisms spanning multiple taxa [23], we chose to use an obligate chiasma meiosis model with a weak level of interference (gamma distribution with shape = 2, scale = 0.38) for all of our transmission simulations.
We then designed a transmission model that partitions transmission into three steps: 1) The host-vector sampling of gametocytes from an initial host infection 2) the sequence of events starting from gamete fusion and meiosis to the development of the oocyst within the mosquito vector, and 3) the vector-host injection of sporozoites and subsequent invasion of the liver to determine the genetic composition of the next human host (Fig 3). We initiate our model by simulating a mosquito blood-feeding event on a polygenomic infection comprised of unrelated strains and parameterized by 1) COI, 2) oocyst count, and 3) the infected hepatocyte count. The number of unique strains present in the initial infection is determined by COI. In our model, we consider oocyst formation as the final outcome of gamete fusion and subsequent meiosis. Based on the oocyst count, our model samples gamete pairs, which fuse and undergo meiosis to create an oocyst consisting of four unique meiotic products. Competition within the oocyst is not modeled and we assume that each meiotic product is present at equal proportion in the oocyst. After all oocysts are created, the model samples sporozoites according to the infected hepatocyte count to determine the genetic composition of the subsequent host infection. If the resulting infection harbors multiple strains, we calculated the relatedness of cotransmitted strains as the average pairwise relatedness between each of the unique genotypes present in the final host infection.
The values for the infected hepatocyte count are pre-specified and drawn from the set {1, 2, 3, 4, 5, 10, 20}. Simulations with COI = 1 were excluded because they always result in selfing and the transmission of genetic clones. Simulations with an infected hepatocyte count = 1 were also excluded, as they cannot result in cotransmission. Small values are overrepresented to reflect the right-skewed distributions of oocyst counts observed in mosquito feeding assays and infected hepatocyte counts estimated from a malaria-challenge study [27–29]. These values also include the COI observed in naturally occurring polygenomic infections from mid-to-low endemic settings (COI ranging from 2–6 in polygenomic infections).
From our pedigree/kinship framework, we knew that sporozoites sampled from a single oocyst would be either genetic clones or meiotic siblings. Our transmission simulation confirmed this prediction and found that the expected relatedness of cotransmitted strains in single-oocyst transmission simulations was always 0.33 (Fig 4), which is the expected relatedness of genetically distinct meiotic siblings. In single oocyst transmission simulations, cotransmission can only be achieved by the transmission of two or more genetically distinct meiotic siblings. The distinction between genetically distinct and genetically identical meiotic siblings is relevant in the context of cotransmission, as the transmission of clonal meiotic siblings cannot result in cotransmission. Changes to the infected hepatocyte count do not affect the expected relatedness values, but higher infected hepatocyte counts caused the distribution to be more concentrated around the mean.
In multiple oocyst transmission simulations, the relatedness of cotransmitted strains is not as easy to predict, since multiple kinships can be transmitted. Based on our pedigree/kinship framework, we hypothesized that COI modulates the expected relatedness of cotransmitted strains by limiting the transmission of half-siblings and unrelated strains; the transmission of half-siblings and unrelated strains described by pedigrees 6 are only possible when COI ≥ 3. The transmission of unrelated strains described by pedigree 9 only applies when COI ≥ 4.
Our transmission simulations confirmed these predictions and revealed a simple relationship between COI, oocyst count, and the relatedness of cotransmitted strains (Figs 5 and 6): the relatedness of cotransmitted strains declines with increasing COI. All COI = 2 simulations have an expected relatedness > 0.33, with a larger increase in high oocyst count simulations. The increase in relatedness is a reflection of the increased transmission of full-siblings and parent-offspring strains. When COI = 3, increasing oocyst counts no longer increased the expected relatedness of cotransmitted strains due to the additional transmission of half-siblings. Once COI > 4, increasing oocyst counts decreased the expected relatedness of cotransmitted strains. This was due to the increased transmission of unrelated strains, particularly those described by pedigree 9 (outcrossed oocysts that do not share any parental strains) (Fig 6C and 6D). When COI = 20, the majority of transmitted parasites are either meiotic siblings or unrelated strains described by pedigree 9.
We found that different infected hepatocyte counts altered the distribution of relatedness (S4 Fig & S5 Fig) but had no effect on the trends established by either COI or oocyst count. Again, simulations with a COI = 2 consistently had the highest expected relatedness values while simulations with higher COIs had lower expected relatedness values, regardless of the infected hepatocyte count.
Thus far, our simulations have assumed that the strains making up polygenomic infection are present and sampled in equal proportions. However, strain proportions in natural polygenomic infections can be highly skewed. Furthermore, different strains can have different transmissibility relating to factors such as gametocyte production. To investigate how skewed gametocyte sampling probabilities could affect the relatedness of cotransmitted strains, we devised a weighted sampling scheme defined by the ratio of the most frequent to the least frequent strain in the infection (Methods).
Predictably, skewing the gametocyte strain ratios increased the rate of selfing and the transmission of genetic clones (S6 Fig). Skewed ratios of up to 10:1 increased relatedness of cotransmitted strains by a small amount. Ratios ranging from 1:1 to 10:1 increased the expected relatedness of cotransmitted strains by 0.01–0.10. This increase depended on both COI and the magnitude by which strains proportions differed. The relatedness of cotransmitted strains from high COI infections was more robust to differences in strain proportions; a 10:1 ratio in a COI = 20 infection increased relatedness by only 0.02 while a 10:1 ratio in a COI = 3 infection increased relatedness by 0.03–0.06.
The genetic composition of natural polygenomic infections can result from multiple transmission events and influenced by population-level transmission dynamics. However, developing a model that take into account all possible population-level transmission dynamics is beyond the scope of this paper. Instead, we used our model to quantify the relatedness of polygenomic infections in three different multiple transmission simulations, which we refer to as transmission lineages. Each transmission lineage is designed to resemble transmission chains that occur in natural populations and initiated by simulating a mosquito blood-feeding event on a polygenomic infection comprised of unrelated strains. The first transmission lineage does not allow superinfection; all subsequent transmission events in the chain must infect uninfected hosts. The second and third transmission lineages allow superinfection and are differentiated by the nature of the resident strain in the soon-to-be superinfected host. For the second transmission lineage, the resident strain is identical to one of the parental strains in the initial polygenomic infection (resembling natural backcrossing events). For the third transmission lineage, the resident strain is not related to any of the parental strains in the initial polygenomic infection but is the same in all transmission events. In the last transmission lineage, the resident strain is not related to any of the parental strains in the initial polygenomic infection and is different in all transmission events.
For our transmission lineage simulations, we modified our cotransmission model so that oocyst and infected hepatocyte counts are determined by randomly sampling from distributions reflecting those of found in previous studies [29,30]. Subsequent transmission events sample parasites from the infection generated by the previous transmission event. Allowing oocyst and infected hepatocyte counts to be chosen from these distributions did not affect the previously observed relationship between COI and the relatedness of cotransmitted strains (S7 Fig). The relatedness of cotransmitted strains following single cotransmission events from infections COI = 2 had an expected relatedness greater than 0.33 while those with a COI > 3 had an expected relatedness less than 0.33.
As expected of serial cotransmission chains, we found that the relatedness of polygenomic infections increases with each transmission event (Fig 7). Transmission lineages with superinfection had lower relatedness values and smaller proportions of serial transmission simulations that converged to the transmission of single strains. The reduction in relatedness was greatest in those where the resident strain was unrelated to the parental strains of the original infection (Fig 7, purple). Changing the resident strain after each transmission event prevented the relatedness of polygenomic relatedness from increasing beyond 0.10 even after five transmission events. We also saw that the COI of the initial infection could have a lasting effect on the relatedness of polygenomic infections. Transmission lineages initiated with low COI polygenomic infections had higher relatedness values than those initiated with high COI polygenomic infections. This effect was weaker in superinfection lineages with unrelated resident strains. While skewed gametocyte-sampling ratios had a modest effect on the relatedness of polygenomic infection, it drastically increased the rate with which transmission lineages converged to the transmission of single strains for all transmission lineages except the one where unrelated resident strains were changed after each transmission event (Fig 7B and 7D).
Parasite strains in polygenomic infections are often genetically related, but it is unclear why there is so much variation between infections or whether the relatedness of polygenomic infections can be used to understand parasite transmission. In order to help bridge the gaps in our understanding, we developed a pedigree/kinship framework for understanding how COI and oocyst counts affect the relatedness of cotransmitted strains. We then tested the predictions of this framework using a parasite transmission model to quantify changes in the relatedness of cotransmitted strains. We demonstrated that multiple oocyst simulations in low COI conditions favor the transmission of full-siblings / parent-offspring strains and limit the transmission of half-siblings and unrelated strains, causing an increase in the expected relatedness of cotransmitted strains. Multiple oocyst simulations in high COI conditions decrease the relatedness of cotransmitted strains by favoring the transmission of half-siblings and unrelated strains. Alterations to the number of sporozoites that invade the liver have little effect on relatedness, conditioned on the fact that multiple sporozoites invade.
We also examined how non-uniform gametocyte-sampling probabilities could affect the relatedness of cotransmitted strains. Previous studies have established that intra-host parasite dynamics depend on patient age [31,32] disease severity (reviewed in [33]), and eco-epidemiological factors such as seasonal transmission [34,35]. These dynamics are strongly influenced by host immunity [36] and can fluctuate over the course of a single infection [32,37–40]. Furthermore, gametocyte sampling is not completely random [41] and not reliant on peripheral blood gametocyte densities at low parasitemias [34,42]. Our results show that the relatedness of cotransmitted strains is robust to variations in intra-host strain proportions and gametocyte-sampling probabilities. Even infections where the ratio of the most frequent to least frequent strain is 10:1 do not result in drastic changes to that observed from infections with even strain proportions. This suggests that the relatedness of cotransmitted strains is consistent across differences in patient-age, disease severity, and host immunity.
Our results are in agreement with the frequent assumption that cotransmission events are comprised of genetically related parasite strains [5–8]. A large fraction of simulated cotransmission events result in the transmission of genetically distinct meiotic siblings, as evidenced by the peaks at 0.33 for all simulations where oocyst counts and hepatocyte counts were randomly sampled. However, we also found that the transmission of unrelated strains is a major aspect of cotransmission. The cotransmission of unrelated strains was present in all multiple oocyst simulations and increased in frequency with COI. Polygenomic infections comprised of unrelated strains are typically assumed to be the result of superinfection, but these findings suggest that some are the result of cotransmission. Current estimates of the prevalence of cotransmission are underestimates, since they rely on the subset of cotransmission events resulting in polygenomic infections comprised of genetically related strains [7].
Our results reveal an inverse relationship between the relatedness of cotransmitted strains and COI. COI is correlated with high entomological inoculation rates [43,44] and a known genetic correlate of transmission intensity [43,44]. COI is higher in high transmission areas than in low transmission areas due to increased superinfection rates. The association between the relatedness of cotransmitted strains and COI suggests that polygenomic infections in low transmission areas are comprised of more related strains than those in high transmission areas. We previously found that the average relatedness of 32 symptomatic polygenomic patients collected from a clinic in a low transmission region of Senegal (mean COI of two) was 0.38 [7]. This value exceeds the expected relatedness of meiotic siblings and may reflect an increase in the transmission of full-siblings / parent-offspring parasites but could also result from factors such as population structure. Previous studies of genetic relatedness have focused on areas of mid-to-low transmission setting [5–8] and a comparison of genetic relatedness of polygenomic infections across transmission settings have yet to be performed. High relatedness from low COI infections could have implications for the spread of drug resistance traits in low transmission settings, as the increased relatedness could increase the chance that multi-locus drug resistant genes are passed on together to the next generation.
It remains to be seen whether the relationship between relatedness and COI can be reflected in polygenomic infections collected from natural parasite populations. If the inverse relationship between COI and relatedness holds, then the relatedness of coinfecting strains could be a potential population genetic correlate of transmission intensity. Population genetic correlates of transmission are valuable in the context of malaria control and can be used to supplement or supplant traditional epidemiological measures, which can be difficult to collect in low transmission areas [44,45]. With regards to polygenomic infections, only the frequency and COI of polygenomic infections are known to correlate with transmission intensity [4,44,46]. Other population genetic metrics, such as parasite clonality [44], currently rely on data obtained from monogenomic infections, which are limited in high transmission areas where polygenomic infections are frequent. By providing an additional source of information, genetic relatedness could increase the granularity by which we use genetic signals to monitor changes in transmission. However, spatial-temporal transmission, such as the seasonality or the existence of transmission hotspots, and host immunity can influence population genetic structure [36]. Neither of these are taken into consideration in this study, and it is unclear how these might affect polygenomic relatedness. Population-level models and epidemiological sampling will be needed to understand the effects of cotransmission and establish whether the relatedness of polygenomic infections correlates with transmission intensity.
An alternative method of dissecting population-level dynamics is to focus on the characterization of transmission lineage. Transmission lineages consist of chained transmission events and are a simplification of the transmission processes within populations. Our transmission lineages were designed to examine the effect of multiple transmission events and to examine how the co-occurrence of superinfection affects the relatedness of polygenomic infections. They show that superinfection depresses the relatedness of polygenomic infections, but also show how sensitive these lineages are to the conditions of the host infection. Strikingly, they show that cotransmission fails to increase the relatedness of polygenomic infections if each host in the transmission chain harbors a different, genetically unrelated parasite strain. They also reveal the fragility of serial cotransmission chains. In the absence of superinfection, serial cotransmission chains quickly converge to the transmission of single strains. High COI in the initial infection delays this process but a large fraction of serial cotransmission chains still converge within five transmission events. Because these transmission lineages are analogous to the introduction of a polygenomic infection to a new population, polygenomic relatedness could be useful for studying transmission in import scenarios.
In conclusion, our study uses a model of parasite transmission to provide mechanistic insight into the process of cotransmission to help understand the factors that influence the relatedness of cotransmitted strains. Understanding the effects of sexual recombination and transmission on malaria population genomics is of key public health interest in an era where parasite populations are experiencing rapid declines in transmission intensity. We believe mechanistic models such as the one used in this study reveal new insights that can be applied to the results obtained from more complicated conditions. Our model highlights the importance of COI in influencing the relatedness of cotransmitted strains, but future models and epidemiology studies are needed uncover how transmission intensity and cotransmission affects the genetic composition of strains in polygenomic infections in natural populations. These models should incorporate background parasite population structure and genetic diversity to understand the effects of cotransmission and establish whether the relatedness of polygenomic infections correlates with transmission intensity. Such models will rely on genetic data collected from well-characterized epidemiological settings to determine whether the relatedness of polygenomic infections is a potential population genetic correlate of transmission.
We simulated meiosis under two different frameworks: one with and one without obligate chiasma formation. Both frameworks sample from a constrained gamma distribution where the average distance between randomly sampled distances is 50 centimorgans to determine the location of chiasma along a bivalent [47,24]. For each placed chiasma, our meiosis model chose one sister chromatid from each homolog to undergo recombination. Sister chromatids were independently chosen for each recombination event. Once all recombination events were complete, the model independently segregated and randomly combined sister chromatids from other bivalents to create haploid parasite genomes.
For the non-obligate chiasma framework, our meiosis model placed the first chiasma 105 base pairs before the beginning of each chromosome. It then drew a distance, d, from a gamma distribution with shape = v and scale = 1/(2v) [47] to determine the location of the next chiasma. New chiasma were placed d units after the previous chiasma and a new distance was drawn for each chiasma. Chiasma locations were filtered to include only those that fell within the boundaries of the chromosome under consideration.
For the obligate chiasma framework, the position of the first chiasma was determined by drawing from a uniform distribution that spans the length of each chromosome. Subsequent chiasma were placed by drawing distances from a constrained gamma distribution (described in the next paragraph) and placing the next chiasma d units before it. This was repeated until the start of the chromosome was reached. Afterwards, the process was repeated in the other direction until the end of the chromosome was reached.
Due to the forced placement of chiasma, we could not use the formulas used in the non-obligate chiasma framework to generate appropriately constrained gamma distributions. We used an approximate Bayesian computation (ABC) Markov chain Monte Carlo (MCMC) to solve the appropriate scale parameter and shape parameters. Shape parameters varied from 1–9 and scale parameters were sampled from a uniform distribution with a range of 0–5. For each set of scale and shape parameters, we counted the number of chiasma on a bivalant 100 centiMorgans (cM) in length and repeated this process 1000 times to estimate the average and standard deviation. We evaluated the fit of each proposed parameter using the following distance metric:
D′=(2−u)20.052+δ2
where u and δ are the simulated mean and standard deviations of the number of chiasma, 2 represents the desired number of chiasma per 100 centiMorgans, and 0.05 represents a small error term. We then constructed an estimate of the pseudo-likelihood as:
L=1eD′
The proposed scale parameter was accepted if the proposed pseudo-likelihood was greater than the pseudo-likelihood of the previously proposed parameter. If the new pseudo-likelihood was smaller, then the probability of rejection was decided by the ratio of the current pseudo-likelihood over the previous pseudo-likelihood. This process was repeated 2,500 times to form a MCMC chain. After our MCMC chain was completed, we calculated the mean of the accepted scale parameters from the last 1500 steps to serve as our estimate of the scale for each shape parameter.
We calculated the average number of crossover events and intercrossover distances for each chromosome in the genome using SNP data from 69 genetically distinct progeny generated from 3 different laboratory crosses [15,17–19]. These data were previously generated by the Pf3k project (https://www.malariagen.net/projects/pf3k) [15–19]. VCF files were downloaded and filtered based on the available INFO strings. We removed non-Mendelian sites, sites that did not pass the quality filters used, and sites that were invariant between the parental strains used in the cross. Samples from each laboratory cross were represented by an average of 1028 SNPs. From this filtered dataset, we performed pairwise calculations of percent similarity to identify and remove duplicate strains. Duplicate strains were defined as those having greater than 90% SNP similarity.
For each chromosome, we used a modified version of an IBD Hidden Markov Model (HMM) [4,22] to quantify the average number of crossover events and the average intercrossover distance for each chromosome. Our previously published HMM relied on population SNP frequencies to infer IBD, which is problematic when using cultured strains with vague demographic histories. For each laboratory cross, we used SNP data to infer IBD between progeny and parental strains using the following emission probabilities:
P(Concordance|IBD)=(1−ε)2+(ε)2
P(Concordance|non−IBD)=2ε(1−ε)
P(Discordance|IBD)=2ε(1−ε)
P(Discordance|non−IBD)=1−2ε(1−ε)
where ε refers to the rate of sequencing error, concordance refers to having the same SNP identity, discordance refers to having different SNP identities, and IBD refers to identical-by-descent.
The resulting IBD maps closely mirror the parental inheritance boundaries specified in [16], but sometimes identifies very short IBD fragments that are unlikely to be real (S5 Fig). Crossover events were identified as the points in the chromosome where the IBD map switches from IBD to non-IBD and intercrossover distance was calculated as the distance (in cM) between each of the identified crossover points. Intercrossover distances were converted to centiMorgans using the estimates reported in [15,16] (15 kb/cM). If no crossovers were observed, then the intercrossover distance was defined as the length of the entire chromosome.
We then used the average number of crossovers and intercrossover distances to determine whether a non-obligate or obligate chiasma model of meiosis would fit the data better. Each simulation was run 20 times to get an average and standard deviation of the number of crossover events and crossover distances per chromosome. We then devised a distance metric defined as:
Dj=∑i14(ui,sim−ui,observed)2δ2i,sim+δ2i,observed
where u is the mean, δ is the standard deviation, j is the feature (number of crossover events or intercrossover distance), i is the chromosome number, sim indicates the simulation result, and observed indicates the value observed in the 69 progeny strains. We defined a pseudo-likelihood as
L=∏j21eDj
and used it to determine the model that fit the data better.
To quantify the average relatedness of cotransmitted strains, we developed an agent-based mosquito transmission model that simulates the sampling processes that occur as parasites enter and exit the mosquito vector and parameterized by COI, oocyst count, and infected hepatocyte count. The values for oocyst count and infected hepatocyte count were drawn from the set {1, 2, 3, 4, 5, 10, 20} while the values for COI were drawn from the set {2, 3, 4, 5, 10, 20}. Each set of parameters was run 2000 times. Each simulation was initiated by creating an initial infection comprised of unrelated parasite strains; the number of strains within the initial infection was determined by COI.
To model differences in intra-host strain proportions and differences in sampling probabilities, we assumed that strain proportions followed an exponential equation of the form:
f(x)=AeBx
where x is a discrete variable representing each strain in the infection. We used an exponential equation to magnify the difference in frequency between the most frequent strain and the other strains present in the infection.
For an infection with COI = n, x ranges from 0 to n -1. We fit this equation to two points, (0, f(0)) and (n—1, f(n—1)), based on the ratio of the most frequent to the least frequent strain in the infection. These ratios ranged from 1:1 to 10:1, reflecting the observed strain proportions in a set of polygenomic infections collected from Thiès, Senegal (S8 Fig). f(0) is the ratio of the most frequent to the least frequent strain. f(n—1) is the ratio of the final strain to the least frequent strain and always equal to one. The ratios of all other strains present in the infection was determined by f(1), f(2), …f(n-1). We then drew from a Dirichlet distribution with a concentration parameter = {f(0), f(1), f(2), … f(n—1)} 1000 times to calculate the expected frequency of each strain in the infection.
Based on the specified oocyst count, our model sampled gametocyte pairs by their intra-host strain proportions to create oocysts, allowing for multiple samplings of the same strain. Each sample pair underwent meiosis to create four meiotic products. The progeny from all the meiotic events were combined without the removal of repeat strains to represent the sporozoites within the mosquito vector. Our model assumed mating success and oocyst formation could be simulated as the random sampling of gametocytes from the human host. It is unclear whether the parasite has a preference for self-fertilization or outcrossing. Evidence for non-random mating is based on the observation of highly inbred oocysts within the mosquito midgut [48], but it is unknown to what extent self-fertilization occurs more frequently than expected by chance.
We then sampled sporozoites to represent the strains in the infected hepatocytes. Multiply-infected hepatocytes were not allowed. At this point, our model performed pairwise comparisons between all the parasites in the infected hepatocytes, regardless of whether or not the pair consisted of genetically distinct parasites, to determine the frequency of the different pedigrees specified in Fig 2. The expected relatedness of cotransmitted strains was calculated as the average pairwise relatedness between genetically distinct strains. This average is not weighted by the frequency of strains within the infected hepatocytes. Because cotransmission must result in the creation of polygenomic infections, we excluded infections where the infected hepatocytes consisted of a single strain. When an infected hepatocytes consisted of two or more genetically distinct strains, the relatedness of cotransmitted strains was calculated as the relatedness between the two strains; when an infection was comprised of 20 genetically distinct strains, the relatedness of cotransmitted strains is calculated as the average pairwise relatedness from all 20-choose-2 comparisons.
Source code is available on GitHub, under the project name Cotransmission (https://github.com/weswong/Cotransmission). The code is written using Python 2.7.0 and is platform independent.
We defined relatedness as the proportion of the genome that is identical-by-descent (IBD) owing to inheritance from the same common ancestor. Because the genetic ancestry of all input strains was known and assumed to be genetically unrelated, IBD segments were identified as segments of the genome that originated from the same parental input strain.
To calculate the expected relatedness of parasites described by our 9 pedigrees, we generated simulations with the appropriate number of oocysts (1 or 2), the appropriate pedigreess for each oocyst, and the appropriate method of sampling parasite pairs (within or between oocysts) for each pedigree and quantified the relatedness of a single randomly drawn parasite pair. This process was repeated 800 times to generate distributions of relatedness and to get an estimate of the mean.
|
10.1371/journal.pgen.1003350 | A Novel Mutation in the Upstream Open Reading Frame of the CDKN1B Gene Causes a MEN4 Phenotype | The CDKN1B gene encodes the cyclin-dependent kinase inhibitor p27KIP1, an atypical tumor suppressor playing a key role in cell cycle regulation, cell proliferation, and differentiation. Impaired p27KIP1 expression and/or localization are often observed in tumor cells, further confirming its central role in regulating the cell cycle. Recently, germline mutations in CDKN1B have been associated with the inherited multiple endocrine neoplasia syndrome type 4, an autosomal dominant syndrome characterized by varying combinations of tumors affecting at least two endocrine organs. In this study we identified a 4-bp deletion in a highly conserved regulatory upstream ORF (uORF) in the 5′UTR of the CDKN1B gene in a patient with a pituitary adenoma and a well-differentiated pancreatic neoplasm. This deletion causes the shift of the uORF termination codon with the consequent lengthening of the uORF–encoded peptide and the drastic shortening of the intercistronic space. Our data on the immunohistochemical analysis of the patient's pancreatic lesion, functional studies based on dual-luciferase assays, site-directed mutagenesis, and on polysome profiling show a negative influence of this deletion on the translation reinitiation at the CDKN1B starting site, with a consequent reduction in p27KIP1 expression. Our findings demonstrate that, in addition to the previously described mechanisms leading to reduced p27KIP1 activity, such as degradation via the ubiquitin/proteasome pathway or non-covalent sequestration, p27KIP1 activity can also be modulated by an uORF and mutations affecting uORF could change p27KIP1 expression. This study adds the CDKN1B gene to the short list of genes for which mutations that either create, delete, or severely modify their regulatory uORFs have been associated with human diseases.
| Gene expression can be modulated at different steps on the way from DNA to protein including control of transcription, translation, and post-translational modifications. An abnormality in the regulation of mRNA and protein expression is a hallmark of many human diseases, including cancer. In some eukaryotic genes translation can be influenced by small DNA sequences termed upstream open reading frames (uORFs). These elements located upstream to the gene start codon may either negatively influence the ability of the translational machinery to reinitiate translation of the main protein or, much less frequently, stimulate protein translation by enabling the ribosomes to bypass cis-acting inhibitory elements. CDKN1B, which encodes the cell cycle inhibitor p27KIP1, includes an uORF in its 5′UTR sequence. p27KIP1 expression is often reduced in cancer, and germline mutations have been identified in CDKN1B in patients affected with a syndrome (MEN4) characterized by varying combinations of tumors in endocrine glands. Here we show that a small deletion in the uORF upstream to CDKN1B reduces translation reinitiation efficiency, leading to underexpression of p27KIP1 and coinciding with tumorigenesis. This study describes a novel mechanism by which p27KIP1 could be underexpressed in human tumors. In addition, our data provide a new insight to the unique pathogenic potential of uORFs in human diseases.
| CDKN1B encodes the cyclin-dependent kinase (CDK) inhibitor, p27KIP1, which negatively regulates the Cdk2/cyclin E and Cdk2/cyclin A protein complexes, thereby preventing the progression from the G1 to the S phase of the cell cycle [1]. In G0 and early G1, p27KIP1 expression and stability are maximal. During the G1 phase gradual degradation of p27KIP1 is associated with an increased activity of Cdk2/cyclin E and Cdk2/cyclin A complexes to stimulate cell proliferation [2], [3]. Several mitogenic (i.e. MAPK, PI3K/AKT) and anti-proliferative (i.e. TGFβ/SMAD) signal transduction pathways regulate p27KIP1 expression and activity, making it a central integration point for cell-fate decision [4]. These pathways can regulate p27KIP1 at different levels, including transcription, translation, intracellular localization or ubiquitin-mediated proteasomal degradation [5].
p27KIP1 acts as an atypical tumor suppressor as it is rarely mutated in human cancers, but frequently underexpressed or mislocalized in human malignancies [4]. Although an augmented proteolysis was initially suggested as the major cause of p27KIP1 loss in human tumors [6], recent findings propose that reduced translation and/or transcription of CDKN1B also contributes to p27KIP1 deficiency [7]–[9].
Translation of CDKN1B may involve regulatory elements within its 5′UTR, including an internal ribosome entry site (IRES) and an upstream ORF (uORF) [10], [11]. The IRES supports p27KIP1 expression when cap-dependent translation is reduced, such as during quiescence or stress conditions [10], [12]. Reduced IRES-mediated translation, due to mutations in the pseudouridine synthase that alters the ribosome's ability to efficiently engage the CDKN1B IRES element, may contribute to the increased predisposition to cancer in X-linked congenital dyskeratosis [13].
Germline mutations in the CDKN1B gene have been recently associated with the development of a multiple endocrine neoplasia syndrome both in humans (MEN4, MIM 610755) and in rats (MENX) [14]. Multiple endocrine neoplasias, including type 1 (MEN1, MIM 131100) and type 2 variants, (MEN2, MIM 171400, MIM 162300), are a group of autosomal dominant syndromes characterized by varying combinations of tumors affecting at least two endocrine organs [15].
To date, seven CDKN1B germline mutations have been identified in MEN4 patients primarily associated with MEN1-related lesions, including parathyroid and pituitary tumors, but the presence of other malignancies such as renal angiomyolipoma, papillary thyroid carcinoma and pancreatic masses has also been reported [8], [9], [14], [16], [17]. Two further germline mutations have been more recently associated with sporadic hyperparathyroidism [18]. In MEN4, CDKN1B mutations either affect p27KIP1 cellular localization, protein stability or the binding with functional partners such as Cdk2 or Grb2 [8], [17]. Reduced transcription/translation efficiency due to mutations in elements regulating translation initiation (i.e., in the Kozak sequence, or forming a secondary stem loop structure within the CDKN1B 5′UTR), has also been described [8], [9].
Germline CDKN1B mutations are hence rare events in MEN1-like subjects (individuals with MEN1-related lesions, without MEN1 inactivating mutations), being identified in less than 3% of cases [8], [16] and a clear genotype-phenotype correlation has not been established to date.
In the present paper we analyzed the CDKN1B gene looking for point mutations and large rearrangements in order to determine the possible cause of multiple endocrine tumors in 25 consecutive sporadic and familial patients with typical MEN1-related symptoms. We identified a 4-bp deletion that modifies the regulatory uORF in the 5′UTR of the CDKN1B gene in a patient with tumors in the pituitary gland and the endocrine pancreas. Functional studies based on dual-luciferase assay and site-directed mutagenesis further support the deleterious influence of this deletion on translation reinitiation at the CDKN1B starting site, with a consequent reduction of p27KIP1 expression both in vitro and in vivo.
Among the 25 patients with MEN1-related symptoms, a 4-bp deletion (c.-456_-453delCCTT, NM_004064) within the 5′UTR of CDKN1B in a 62 year old female patient with acromegaly and a well-differentiated non-functioning pancreatic endocrine neoplasm has been identified. This sequence variant was not detected in either 600 chromosomes or in the dbSNP/1000 genomes databases.
The 5′UTR of the CDKN1B gene is highly structured, containing several translational regulatory elements. An IRES element sustains p27KIP1 translation under poor growth conditions [10], [12], while a G/C-rich hairpin domain contributes to cell-cycle dependent regulation of CDKN1B translation [11]. Downstream the G/C-rich domain and encompassing the c.-456_-453delCCTT, an uORF coding for a 29 amino acid-long peptide has been described that has been suggested to inhibit the in vitro synthesis of p27KIP1 and to enhance its cell cycle-dependent translation [11]. An extensive comparative analysis of DNA and protein sequences from multiple species (Figure 1) confirmed previous data of high evolutionary conservation among vertebrates of the uORF [11], and support the hypothesis of a functional role of this element [11].
In general, uORFs are small open reading frames located in the 5′UTR of genes that influence translation during ribosome scanning, thus modulating gene expression. A scanning ribosome encountering an uORF has multiple fates: it can i) translate the uORF; ii) scan through the sequence (leaky scanning) and reinitiate translation further downstream at a proximal or distal ATG; iii) induce ribosome stalling or premature dissociation at the uORF stop codon, thus reducing downstream-cistron translation [19] or down-regulating gene expression by promoting mRNA decay [20].
In our case the 4-bp deletion shifts the uORF termination codon, thus lengthening the uORF encoded peptide from 29 to 158 amino acids and shortening the intercistronic space from 429 to 38 bp, with a possible negative influence on translation reinitiation from the main ATG (Figure 2). Long uORFs and short intercistronic regions may indeed prevent the 40S ribosomal subunits from keeping and/or re-acquiring appropriate cofactors for translation resumption/reinitiation at the downstream ATG [21], [22].
To address the possibility that the 4-bp deletion affects transcription and/or mRNA stability, making a decreased translation rate due to reduced reinitiation efficiency biologically irrelevant, or alters the promoter usage pattern preventing transcription of the uORF-containing isoform [10], we measured the steady state levels of CDKN1B allelic mRNAs from whole blood by 5′RACE and allele-specific qPCR. As reported in Figure 3a, both wild type and mutated alleles were expressed in blood cells in almost equal amounts, suggesting that the identified deletion does not alter mRNA steady state levels, and therefore probably does not alter either transcription or mRNA stability. An apparently unique 5′UTR of >530 bp has been identified in the c.-456_-453delCCTT carrier and in healthy controls, supporting the concept that the transcription pattern is preserved in the mutated subject (Figure 3b).
The pancreatic tumor of the mutated patient was then analyzed by immunohistochemistry for p27KIP1 expression and for the proliferation antigen Ki-67, and compared with similar tumors from CDKN1B-mutation negative subjects. Parallel differences in expression level and localization were found. We observed weak cytoplasmic staining in tumor cells and very strong nuclear staining in the interspersed normal endothelial cells in the MEN4 patient (Ki67<1%), while in contrast p27KIP1 nuclear staining was found in a high proportion of sporadic well-differentiated pancreatic tumors examined (Figure 4). The reduction in nuclear p27KIP1 and/or its cytoplasmic mislocalization has been reported in different cancers including breast, colon and prostate [4]. Loss of p27KIP1 may occur through different mechanisms, including augmented proteasome-mediated proteolysis and impaired translation [23]. On the other hand, the cytoplasmic mislocalization may be associated with imbalanced p27KIP1 phosphorylation due to the oncogenic activation of PI3K- and MEK-dependent kinases, mimicking protein loss [4]. Indeed, in the cytoplasm p27KIP1 is unable to exert its inhibitory activity on CDK even in the presence of anti-mitogenic stimuli.
On the same lesion loss of heterozygosity (LOH) analysis was then performed. No loss of the wild type allele was observed (Figure 2). Moreover, the biallelic expression of an uORF-containing transcript has been observed (Figure 3c), further confirming that p27KIP1 may act as a haploinsufficient tumor suppressor [24].
To identify possible additional uORF mutations, we extended the CDKN1B 5′UTR analysis to additional 41 patients with typical MEN1-like features previously reported negative for mutations in the CDKN1B coding sequence [17], [25]. A c.-469C>T substitution resulting in a silent change in the uORF was detected in a single patient but not in healthy controls (see above).
To determine whether the two identified substitutions negatively affect CDKN1B translation, the wild type and mutated 5′UTRs were cloned upstream of the firefly luciferase gene (Figure 5a). By transfecting lovastatin G1-synchronized or asynchronous HeLa and GH3 cells, we demonstrated that the c.-456_-453delCCTT, but not the c.-469C>T variant, significantly reduced luciferase activity in a cell cycle phase-independent manner (Figure 6a, 6b). When we analyzed the luciferase mRNA from the transfected cells by quantitative real-time RT-PCR, we demonstrated that the effects of the 4-bp deletion are largely due to reduction in translation rate rather than to changed steady-state mRNA levels (Figure 6c), in agreement with our observation on blood CDKN1B mRNA (Figure 3a) and with the trend observed in large-scale datasets [26].
We then evaluated the effect of the 4 bp deletion on p27KIP1 translation by transfecting HEK293 cells with vectors with either the wild type or the mutated 5′UTRs cloned upstream the CDKN1B gene (Figure 5b). We confirmed a significant reduction in p27KIP1 protein levels as a consequence of the 5′UTR c.-456_-453delCCTT mutation (Figure 6d).
In a previous study on HeLa cells using an identical wild type construct, the CDKN1B 5′UTR induced luciferase expression only during G1 progression or in lovastatin-arrested cells [11]. Although we cannot exclude the presence of DNA variations on regulatory elements between the two cloned sequences, a possible biological variability between batches of cells from the same cell line seems the more plausible explanation. However, similar cell-cycle independent luciferase activation was observed under our experimental conditions in three additional cell lines, namely GH3 (Figure 6a), SH-SY5Y and HEK293 (Figure S1). Based on such observation we may therefore suggest the need for further studies for better clarifying the cell-cycle dependent translation of p27KIP1 regulated by the CDKN1B 5′UTR.
Site-directed mutagenesis (c.-428A>T) was then used to reintroduce a stop codon in the c.-456_-453delCCTT containing vector, thus restoring both uORF length and intercistronic distance (Figure 5c). After transfection, the uORF regulatory properties were almost completely rescued in the double mutant compared to the c.-428A>T construct (Figure 7a), further supporting the hypothesis that the 4 bp deletion affects translation reinitiation of the downstream CDKN1B ORF. In addition, the lack of complete recovery of the uORF modulatory activity, possibly due to differences in the C-terminus of the uORF-encoded peptide (Figure 7b), further confirms that the CDKN1B uORF belongs to the class of sequence-dependent uORFs that exert their inhibitory role by acting in cis to regulate components of the translation apparatus [19].
To evaluate the ability of the uORF to be translated, which represents the central point of our hypothesis on the possible deleterious effects of the c.-456_-453delCCTT change, the wild type or the mutated 5′UTRs were placed upstream of the CDKN1B open reading frame and the c.-74insC mutation was introduced by site-directed mutagenesis. This additional DNA variant leads to the in-frame fusion of the mutated uORF with the main gene (Figure 5d). As expected, the chimeric product was detected only in the c.-74insC+c.-456_-453delCCTT transfected HEK293 cells, and was again associated with a significant reduction of p27KIP1 expression (Figure 6e). To our knowledge, this is the first direct evidence of the translation of the CDKN1B uORF in a cellular system. However, it remains to be clarified if this peptide has additional biological functions other than repressing translation of the CDKN1B ORF as an effect of impaired reinitiation, as suggested for a subset of uORFs [27].
To elucidate the molecular mechanism by which c.-456_-453delCCTT determines a decrease in p27KIP1 translational efficiency, we estimated the relative proportion of the two allelic mRNAs engaged in translation in the immortalized lymphoblastoid cells of the heterozygote patient. To this aim, the cell lysates were subjected to polysome fractionation through sucrose gradient ultracentrifugation [28] and we determined the level of each of the two allelic mRNAs for each fraction. Figure 8a reports the distribution of ribosomal RNA in the different fractions, showing a typical distribution with polysomes reproducibly spanning fractions 7–11. The distributions of the wild type and c.-456_-453delCCTT transcripts present an almost superimposable pattern, being for both about 90% of the total detectable mRNA localized in polysomes with a peak corresponding to fraction 9 (compare Figure 8b with Figure 8a). However, when the amounts of both alleles were expressed as differences between Cq values for mRNA and for genomic DNA for removing the intrinsic variation between the two qPCR assays, a clear preponderance on polysomes of the wild type CDKN1B mRNA could be observed (Figure 8c), which we can estimate to be of the order of about three times. Since the levels of the two allelic mRNAs in the cells are the same (Figure 3a), this implies that the c.-456_-453delCCTT mRNA suffers decreased average polysomal loading with respect to the wild type mRNA. Therefore, the two different CDKN1B mRNAs are differentially loaded in polysomes despite being present in the cells in the same relative amounts, and despite the fact that they share a distribution profile on polysomes of different molecular weights. The result is compatible with a decreased efficiency of translation reinitiation of the CDKN1B ORF due to the c.-456_-453delCCTT mutation.
The data we presented here further confirm the role of CDKN1B germline mutations in predisposing to a MEN4 syndrome. Furthermore, they demonstrate that a reduced translation initiation rate of p27KIP1 due to the ineffective regulatory activity of its uORF may be associated with transformation. Based on our data, mutations in the CDKN1B-regulating uORF seem to be rare. However, previous studies on CDKN1B germline mutations in MEN1-like patients did not consider the uORF region [8], [9], [14], [16]–[18], [25], [29], and therefore the prevalence of this type of mutation remains to be established. Our results emphasize thus the need for the inclusion of the entire 5′UTR region of CDKN1B in molecular testing for MEN4.
Increasing evidence suggests that uORF-mediated translational control may represent an important mechanism in the regulation of gene expression. This is supported by the close relationship of mutations that introduce or disrupt uORFs and the pathophysiology of several human diseases, including cancer [30]. To date, only three well-known hereditary diseases have been associated with uORF-affecting mutations: i) thrombocythemia due to thrombopoietin mutation [31], ii) melanoma due to CDKN2A mutation [32] and iii) Marie Unna hypotrichosis due to mutations in the hairless gene [33]. Other diseases, such as breast cancer, Alzheimer's diseases, arrhythmogenic right ventricular cardiomyopathy have also been suggested to be associated to genes which have uORF-related control [34]–[36]. However, the pathogenic effects of deregulated uORF-mediated translation in these cases remain to be clarified [37].
Many important genes involved in controlling cell growth (i.e. receptors, oncogenes, growth factors) harbor uORF in their 5′UTR [38]. Some of these genes override the uORF-mediated translational repression and accumulate their protein product in cancer cells [39]. Translational derepression elements in the 3′UTR may counteract the inhibitory activity of uORFs on translation [39]; however, mutations inducing loss of uORF function in oncogenes might lead to a similar increase of translation rate and consequently to malignant transformation. Conversely, gain of function mutations in uORFs regulating tumor suppressor genes may reduce translation of protective proteins leading to tumor formation [32]. Similarly to a point mutation introducing a regulative uORF in the leader sequence of the tumor suppressor gene CDKN2A in hereditary melanoma [32], the 4-bp deletion in CDKN1B gene we describe here led to the reduced production of CDKN1B-encoded protein p27KIP1, probably due to a decreased translation reinitiation rate, which then results in predisposition to tumor development.
In conclusion, the CDKN1B mutation functionally characterized in this study represents a novel example of an uORF-affecting mutation. Our functional studies show the negative influence of this deletion on the translation reinitiation at the CDKN1B starting site thus providing novel insights into the role of uORFs in the pathogenesis of human diseases.
In addition to the classical mechanisms of degradation by the ubiquitin/proteasome pathway and by non-covalent cytoplasmic sequestration, our findings demonstrate that p27KIP1 activity can also be modulated by its uORF, and mutations affecting this sequence may lead to reduced expression of p27KIP1 protein.
The cohort of patients screened for mutations in the entire CDKN1B gene consisted of 25 consecutive patients with two or more typical MEN1-related symptoms (hyperparathyroidism, neuroendocrine tumors, pituitary adenoma). Patients were collected and diagnosed at the Division of Endocrinology (University/Hospital of Padova) and at the Familial Cancer Clinic and Oncoendocrinology (Veneto Institute of Oncology), Padova, Italy, following the recognized clinical practice guidelines [40]. All patients had negative mutational screening for MEN1, PRKAR1A and AIP genes. A second group of additional 41 patients with similar phenotype has been analyzed only for the uORF sequence since the rest of the gene has been analyzed and published previously without finding any pathogenic mutations [17], [25]. The study was conducted in accordance with the Helsinki declaration. Local ethical committees from each referring center approved the study, and all subjects gave written informed consent.
The whole coding region, intron–exon boundaries, and 5′- and 3′-UTRs of CDKN1B were amplified and directly sequenced as reported elsewhere [41]. All primer pairs used were designed by PRIMER3 (http://primer3.sourceforge.net/) and synthesized by IDT (Leuven, Belgium). Primers for point mutation analysis of the entire human CDKN1B gene were P0F, 5′-agcagtacccctccagcagt-3′; P0R, 5′-aaagcccgtccgagtctg-3′; P1F, 5′-ccaatggatctcctcctctg-3′; P1R, 5′-ggagccaaaagacacagacc-3′; P2F, 5′-ccatttgatcagcggagact-3′; P2R, 5′-gccctctaggggtttgtgat-3′; P3F, 5′-gagttaacccgggacttggag-3′; P3R, 5′-atacgccgaaaagcaagcta-3′; P4F, 5′-tgactatggggccaacttct-3′; P4R, 5′-tttgccagcaaccagtaaga-3′; P5F, 5′-ccccatcaagtatttccaagc-3′; P5R, 5′-cctcccttccccaaagttta-3′; P6F, 5′-tgcctctaaaagcgttggat-3′; P6R, 5′-tttttgccccaaactacctg-3′; P7F, 5′-gccctccccagtctctctta-3′; P7R, 5′-ggtttttccatacacaggcaat-3′; P8F, 5′-tctgtccatttatccacaggaa-3′; P8R 5′-tgccaggtcaaataccttgtt-3′.
Previously unreported nucleotide changes were screened in 300 healthy, anonymous, unrelated individuals by Tetra-primer ARMS-PCR [42] and searched in the dbSNP and 1000 genomes databases (http://www.ncbi.nlm.nih.gov/projects/SNP/; http://www.1000genomes.org/). The NHLBI Exome Sequencing Project - Exome Variant Server database (http://evs.gs.washington.edu/EVS) has been queried for the c.-469C>T. Primers for Tetra-primer ARMS-PCR were: hp27delOUTR, 5′-agccgctctccaaacctt-3′; hp27delOUTF, 5′-caatggatctcctcctctgttt-3′; hp27delINF, 5′-cttcttcgtcagcctcccac-3′; hp27-469INR, 5′-tggcggtggaagggaggctgacgcaa-3′; hp27-469INF, 5′-gactcgccgtgtcaatcattttcgtc-3′.
Gene dosage alteration was assessed by the quantitative multiplex PCR of short fluorescent fragments (QMPSFs) and by long-range PCR (LR-PCR) as previously described [41] using the following primers: CLIF, 5′-tggtcagagagtggcctttctc-3′; CLIR, 5′-tgccgagtagaggcatttagtca-3′; CLIIF, 5′-tgtctgtgacgccgttgtct-3′; CLIIR, 5′-aagggttttctagcacacataggaa-3′; 1IF, 5′-gccgcaaccaatggatctc-3′; 1IR, 5′-acgagccccctttttttagtg-3′; 1IIF, 5′-ctctgaggacacgcatttggt-3′; 1IIR, 5′-aaatcagaatacgccgaaaagc-3′; 2F, 5′-tttcccctgcgcttagattc-3′; 2R, 5′-ccaccgagctgtttacgtttg-3′; 3IF, 5′-ccccatcaagtatttccaagct-3′; 3IR, 5′-gttattgtgttgttgtttttcagtgctta-3′; 3IIF, 5′-aacttccatagctattcattgagtcaaa-3′; 3IIR, 5′- tgagcgatgtggctcggct -3′.
Sequence based-LOH analysis was performed on the pancreatic lesion by direct analysis of the CDKN1B mutation.
The human cervical carcinoma HeLa, the rodent p27KIP1-negative GH-secreting pituitary adenoma GH3, the human embryonic kidney HEK293 and the human neuroblastoma SH-SY5Y cell lines (American Type Culture Collection, Manassas, VA), were maintained at 37°C in a 5% CO2 in complete 10% FCS DMEM.
GH3 (1.5×105 cells/well), HeLa (1.0×105 cells/well), SH-SY5Y (1.30×105 cells/well) and HEK293 (2.5×105 cells/well) cells were plated 24 hours before transfection into 12-well plates. When necessary, 24 hours after seeding cells have been arrested in G1 phase by a 36-hour treatment with either 10 µM (GH3) or 20 µM (HeLa) lovastatin. In all cell lines but HEK293 (see below) transient transfection was performed by Superfect (Qiagen, Milan, Italy). 1.5 µg plasmid and a ratio µg DNA/µl Superfect of 1∶6 following manufacturer's protocol were used. The pRL-TK plasmid (Promega) encoding Renilla luciferase was cotransfected and used for normalization of transfection efficiency. After 3 hours, the medium was changed to DMEM with 2% FCS and incubated for further 24 hours. Cells were then harvested in passive lysis buffer (Promega) and the relative luciferase activity was measured using the Dual-Luciferase Assay System and a GloMax 20/20 luminometer (Promega) according to the manufacturer's instructions.
For expression experiments, HEK293 cells were seeded into 12-well plates, grew to 95% confluence and transfected with Lipofectamine 2000 (Invitrogen, Milan, Italy) following the manufacturer's protocol. Cells were harvested 24 hours post-transfection, lysed in RIPA Buffer supplemented with proteases inhibitors (MgCl2 10 mM, Pepstatin 1 µM, PMSF 1 mM, cOmplete 1X (Roche, Monza, Italy)). Samples were clarified by centrifugation at 13,000 rpm for 5 min at 4°C.
Concentrations of the HEK293 extracted proteins were determined using the Bio-Rad DC protein assay kit (Bio-Rad Italia, Milan, Italy) following the manufacturer's instructions. For each sample, 20 µg were resuspended in NuPAGE LDS sample buffer and NuPAGE sample reducing agent (Invitrogen), boiled for 10 min at 70°C and resolved by SDS-PAGE on 4–12% NuPAGE gels (Invitrogen) and Mes buffer (Invitrogen). Separated proteins were transferred onto nitrocellulose membrane by Trans-Blot Turbo transfer system (BioRad) that was blocked for 2 hours with 5% non-fat dry milk (BioRad). The membrane was incubated overnight at 4°C with anti-p27KIP1 monoclonal antibody (BD Bioscience Heidelberg, Germany) used at 1∶300. Expression was corrected for differences in protein loading by probing blots for 1 hour at RT with mouse anti-ß-actin antibody (clone AC-15 1∶5,000, Sigma-Aldrich, Milan, Italy). Blots were developed using Pierce ECL Substrate (Part No. 32106, Thermo Scientific, Rockford, IL USA) and exposed to CL-XPosure Film (Thermo Scientific).
For total RNA extraction, HEK293 cells were resuspended in TRIzol (Invitrogen) and processed according to the manufacturer's instructions. Plasmid DNA contamination was removed by DNase, treating total RNA twice with Turbo DNA free kit (Applied Biosystems, Milan, Italy). One µg of DNase-treated RNA was reverse-transcribed using M-MuLV Reverse Transcriptase RNase H- (F-572S, Finnzymes, Espoo, Finland). qPCR was done with Platinum SYBR Green qPCR SuperMix-UDG (Invitrogen) in an ABI PRISM 7900HT Sequence Detector (Applied Biosystems). A final concentration of 300 nM for both forward and reverse primers was used. Primers for qPCR were qLUCF, 5′-gcctgaagtctctgattaagt-3′; qLUCR, 5′-acacctgcgtcgaaga-3′; qrBActF, 5′-agattactgccctggctcct-3′; qrBActR, 5′-aacgcagctcagtaacagtccg -3′; qhGAPDHF, 5′- ctctctgctcctcctgttcgac-3′; qhGAPDHR, 5′- ctctctgctcctcctgttcgac-3′.
Threshold levels were set at the exponential phase of qPCR using Sequence Detection software, version 2.4 (Applied Biosystems). The amount of each target gene relative to the proper housekeeping gene (HK, rat β-actin or human GAPDH) was determined using a relative standard curve method and the results were expressed as a ratio of target gene/HK. A 38-cycle threshold was set, beyond which the gene was considered undetectable.
Total RNA from whole blood samples was obtained using Paxgene Blood RNA Kit (Qiagen) following manufacturer protocol, while RNA from paraffin-embedded pancreatic tumor tissue was extracted using a modified RNAzol method, as previously described [43]. RNA was reverse-transcribed as described above. Allele-specific analysis was evaluated by qPCR as described above using two different SYBR assays. A final concentration of 300 nM for both forward primers and 50 mM for the unique reverse primer was used (-456_-453del_wtF 5′- cttcttcgtcagcctccctt-3′; -456_-453del_mutF 5′- cttcttcgtcagcctcccac-3′; -456_-453del-R 5′-agccgctctccaaacctt-3′). Given the different efficiency that may characterize the two different assays, the value of each allele was referred to the genomic DNA expressed as ΔCq (Cq value obtained for mRNA minus Cq value for genomic DNA). The lack of a possible deletion/duplication of the corresponding genomic locus was proven by QMPSFs as described above.
5′RACE was performed using 5′RACE System 2.0 kit (Invitrogen) on whole blood derived total RNA following manufacturer's instructions. Briefly, first strand cDNA was synthesized from 2 µg of mRNA by SuperScriptII RNA polymerase reaction using the specific primers GSP1R (5′- gttaactcttcgtggtcc -3′). After adding an oligo-dC tail to the cDNAs 3′-ends, a PCR reaction has been performed with GSP2R primer (5′-ttctcccgggtctgcacg-3′), coupled with an Abridged Anchor Primer (AAP). The resulting DNA fragments were eluted from agarose gel and analyzed by direct sequencing, as reported above.
Immunohistochemistry was performed on an automated immunostainer (Ventana Medical Systems, Frankfurt am Main, Germany), according to the manufacturer's protocols with minor modifications [44] using the monoclonal anti-p27KIP1 antibody cited above (1∶1,000). The monoclonal MIB5 antibody (1∶500, Dako, Hamburg, Germany) was used to detect the proliferation antigen Ki-67. Positive controls were used to confirm the adequacy of the staining.
PCR fragments were obtained by amplification of the mutation carrier with forward and reverse primers containing extra HindIII and NcoI sites, respectively (clonF, 5′- catcataagcttccaccttaaggccgcgct -3′; clonR, 5′- catcatccatggttctcccgggtctgcacg -3′). The PCR product was digested and inserted upstream the luciferase reporter gene into the pGL3 Control Vector (Promega). For expression studies the wild type and mutated 5′UTRs were subcloned into pcDNA3.1/p27HA (kind gift of Prof. Sylvain Meloche, Institute for Research in Immunology and Cancer, Université de Montréal, Canada). The c.-469C>T, c.-428A>T and c.-74insC modifications were introduced by QuikChange II XL kit (Stratagene, La Jolla, CA USA) following manufacturer's protocol.
EBV-transformed lymphoblastoid cells were generated by infection of peripheral blood mononuclear cells from the c.-456_-453delCCTT mutation carrier with culture supernatant from the EBV-producing marmoset cell line B95.8 (American Type Culture Collection) and maintained in RPMI 1640 medium (Euroclone, Milano, Italy) supplemented with 10% FBS, 1 mM Na Pyruvate, 10 mM Hepes Buffer, 2 mM Ultraglutamine (Lonza BioWhittaker, Basel, Switzerland), 1% Antibiotic/antimycotic (Gibco, Invitrogen Corporation). Cyclosporin A (CsA, Sandoz Pharmaceuticals AG; Cham, Switzerland) was initially added to the cultures to inhibit T cell growth (final concentration, 0.7 µg/ml).
For polysomal RNA extraction lymphoblastoid cells (25×106) were incubated with 100 µg/ml cycloheximide for 4 minutes, washed once with phosphate buffer saline (PBS), resuspended in lysis buffer [10 mM NaCl, 10 mM MgCl2, 10 mM Tris–HCl, pH 7.5, 1% Triton X-100, 1% sodium deoxycholate, 100 µg/ml cycloheximide, 0.2 U/µl RNase inhibitor, 1 mM DTT] and transferred to a microcentrifuge tube. After 5 minutes incubation on ice, the extracts were centrifuged for 10 min at 12,000 g at 4°C. The supernatant was collected and stored at −80°C. The cytoplasmic lysates were fractionated by ultracentrifugation (Sorvall rotor, 100 min at 180,000 g) trough 15–50% linear sucrose gradient containing 30 mM Tris–HCl, pH 7.5, 100 mM NaCl, 10 mM MgCl2. Eleven fractions were collected monitoring the absorbance at 254 nm. The RNA in each fraction was isolated after proteinase K treatment, phenol–chloroform extraction and isopropanol precipitation. RNA was resuspended in 30 µl of water. For each fraction 1 µg RNA was reverse-transcribed and analyzed by qPCR using allele-specific assays as reported above.
|
10.1371/journal.pntd.0001365 | Hidden Sylvatic Foci of the Main Vector of Chagas Disease Triatoma infestans: Threats to the Vector Elimination Campaign? | Establishing the sources of reinfestation after residual insecticide spraying is crucial for vector elimination programs. Triatoma infestans, traditionally considered to be limited to domestic or peridomestic (abbreviated as D/PD) habitats throughout most of its range, is the target of an elimination program that has achieved limited success in the Gran Chaco region in South America.
During a two-year period we conducted semi-annual searches for triatomine bugs in every D/PD site and surrounding sylvatic habitats after full-coverage spraying of pyrethroid insecticides of all houses in a well-defined rural area in northwestern Argentina. We found six low-density sylvatic foci with 24 T. infestans in fallen or standing trees located 110–2,300 m from the nearest house or infested D/PD site detected after insecticide spraying, when house infestations were rare. Analysis of two mitochondrial gene fragments of 20 sylvatic specimens confirmed their species identity as T. infestans and showed that their composite haplotypes were the same as or closely related to D/PD haplotypes. Population studies with 10 polymorphic microsatellite loci and wing geometric morphometry consistently indicated the occurrence of unrestricted gene flow between local D/PD and sylvatic populations. Mitochondrial DNA and microsatellite sibship analyses in the most abundant sylvatic colony revealed descendents from five different females. Spatial analysis showed a significant association between two sylvatic foci and the nearest D/PD bug population found before insecticide spraying.
Our study shows that, despite of its high degree of domesticity, T. infestans has sylvatic colonies with normal chromatic characters (not melanic morphs) highly connected to D/PD conspecifics in the Argentinean Chaco. Sylvatic habitats may provide a transient or permanent refuge after control interventions, and function as sources for D/PD reinfestation. The occurrence of sylvatic foci of T. infestans in the Gran Chaco may pose additional threats to ongoing vector elimination efforts.
| Triatoma infestans, a highly domesticated species and historically the main vector of Trypanosoma cruzi, is the target of an insecticide-based elimination program in the southern cone countries of South America since 1991. Only limited success has been achieved in the Gran Chaco region due to repeated reinfestations. We conducted full-coverage spraying of pyrethroid insecticides of all houses in a well-defined rural area in northwestern Argentina, followed by intense monitoring of house reinfestation and searches for triatomine bugs in sylvatic habitats during the next two years, to establish the putative sources of new bug colonies. We found low-density sylvatic foci of T. infestans in trees located within the species' flight range from the nearest infested house detected before control interventions. Using multiple methods (fine-resolution satellite imagery, geographic information systems, spatial statistics, genetic markers and wing geometric morphometry), we corroborated the species identity of the sylvatic bugs as T. infestans and found they were indistinguishable from or closely related to local domestic or peridomestic bug populations. Two sylvatic foci were spatially associated to the nearest peridomestic bug populations found before interventions. Sylvatic habitats harbor hidden foci of T. infestans that may represent a threat to vector suppression attempts.
| Disease eradication or elimination programs depend on time-limited intensive campaigns and are likely to fail if resistance to insecticides or drugs (i.e., malaria) or sylvatic transmission cycles (i.e., yellow fever) occur. Chagas disease is the most important vector-borne disease in Latin America in terms of disability-adjusted lost years, with an estimated 10–18 million people infected with Trypanosoma cruzi [1]. Elimination of domestic or peridomestic (hereafter abbreviated D/PD) populations of the insect vectors of T. cruzi through residual spraying with insecticides has shown varying degrees of success depending on the species and the occurrence of sylvatic foci. Several vector species occupy sylvatic habitats and show different degrees of domestication, such as T. dimidiata in Central America, Panstrongylus megistus, T. brasiliensis and T. pseudomaculata in Brazil, Rhodnius ecuadoriensis in northern Peru and Ecuador, and T. pallidipennis and related species in Mexico [2]–[4]. Species of sylvatic or peridomestic triatomines that were not recognized as control targets have emerged as primary vectors of T. cruzi in geographically defined areas over the last two decades [e.g., 5]. For species such as R. prolixus, house reinfestations may also be driven by invasion from peridomestic or sylvatic foci [6].
Triatoma infestans historically is the main vector of human T. cruzi infection. In 1991, this species was the target of a regional elimination program (the Southern Cone Initiative) that interrupted vector- and blood-borne transmission to humans in Chile, Uruguay, Brazil, eastern Paraguay and parts of Argentina [7]. However, only limited success in the elimination of T. infestans and interruption of vector-borne transmission has been achieved in the Gran Chaco region due to repeated reinfestations even in areas under intensive professional vector control [8]. The Gran Chaco, an ecoregion of 1.3 million km2 mainly spanning northern Argentina, Bolivia and Paraguay, has high levels of poverty and is hyperendemic for Chagas disease [9]. Recurrent reinfestation after residual spraying with insecticides and lack of a sustainable vector surveillance program result in renewed parasite transmission 3–5 years after community-wide vector control campaigns [10]–[13]. The obstacles to the elimination of T. infestans in the Gran Chaco may stem from different processes yet to be identified conclusively.
The Southern Cone Initiative for the elimination of T. infestans was based on two major assumptions with wide consensus and limited supporting evidence [14], [15]: (i) the species was restricted to D/PD habitats [16]–[19], with true sylvatic foci only occurring in rock piles associated with wild guinea pigs in the Cochabamba and Sucre Andean valleys in Bolivia [20]–[22], and (ii) T. infestans had low genetic variability and therefore was very unlikely to develop resistance to modern pyrethroid insecticides. Rare findings of T. infestans in sylvatic habitats up to the early 1980 s were judged to be of little relevance by several investigators (reviewed in [23], [24]). The surprising finding of melanic forms (“dark morphs”) in isolated dry forests in the Bolivian [23], [25] and Argentine Chaco [24], and more recently in the Paraguayan Chaco [26], combined with the discovery of sylvatic foci with normal phenotypes in Chile and Bolivia [27]–[29] challenged the highly domesticated status of T. infestans. In addition, recent evidence showed T. infestans had richer genetic variability than previously assumed [30]–[33], with strong chromosomal and DNA content differences between T. infestans from different sources [34], whereas pyrethroid resistance emerged in northwestern Argentina and throughout Bolivia since the late 1990 s [35], [36]. Understanding the ecological dynamics of reinfestation in insecticide-treated villages and untangling the mechanisms underlying the observed patterns is crucial for devising improved vector control tactics and the eventual elimination of T. infestans and other major triatomine vectors [18], [37]. Genetic [38] and phenetic [39], markers combined with carefully georeferenced bug samples collected before and after control interventions, a geographic information system (GIS) and spatial statistics [41] provide the means to better understand reinfestation dynamics. Here we first integrate the use of all these tools to investigate house reinfestation dynamics in the context of control interventions.
As part of a longitudinal project on the eco-epidemiology and control of Chagas disease in a well-defined rural area in the dry Argentine Chaco [8], we detected isolated findings of adult T. infestans and recently established, very low-density D/PD colonies during two years after a community-wide residual spraying of pyrethroid insecticides of all houses. To identify the putative sources for such occurrences and the sylvatic vectors of T. cruzi [42], we conducted intensive surveys for triatomine bugs in diverse sylvatic habitats after interventions and surprisingly found various sylvatic foci of T. infestans. Using fine-resolution satellite imagery, GIS, spatial statistics, genetic markers and wing geometric morphometry, we investigated the relatedness between sylvatic and D/PD populations of T. infestans and the threat that they may represent to vector control and elimination attempts in the Argentinean Chaco. Based on previous findings of sylvatic T. infestans in the Bolivian Chaco [43] and of an isolated adult specimen of T. infestans infected with T. cruzi in semi-sylvatic habitats of our study area in the mid-1980 s [44], we speculated that similar foci might exist in the Argentinean Chaco and that Triatoma guasayana was a likely candidate sylvatic vector of T. cruzi given its high abundance, widespread occurrence and occasional infection with the parasite [43], [45], [46].
Field studies were carried out in Amamá (27° 12′ 30″S, 63° 02′ 30″W) and neighboring rural villages in a 650 km2 area situated in the Moreno Department, Province of Santiago del Estero, Argentina (Figure 1). This area is located in the dry Chaco ecoregion [42] and its history of infestation since the mid-1980 s has been described elsewhere [8]. Based on the history of control interventions, the study area was subdivided into core (5 villages, 143 domiciles and 790 peridomestic sites) and peripheral (7 villages, 132 houses and 709 peridomestic sites) areas with all sites georeferenced. In April 2004, community-wide residual spraying with 2.5% deltamethrin (K-Othrin, Bayer) of nearly all houses was conducted by professional vector-control personnel using a standard insecticide dose in domiciles (25 mg/m2) and standard or double dose in peridomestic sites for enhanced impact. Here we only report results from the core area (villages of Amamá, Trinidad, Mercedes, Villa Matilde and Pampa Pozo; Figure 1) because no systematic searches for bugs were performed in sylvatic habitats around the peripheral communities.
Timed manual searches for triatomine bugs with a dislodging spray (0.2% tetramethrin, Espacial) were conducted in all domestic (0.5 person-hour) and peridomestic sites (one person-hour per house compound) from all study villages in October 2004, April and December 2005, and November 2006 as described before [10]. In the core area, 143 domiciles and 764 peridomestic sites were inspected for triatomine bugs at least once between 2004 and 2006. All detected foci were immediately sprayed with deltamethrin using the same procedures. As part of an ongoing monitoring program, discriminant dose assays demonstrated that no pyrethroid resistance occurred in local populations of T. infestans (María Inés Picollo, unpublished results).
We conducted four intensive surveys of triatomine bugs in sylvatic habitats using mouse-baited (Noireau) traps fitted with adhesive tape (Plasto®, Brazil) [47] in October and November 2005, April and November-December 2006 as described before [24]. Mean temperatures varied between 23°C and 26°C in October-December (spring) surveys, and were below 20°C in April (fall). Searches for sylvatic triatomine foci were conducted in 15 sampling areas that included representative forest sections with different degrees of disturbance (i.e., degraded forest under logging operations, cleared sections, ecotones, and implanted grasslands preceded by selective deforestation) and in all sorts of refuges potentially suitable for triatomine bugs. The total capture effort was 598 trap-nights (range per survey, 129 to 169). Traps were usually placed far from houses in holes of fallen or standing trees (live or dead), trunks or tree stumps and in between terrestrial bromeliads (Bromelia serra and Bromelia hieronymi), cacti (Opuntia quimilo and Opuntia ficus-indica) or piles of shrubs (Figure S1). Traps were deployed when the weather was warm and not rainy approximately between 17.00–18.00 hs and retrieved before 10.00 hs to protect mice from exposure to extreme temperatures. All trap locations were georeferenced using a GPS (Garmin, Etrex Legend C). All sylvatic sites surveyed in October and November 2005 were different except one, and 98% of them were re-inspected with mouse-baited traps on April 2006 to assess bug occurrence, persistence and invasion. The survey conducted in November-December 2006 only included sites that had not been surveyed previously.
Flight-dispersing triatomine bugs were collected using black-light traps [48] placed in 36 georeferenced sylvatic sites where concurrent searches with mouse-baited traps were made (i.e., in the same areas). Light traps were deployed away from houses in habitats where there was a wide opening in the forest that allowed at least a 100 m visibility. Light traps were operated from approximately 19:45 (i.e., 15 min before sunset) to 22:00–23:00 hs because the flight activity of T. infestans peaks during the first hour after sunset, and is more likely to occur when air temperature exceeds 20°C and wind speed is <5 km/h [48]–[50]. Suitable conditions for flight initiation of T. infestans occurred during the surveys conducted in October-November 2005 but not in April 2006.
All collected bugs were kept alive in plastic vials with folded filter paper, identified to species following Lent and Wygodzinsky [19] and counted. Species identification of very small first- or second-instar nymphs sometimes was considered tentative depending on the integrity of the material; no such doubts remained for third-instars or later stages. Fourth- or fifth-instar nymphs and adult bugs collected in 2005 were individually weighed on an electronic balance (OHAUS, precision, 0.1 mg) and total body length (L) measured from the end of the clipeus to the end of the abdomen with a vernier caliber (precision, 0.02 mm) to estimate a weight-to-length ratio (W∶L) –a quantitative index of nutritional status. The qualitative nutritional status of nymphs was determined by a cross-sectional view of the abdomen and cuticle distension and classified into four categories that ranged from unfed to large blood contents [51]. Feces from live third-instars and larger stages were examined microscopically for T. cruzi infection at 400× magnification.
DNA from bugs assigned to T. infestans (based on morphological evidence) was obtained, PCR-amplified, and sequenced for a 661 bp fragment of the mitochondrial genes cytochrome oxidase I (mtCOI) [32] and a 572 bp fragment of the cytochrome B (mtcytB) gene [52]. Sequences from sylvatic T. infestans were compared with Triatoma spp sequences available at Genbank and from previous surveys on the instraspecific variability of T. infestans [32], [53]–[56].
Sylvatic T. infestans mtCOI plus mtcytB composite haplotypes were compared with previously recorded haplotypes of D/PD T. infestans from the study villages (collected in 2001–2002), from other more distant (40 km) localities within Santiago del Estero Province (Quilumpa, Km 40, La Loma and Invernada Norte, collected in 2003–2004), and from other Argentinean Provinces more than 300 km apart (Salta, La Rioja, Tucumán and Formosa, collected in 2000–2005). A detailed description of the source localities was published elsewhere [32]. Genetic variability was estimated as the mean number of pairwise differences per site (π), Watterson's estimator (θW) and the haplotype diversity (Hd) with DnaSP 5.0 [57] and a statistical parsimony haplotype network was built with TCS 1.21 [58].
For higher resolution of the relationships between sylvatic and D/PD populations of T. infestans, the multilocus (ML) genotype for 10 microsatellite loci was obtained for sylvatic T. infestans using primers and PCR conditions previously described [59]. ML genotypes were compared with those from T. infestans captured in D/PD sites from Amamá and neighboring villages in October 2002 and April 2004 before full-coverage insecticide spraying [60]. Inter-individual genetic distance based on the complement of the proportion of shared alleles [61] was estimated with MICROSAT 1.5d (http://hpgl.stanford.edu/projects/microsat/), and a neighbor-joining (NJ) tree was built with MEGA 3.0 [62].
Using the genotypes of local D/PD T. infestans as reference populations, we applied the Bayesian based assignment-exclusion test implemented in GENECLASS 2 [63] to individually assign sylvatic individuals to the local pre-spraying D/PD populations (defined as the total gene pool at a given community in each capture date). No post-spraying reference groups could be formed because after community-wide insecticide spraying (2004–2006) most bug collections contained one or a few insects per site that were sparsely distributed throughout the communities (i.e., no established populations of T. infestans were detected). Reference populations were not excluded as the putative origin of the sylvatic insects when the marginal probability exceeded 0.05. We used 100,000 replications and a simulation algorithm [64].
Sibship of T. infestans bugs collected in traps with more than one individual (TN-92 and TN-139) was inferred with the maximum likelihood approach implemented in COLONY 2.0 [65] performing two independent runs and assuming a probability of null alleles of 0.05 in loci ms42, ms64 and ms65 due to departures from Hardy-Weinberg expectations.
The wing geometric morphometry of the only sylvatic T. infestans male collected was compared with T. infestans males captured in D/PD sites from Amamá and neighboring study villages in October 2002 (n = 87) and April 2004 (n = 74) as described elsewhere [66]. The geometric coordinates of 11 type-I landmarks (venation intersections) from all right wings were digitized by the same user (JSB). After performing the generalized Procrustes superposition (GPA, [67]), the residual coordinates of the total sample (including the sylvatic specimen) were transformed into partial warps (PW). These shape variables allow standard statistical analyses such as principal component (PCA) or discriminant analyses (DA). To cope with small sample sizes in some villages, the first nine principal components of the PW were used as input for a DA performed on the village samples (excluding the sylvatic specimen). These principal components are also called relative warps (RW). The sylvatic specimen was then used as supplementary data and its position in the morphospace examined in terms of Mahalanobis distances. Digitization, GPA, PCA and DA were performed using the corresponding modules of the CLIC package [68].
Global positioning system readings from all sampling sites (with mouse-baited and light traps) were integrated into a Geographic Information System (ArcGIS 9.1, ESRI, Redlands, CA, U.S.A.) of the study communities containing a georeferenced satellite image (Ikonos2, Space Imaging Inc., Atlanta, GA, U.S.A.) and the position of all houses and peridomestic sites sprayed with insecticides in 2004. Cartesian coordinates (Universal Transverse Mercator, UTM, Zone 20S) were calculated for each D/PD site and trapping location in order to perform spatial analysis. A focal spatial statistic (Gi(d)) [69] was used to determine the presence and extent of spatial clustering of T. infestans D/PD abundance (average of timed manual catches of bugs per site in 2002 and 2004] around each T. infestans-positive sylvatic focus (point i). This local statistic is additive in the sense that it focuses on the sum of the j values in the vicinity of point i. Hence, we took each T. infestans-positive sylvatic focus, one at a time, and searched the nearby area for occurrences of more or fewer D/PD T. infestans bugs collected before full-coverage insecticide spraying than expected by random. This procedure identified specific trap locations as members or non-members of infestation clusters. We used a binary weight wij based on a distance threshold (d) scheme. Clustering of D/PD T. infestans abundance around a positive sylvatic site occurred when the observed Gi was higher than 2.32 (the expected value at P<0.01). We evaluated the value of Gi up to 3 km from each sylvatic site with T. infestans –a tentative upper bound of the flight range of T. infestans. Analyses were performed using the software Point Pattern Analysis [70].
Humane care and use of laboratory animals were performed according to Institutional Animal Care and Use Committee (IACUC, CICUAL in Spanish) guidelines at UBA's Faculty of Exact and Natural Sciences. Animal care and use is guided by the International Guiding Principles for Biomedical Research Involving Animals developed by the Council for International Organizations of Medical Sciences.
A total of 13 (9.1% of 143 domiciles) domestic foci of T. infestans with 23 bugs and 38 (5.0% of 764 sites) peridomestic foci with 223 bugs were detected between 2004 and 2006 after full-coverage spraying with deltamethrin. Nearly 25% of all collected T. infestans were adult bugs.
Only 30 (5%) of 598 mouse-baited traps deployed overnight in sylvatic habitats were positive for triatomine bugs (Table 1). Six sylvatic foci of T. infestans with normal chromatic characters (totaling 23 nymphs and 1 male; range per site, 1–17) were found in tree holes or trunks (Figures S1 and S2). Another probable sylvatic foci of this species with two first- or second-instar nymphs was conservatively excluded because the morphological identification of these stages was uncertain and mtDNA markers did not amplify; this probable focus occurred in the vicinity of the largest sylvatic colony of T. infestans (trap TN-139, Figure S2). The apparent density of sylvatic T. infestans was 4 per 100 trap-nights (24 bugs in 598 trap-nights; mean ± SD, 3.8±6.4 bugs per site). One sylvatic focus located west of Amamá (trap TN-139) was infested both in October (1 male) and November 2005 (14 first- or second-instar nymphs and 2 fourth-instars) and was taken as one colony. No T. infestans bugs were collected with mouse-baited traps in April or November 2006.
T. guasayana occurred more frequently (3.0% of mouse-baited traps in all surveyed habitats) than T. infestans (1.2%, Table 1). Feces and hairs of Didelphis opossums were found in one T. guasayana focus. All first- or second-instars of Triatoma sp. not identified to species level most likely were T. guasayana based on morphology, size and type of habitat. Light-trap collections yielded 110 adult T. guasayana, one specimen of T. garciabesi (female) and one of T. platensis (male), and no T. infestans in 41 light-trap-nights (Table 1). Of the 41 light-trap nights, 28 (68.3%) were positive for triatomine bugs. The adult sex ratio in T. guasayana was 1∶2.2 (male to female).
Sylvatic foci of T. infestans occurred at 5 sampling areas located 2.0–11.5 km apart (Figure 2). Most triatomines (17 or 70.8% of 24 T. infestans and 18 or 64.3% of 28 T. guasayana) caught with mouse-baited traps occurred in areas that had been deforested selectively (totalling 40 bugs at 11 sites); the other seven T. infestans were caught in secondary forest with medium-sized or a few large-sized trees. The only three T. garciabesi found were caught in mature forest under active deforestation. The remaining triatomine bugs were caught in secondary forest with medium- or large-sized trees. The main identified micro-habitats of T. infestans were in holes of fallen trees and decaying tree trunks lying on the ground (21 or 87% of 24 bugs collected), a tree stump and a live standing tree. These ecotopes included 4 ‘quebracho colorado’ (Schinopsis lorentzii) and 2 ‘mistol’ (Zizyphus mistol) trees (Figure S1).
Nearly all triatomine bugs caught with mouse-baited traps and examined for qualitative nutritional status (n = 36) were unfed (61.1%) or had very little remnants of a blood meal (33.3%) and very low W/L ratios (Table S1). Of 140 sylvatic triatomine bugs examined microscopically (10 T. infestans, 21 T. guasayana and 3 T. garciabesi caught with mouse-baited traps and 106 T. guasayana collected with light traps) none was found microscope-positive for T. cruzi.
The morphological identification of 20 sylvatic bugs as T. infestans was confirmed by DNA sequencing of mtCOI and/or mtcytB fragments; DNA from six other bugs (all first- to third-instars identified as T. infestans based on morphological characters) could not be amplified. The two third-instar nymphs not amplified were taken as T. infestans because a morphological misidentification (relative to the locally known species) was considered very unlikely. None of the sylvatic T. infestans bugs carried the T_C change at position 556, which is characteristic of T. platensis and is absent in a large sample of T. infestans from Argentina, Bolivia, Peru, and Uruguay [32].
Sylvatic T. infestans with mtCOI and mtcytB composite haplotypes (n = 16, Table S2) exhibited high nucleotide variability (θW = 0.006, π = 0.007) and haplotype diversity (Hd = 0.901). No shared haplotypes were found among bugs from different traps, whereas traps with more than one bug had one (TN-92, n = 3) and five (TN-139, n = 11) different haplotypes (Table S2). Of eight sylvatic haplotypes identified, six were exclusive of sylvatic bugs whereas two haplotypes were recorded in local peridomestic populations of T. infestans and elsewhere in Argentina (Figure 3). Sylvatic haplotypes were spread along the entire statistical parsimony network; they did not form a unique cluster separated from the rest and were more closely related to D/PD than to other sylvatic haplotypes (Figure 3). One sylvatic haplotype was highly divergent (am-XIV) but also was closely connected to an Amamá peridomestic haplotype (haplotype b-XIV).
The multilocus (ML) genotype for 10 microsatellite loci was obtained for 21 sylvatic T. infestans. We identified a total of 86 different alleles for the 10 loci, of which only 15 (17.5%) and 17 (19.8%) were private alleles not detected in the local D/PD populations in 2002 and 2004, respectively. Sylvatic T. infestans clustered among D/PD bugs with no sharp discontinuity (Figure 4). T. infestans bugs captured concurrently at trap TN-139 clustered together whereas bugs collected there at different times were more closely related to different clusters of Amamá peridomestic bugs (i.e., the closest village). In addition, insects from trap TN-139 had five different mtCOI-mtcytB haplotypes (Table S2). Sibship microsatellite analyses showed that bugs that shared a mitochondrial haplotype (or that had consistent haplotypes because of missing data for mtCOI or mtcytB) were most likely full- or half-sibs whereas bugs with different haplotypes were not (Tables S3 and S4). Bugs from trap TN-92 clustered together and closely to bugs from Mercedes village (where the trap was located) and from another village at ∼5 km (Pampa Pozo). These three bugs were full- or half-sibs and shared the same mitochondrial haplotype (Tables S3 and S4). The bug from site trap TN-182 was grouped with bugs from the nearest village (Mercedes) located at ∼8 km. The bug collected at trap TN-101 (close to Villa Matilde, Fig. 1) clustered with bugs from Amamá and Pampa Pozo.
The Bayesian-based assignment-exclusion test indicated that 18 of 21 sylvatic ML genotypes were not excluded from one or more of the D/PD reference populations (Table 2). D/PD populations were excluded as putative sources for three sylvatic insects captured in two different sites (traps TN-182 and -139). The mtCOI-mtcytB haplotype from the bug in trap TN-182 (al VII, Figure 4) was also genetically distant from the local D/PD populations and was closely related to D/PD populations from La Rioja, more than 400 km far from the study area (Figure 4).
Wing geometric morphometry was used to compare the only sylvatic T. infestans male collected (trap TN-139) with T. infestans males captured in local PD sites in 2002 and 2004. The factorial map showed that the sylvatic bug clearly overlapped with 2002 PD bugs from Amamá –the closest village to its capture site (Figure 5) and it was also assigned to 2004 PD bugs from Amamá (not shown).
All sylvatic foci of T. infestans were located 110–2,300 m from the nearest D/PD sites ever found to be infested by this species after full-coverage insecticide spraying (i.e., detected during the preceding 18 months) (Figure 2). Trap location TN-182 included two T. infestans-positive sites (TN-180 and TN-182) that were analyzed together because their separation (13 m) was smaller than the distance resolution of the Gi(d) test (50 m). The distance between traps positive for T. infestans to the nearest house varied from 125 to 1,900 m. Spatial analysis showed a statistically significant association (Gi(d)>2.32, P<0.01) between two sylvatic foci of T. infestans found within three km of a D/PD site and the average timed-manual catch of bugs before insecticide spraying (Figure 2). Significant clustering occurred up to 1.2 km in Amamá (trap TN-139, with 17 insects) and up to 150 m in Mercedes (trap TN-101, with one third-instar nymph) (Figure 2). The remaining three sylvatic foci of T. infestans (TN-182, TN-180 and TN-92) were located at 430–1,846 m from the nearest infested house, but did not appear to be significantly associated with any of them (Gi(d)>1.96; P>0.05).
We report here the first finding of multiple sylvatic foci of T. infestans: i) with normal chromatic characters (not “dark morphs”) in the Gran Chaco region outside Bolivia; ii) with morphological identification confirmed by DNA sequence information –ruling out taxonomic misdiagnosis of nymphs, and iii) with a genetic makeup indistinguishable from their local D/PD conspecifics in nearly all cases. The discovery of sylvatic foci of T. infestans was made possible by the extensive deployment of mouse-baited sticky traps in a wide diversity of habitats potentially suitable for the species as suggested by surveys in the Bolivian Chaco [43]. Although mouse-baited sticky traps may not achieve perfect detection of sylvatic foci [26], [71], the alternatives of using timed manual collections with a dislodging spray or habitat destruction are even less satisfactory or feasible [24]. Easy-to-use, more sensitive sampling methods for triatomine bugs in sylvatic habitats are crucially needed. Therefore, the actual prevalence of sylvatic foci of T. infestans as determined with mouse-baited traps was most likely underestimated.
Mitochondrial and microsatellite DNA markers coupled with wing geometric morphometry consistently indicated the occurrence of unrestricted gene flow between local D/PD and sylvatic T. infestans populations. In spite of the occurrence of private mitochondrial haplotypes and microsatellite alleles in sylvatic bugs, analyses suggest a strong genetic relationship with D/PD bugs. In the phylogenetic network, mtDNA sylvatic haplotypes were more frequently connected to peridomestic haplotypes rather than to other sylvatic variants ―which indicates that they did not form a population that evolved in isolation for a long period of time. A limitation here is that mtDNA only allows the estimation of historical female-based gene flow. However, microsatellite-based analyses ―a more suitable tool for detecting current gene flow― failed to reject that neighboring villages were the putative sources of sylvatic bugs in most cases. In addition, the small level of differentiation between sylvatic and D/PD specimens fell within the observed levels of within-population diversity [31], [60]. Therefore, there is no sufficient evidence to support restriction of gene flow between sylvatic and D/PD populations of T. infestans from the surrounding villages except in one case (trap TN-182).
Sibship analyses coupled with mitochondrial haplotype information at trap TN-139 over two trapping sessions separated by one month suggest that descendents from five different females were found at this rather remote site. Microsatellite data corroborated the heterogeneous genetic composition of TN-139 bugs as the insects were assigned to different reference populations. In the context of rare, light D/PD infestations after the insecticide spraying campaign, the finding of multiple haplotypes at a defined site was surprising. This sylvatic colony of T. infestans (the largest) was located 1.1 km away from the nearest infested house in an isolated habitat with no signs of current or past human use over the previous two decades (Figure S2). Moreover, another probable sylvatic foci of T. infestans with early-stage nymphs was detected in the vicinity of the largest sylvatic colony. Passive transport of T. infestans in the belongings of rural workers at a transitory camp may have provided an additional means of disseminating bugs within and between communities or the surrounding landscape. This alternative is worth considering because long-distance passive bug transport beyond its distribution range is well known and still occurs [72]. Thus, genetic and morphological evidence combined with the past history of denser D/PD infestations [8] suggests that the sylvatic specimens of T. infestans may have been feral derivatives (“spill-over”) of D/PD populations. Lack of sampling in sylvatic habitats before full-coverage insecticide spraying unfortunately does not allow establishing whether the sylvatic foci of T. infestans existed before or were formed as a consequence of flight dispersal of D/PD adult bugs or human-assisted passive transport of bugs.
These findings question the widely held notion of an unlikely continuous exchange of T. infestans bugs between wild and domestic habitats in the Chaco. Earlier studies using allozymes or morphometrics [20], [21] and mitochondrial DNA [32], [54] did not detect differences between sylvatic and domestic populations of T. infestans in the Andean Bolivian valleys, neither could mitochondrial markers in Chile [73]. In the allozyme-based study, the findings were interpreted as suggesting that sylvatic foci could be recent derivatives from nearby D/PD bug populations or vice versa –a pattern that was also consistent with unrestricted gene flow between domestic and sylvatic T. cruzi [20]. Intense gene flow between both types of bug populations (abundant at that time) could have generated the same patterns. Using head morphometry in the same study area in the Andean Bolivian valleys, reinfestant specimens of T. infestans found six months after house spraying with pyrethroids were considered survivors of the original domestic bug population unrelated to local sylvatic specimens [21]. Microsatellite data comparing the genetic makeup of sylvatic and D/PD populations of T. infestans showed restricted gene flow between sylvatic and peridomestic populations separated by only 300–650 m at 2,700 m altitude in the Andean Bolivian valleys [74], whereas they were highly structured and with evidence of low, asymmetric gene flow in a remote, well-preserved dry forest in the Argentinean Chaco [33]. Our data collected in highly-disturbed dry forest with more scattered houses show a different pattern and suggest that the occurrence of sylvatic foci of T. infestans may explain at least some of the new D/PD foci detected after full-coverage residual spraying of insecticides.
All sylvatic foci of T. infestans were 110–2,300 m from the nearest house or infested D/PD site detected after full-coverage insecticide spraying. These distances are within the estimated flight range of this species (1.5 km) derived from direct and indirect observations [12], [49], [50], [75], [76]. Because T. infestans may sustain tethered flights for >20 min at speeds of 2 m/s [77], the upper bound of its flight range may reach 3 km and remains uncertain. Therefore, the significant spatial associations detected combined with the range of distances between sylvatic and D/PD foci of T. infestans suggest that these habitats were probably connected through flight dispersal of adult bugs.
The identified habitats of sylvatic T. infestans in our study area were nearly all associated with trees at ground level, in fallen trees or tree stumps. No rocky outcrops were available. Potential bug refuges at higher altitude in the canopy ―difficult to spot and sample― were much less represented in our surveys. Compared with other sylvatic foci investigated with mouse-baited traps, the local apparent density of T. infestans (4 bugs per 100 trap-nights) was slightly higher than that recorded in remote dry forest in the Argentine Chaco (1.2 bugs per 100 trap-nights) [24], and substantially lower than in the Bolivian Chaco (17 bugs per 100 trap-nights) [43] or the Andean valleys (8–123 bugs per 100 trap-nights) [56]. The finding of small, malnourished sylvatic colonies with immature stages of T. infestans indicates that despite extensive deforestation and land-use change, the degraded forest still maintained suitable conditions and resources for bug development but at reduced levels: the apparent abundance and availability of blood-meal sources (not identified yet) in local sylvatic habitats was poorer and more unstable than in D/PD habitats. Some of the sylvatic bug foci in our study could be considered “semi-sylvatic”, in the sense that these habitats were intermediate between peridomestic ecotopes (such as pig or goat corrals made with piled thorny shrubs) and sylvatic habitats in terms of resident host species and abundance [46]. Semi-sylvatic habitats also tend to be less used and modified by regular human activities than peridomestic ecotopes. These findings suggest the possibility of sylvatic foci of T. infestans in almost any rural area within its geographic range. The domestication process T. infestans underwent in the past does not prevent the species from surviving at low density in a wide diversity of sylvatic or semi-sylvatic habitats, even after community-wide insecticide spraying.
Unlike previous reports in Argentina where the presence of sylvatic T. infestans may be a result of spill over from heavy D/PD infestations [reviewed in 24], the sylvatic T. infestans of this study occurred in sampling areas around villages under vector surveillance and selective control activities that only allowed the establishment of very few low density D/DP foci for limited time periods between surveys. A relevant question is whether the small-sized sylvatic bug populations we found are viable in the absence of immigration from D/PD sources (i.e., ‘rescue effects’) or they simply are temporary sinks. Our second-year follow-up data raise doubts about their viability over a longer time horizon in the absence of immigration, although removal of bugs may have contributed to apparent local extinctions. However, it is noteworthy that “dark morph” populations of T. infestans in the Bolivian and Argentine Chaco were viable despite having very low density and remote locations excluding them from D/PD ‘rescue effects’ [24], [25], [33].
T. guasayana was far more abundant than T. infestans in sylvatic habitats, and light-trap collections demonstrated the large number of flight-dispersing adult T. guasayana, as was found in the Bolivian and Paraguayan Chaco [78], [79]. Previous studies showed that T. guasayana also colonized peridomestic structures and semi-sylvatic ecotopes where it was associated positively with the local abundance of goats and the density of cacti and bromeliads [46]. Householders frequently collected adult bugs of this species when invading human habitations at sunset but this species was not able to colonize domestic premises before or after apparent suppression of T. infestans [8], [10], [46]. In the present study, the concurrent finding of T. guasayana in a fallen tree with fresh signs of Didelphis opossums suggests a close association with the main local sylvatic reservoir of T. cruzi typically infected with discrete typing unit I [42], [80]. The widespread occurrence and large abundance of T. guasayana combined with its ocassional infection, opportunistic blood-feeding behavior and dispersal ability implicate it as a secondary vector of T. cruzi in the peridomestic environment [45] and sylvatic habitats in the Argentine Chaco.
A long-standing, key scientific question with vast implications for vector control is what is the source of the triatomine bugs appearing after community-wide insecticide spraying [18], [37], [81]. Are they (i) survivors or the offspring of previously existing bugs; (ii) immigrants from untreated D/PD or sylvatic foci; or (iii) migrants brought by passive transport from other villages or elsewhere? This issue is applicable to all major triatomine vector control programs throughout Latin America and the responses may differ between settings and even within the same species, as with T. dimidiata in Central America and Mexico or T. brasiliensis and P. megistus in Brazil –all of which display substantial within-species differences in habitat distribution, invasive capacity and other relevant traits. As with other species of triatomine bugs, T. infestans adults and nymphs are attracted to lights [48], [82]. Sylvatic populations of T. infestans are much more widespread than assumed in the past [23]–[29] and have recently been discovered in the Paraguayan Chaco [26]. Because sylvatic habitats are not targeted for vector control operations, they may provide hidden refuges for T. infestans from which they may reinvade houses in search of more suitable conditions and resources. Our results suggest that in areas with recurrent reinfestation, vector control programs should consider the potential occurrence of external sources (semi-sylvatic or sylvatic) around the target community. The role that sylvatic populations of T. infestans (either with melanic or normal phenotype) play in the process of recolonization of insecticide-treated villages and their invasive capacity needs to be more widely investigated to evaluate the risk they pose to effective vector control and eventual elimination in the Gran Chaco and elsewhere.
|
10.1371/journal.pntd.0006488 | Cost-benefit analysis of intervention policies for prevention and control of brucellosis in India | Brucellosis is endemic in the bovine population in India and causes a loss of US$ 3·4 billion to the livestock industry besides having a significant human health impact.
We developed a stochastic simulation model to estimate the impact of three alternative vaccination strategies on the prevalence of Brucella infection in the bovine populations in India for the next two decades: (a) annual mass vaccination only for the replacement calves and (b) vaccination of both the adult and young population at the beginning of the program followed by an annual vaccination of the replacement calves and, (c) annual mass vaccination of replacements for a decade followed by a decade of a test and slaughter strategy.
For all interventions, our results indicate that the prevalence of Brucella infection will drop below 2% in cattle and, below 3% in buffalo after 20 years of the implementation of a disease control program. For cattle, the Net Present Value (NPV) was found to be US $ 4·16 billion for intervention (a), US $ 8·31 billion for intervention (b) and, US $ 4·26 for intervention (c). For buffalo, the corresponding NPVs were US $ 8·77 billion, US $ 13·42 and, US $ 7·66, respectively. The benefit cost ratio (BCR) for the first, second and the third intervention for cattle were 7·98, 10·62 and, 3·16, respectively. Corresponding BCR estimates for buffalo were 17·81, 21·27 and, 3·79, respectively.
These results suggest that all interventions will be cost-effective with the intervention (b), i.e. the vaccination of replacements with mass vaccination at the beginning of the program, being the most cost-effective choice. Further, sensitivity analysis revealed that all interventions will be cost-effective even at the 50% of the current prevalence estimates. The results advocate for the implementation of a disease control program for brucellosis in India.
| Brucellosis is an endemic zoonosis in India and recent studies demonstrate that the disease results in a median loss of US$ 3.43 billion in livestock populations. Lack of resources to compensate farmers and a ban on cow slaughter means that test and slaughter policy to control brucellosis cannot be implemented in India. This is the first systematic analysis of a brucellosis control program interventions for bovine brucellosis in India. The cost-benefit analysis was successfully conducted and indicated benefits of implementing the intervention policies. For each intervention, our results indicate that the prevalence of Brucella infection will drop below 2% in cattle after 20 years of the implementation of disease control program although some strategies were better than others. The expected net present value (NPV) was found to range from US $ 4·16 to $ 8·31 billion for cattle and from $ 7·66 to $ 13·42 billion in buffalo for the three strategies investigated. The benefit cost ratio (BCR) ranged from 3·16 to 10·62 for cattle and from 3·79 to 21·27 for buffalo. The results advocate for the implementation of a disease control program and will help development of an official health policy for the control of brucellosis in India.
| Brucellosis is an important zoonotic disease causing infertility, repeat breeding, retention of placenta and abortion in cattle. Humans in contact with animals usually get infected by coming in direct or indirect contact with reproductive secretions and excretions from infected animals. The disease is quite painful among humans and causes undulant fever, chills, fatigue, joint and muscle pain. If not treated, the disease can last for months and years and can cause orchitis, epididymis and endocarditis. Successful implementation of disease control programs have resulted in the eradication of brucellosis from domestic livestock in most of the developed countries [1]. However, the disease is still prevalent and classified as a neglected zoonosis in many parts of the developing world [2].
The disease is endemic in most of the production animals in India [3, 4]. With the reported disease seroprevalence of 9.3% in cattle [5] and 16·4% in buffalo populations [6], brucellosis is a serious economic concern for the cattle and buffalo industry [7]. Recent studies in India demonstrate that brucellosis in livestock populations results in a median loss of US$ 3·43 billion, with more than 95% of the losses occurring in the cattle and buffalo industry [7].
Brucellosis can be successfully controlled using appropriate intervention policies. Lack of resources to compensate farmers and a ban on cow slaughter in most parts of the country means that test and slaughter policy to control brucellosis cannot be implemented in India. There is no treatment for the disease in animals. Therefore, vaccination of cattle and buffalo population remains the sole alternative for the prevention and control of brucellosis in livestock populations in the country. However, information of benefits and costs of implementing intervention strategies to control the disease in India are largely unknown.This study aims to assess the costs and benefits of alternative control strategies for brucellosis in India. Initially, a stochastic simulation model was developed to project the course of Brucella infection for the national cattle and buffalo herd, over the next twenty years, under two different vaccination schemes. Subsequently, we performed a cost-benefit analysis to quantify the expected benefits of the proposed alternatives. We anticipate that this study would help policy makers to adopt the best available long-term intervention policy to prevent and control the occurrence of brucellosis in livestock and human populations in India.
Firstly, we developed a stochastic simulation model to estimate the impact of alternative vaccination strategies on the prevalence of Brucella infection for the cattle and buffalo populations in India, for the next two decades. The considered alternatives were based on published literature [8–10]. For the first intervention, we assumed a planned annual livestock mass vaccination campaign using Brucella abortus S19 for the female bovine (cattle and buffalo) replacement calves. For the second intervention, we assumed that all the adult and young female bovine populations will be vaccinated at the beginning of the program, followed by an annual vaccination of only replacements. The third intervention considered the annual mass vaccination of replacements for a decade followed by a decade of a test and slaughter strategy. We quantified the expected benefits and gains of the proposed control programs and performed a benefit-cost analysis to calculate the overall net expected benefit for each intervention.
To estimate the expected benefits from the alternative vaccination strategies, the following dynamic, synchronous, discrete time event stochastic simulation model was setup. First, animals were generated within herds. The time step (t) for this model was one year. At each time step, (a) the life stage (i.e. age) of each animal was determined by a dynamic component that is based on data about the age distribution and the age-specific replacement rate for the cattle and the buffalo populations; and (b) the infection stage for each animal was based on the expected prevalence of brucellosis for each species. Prevalence (P) was simulated at the herd level and animals within the same herd were assumed to attain the same risk of getting infected. For each herd, P was simulated for the first year and for each of next years it was based on the mean prevalence estimate of the previous year Pt-1.
Animals that got infected were assumed to remain infected for life. Replacements and animals that were not infected were assumed to attain a yearly risk (YRt) of getting infected that depended on Pt-1 and the expected mean duration (D) of the disease in the infected animals:
YRt=1−e−Pt−1(1−Pt−1)D
The model was allowed to run for a “burn-in” phase of 50 years and then each of the alternative interventions was considered: (a) annual mass vaccination only for the replacement calves and (b) vaccination of both the adult and young population at the beginning of the program followed by an annual vaccination of the replacement calves and (c) annual mass vaccination of replacements for a decade followed by a decade of a test and slaughter strategy. Vaccination was assumed to provide complete protection from infection although we allowed for a small rate of vaccination failures and we also considered different realistic vaccination coverage rates that in real life affect vaccine efficacy. At each time step, a number of parameters was recorded among which prevalence (P), replacement rate, the number of vaccinations and for the third scenario (i.e. vaccination of replacements followed by a test and cull strategy) the number of tested animals and the number of culled animals.
Estimates were based on the summaries of 4000 simulations of all animals and herds that were run for 40 times, for each species and intervention. A detailed description of the data sources and the input parameters of the model follows. Input parameters and the corresponding distributions are also summarized in Table 1 [5–7, 11–22].
For each intervention we evaluated the effect of varying input parameters on the expected benefits. Specifically, we assessed the impact of (a) reducing the initial prevalence of Brucella infection by fifty percent, (b) reducing the vaccination coverage to 50% and (c) having herds that are consistently unvaccinated (i.e. herds that were more likely to remain unvaccinated the next year). We also assessed the additional benefits of expanding the intervention strategies beyond the twenty year period.
The interventions were considered for t = 20 years and at a discount rate (r) of 5%. Initially, we predicted the annual costs (Ct) and benefits (Bt) for each strategy and subsequently calculated the net present value (NPV) by applying the discount rate:
NPV=∑t=1TBt−Ct(1+r)t
Further, for each intervention the benefit cost ratio (BCR) was estimated as the discounted value of the incremental benefits divided by the discounted value of the incremental costs:
BCR=∑t=1TBt(1+r)t∑t=1TCt(1+r)t
The costs included vaccine costs, service costs of vaccination (transportation, cold chain, and veterinarian fee), animal identification costs (ear tagging), service costs for surveillance and diagnostics, and costs for health education program (Table 1).
The averted losses were considered as benefits for implementing the control programs [27]. Based on our previous study [7], the losses occurring due to brucellosis per infected animal were estimated by dividing total losses for each species with the number of infected animals for that species (Table 1). Due to lack of data, the health and economic burden of human brucellosis could not be accounted into the overall benefits of the control programs.
The analyses were conducted using R-statistical program (R statistical package version 2.12.0, R Development Core Team, http://www.r-project.org) and we run Monte Carlo simulations for 10,000 iterations so as to determine confidence limits for these estimates.
For each intervention, our results indicate that the prevalence of Brucella infection will drop below 2% in cattle after 20 years of the implementation of disease control program (Fig 1) for the cattle population. For buffaloes, a similar trend was observed. However, due to the higher initial prevalence of infection, it only drops below 3% after the twenty year implementation of all interventions (Fig 2).
The NPV during the first 20 years of the program for cattle for scenario 1, 2 and, 3 are presented in Table 2. For cattle, the NPV was found to be US $ 4·16 billion (95% CI: US $ 3·16; 5·39 billion) for the scenario 1, US $ 8·31 (6·40; 9·87) billion for the scenario 2 and, US $ 4·26 (3·26; 5·61) for the third scenario (Table 2). The results indicate that first 20 years of the programme will be cost-effective for all scenarios with the second intervention (vaccination of replacements with mass vaccination at the beginning of the program) being a significantly more cost-effective choice. The BCR for the first, second and the third intervention for cattle were 7·98 (6·29; 10·09), 10·62 (8·33; 12·5) and, 3·16 (2·66; 3·83), respectively. Similar results were obtained for buffaloes (Table 3).
The NPV for the 50% prevalence estimates during the first 20 years of the program for cattle for scenario 1, 2 and, 3 are presented in S1 Table. For cattle, NPV was found to be US $ 1·78 billion (95% CI: US $ 1·07; 2·79 billion) for the scenario 1, US $ 3·27 (2·18; 4·23) billion for scenario 2 and, US $ 0·87 (-0·24; 1·71) for scenario 3 (S1 Table). For buffaloes, the NPV for the 50% prevalence estimates was found to be US $ 3·69 billion (95% CI: US $ 2·74; 4·54 billion) for the scenario 1, US $ 5·96 (4·40; 7·27) billion for the scenario 2 and, US $ 2·09 (1·02; 3·20) for the third scenario (S2 Table).
Further, sensitivity analysis revealed that for either species, disease prevalence will further reduce to less than 1% after 50 years of implementation for either intervention and will virtually lead to eradication of the disease after 100 years of the implementation programme. The long time to eradicate infection is based on the fact that we only considered realistic vaccination coverage rates. Our primary analysis, assumed a vaccination coverage of 70%. Reduction of the vaccination rate led to reduced NPV and BCR values. The same impact had the assumption that herds that were not covered were more likely to remain uncovered the next year.
This is the first systematic analysis of a brucellosis control program interventions for bovine brucellosis in India. Bovine brucellosis is highly prevalent in India and causes significant losses to the livestock industries. The results suggest that all of the three approaches investigated for controlling the disease would be beneficial as the prevalence of Brucella infection will drop below 2% in cattle and 3% in buffalo after 20 years of the implementation of disease control program. All programs had positive NPVs and >1 BCRs indicating the benefits from all programs are higher than their respective costs. The best BCR was obtained in the second intervention, i.e. vaccination of both the adult and young population at the beginning of the program followed by an annual vaccination of the replacement calves. It leads to a significant drop in prevalence at the beginning of the program and hence the risk of transmitting the disease in the subsequent years is lower. This is thus the most cost-effective approach for control of brucellosis in India. Overall, the results advocate the implementation of a disease control program for brucellosis in India.
The results of sensitivity analyses indicated that a positive effect for all interventions and a net benefit of billions of dollars for any intervention remains even after considering significantly reduced initial prevalence and vaccination coverage. This suggests that the control program would be beneficial even if some of the assumptions used in the model are changed, further supporting the implementation of a control program for the disease.
It must be noted that we only considered economic benefits of the control programs for the livestock populations. The benefits such as disability-adjusted life years (DALYs) and social losses averted due to the control programs could not be accounted. Similarly, the extra costs due to increased livestock numbers (feed costs), or unintentional consequences (abortion due to vaccinating a female cow) were not estimated. However, we believe that this will not have a major impact on the results of the current study.
Although the third approach–i.e. annual mass vaccination of replacements for a decade followed by a decade of a test and slaughter strategy–was also found to be cost-effective, it is less likely to be adopted in India because it is a Hindu majority country and Hindus consider cows to be sacred. As a result, cow slaughter is banned in most states of India. Thus it would be difficult to get community support for a strategy involving animal slaughter although it drastically reduces prevalence of the disease if implemented after a decade of vaccination. It will also be more expensive as it would involve testing of animals which would include sample collection, transport and laboratory testing. Also, there would be additional costs involved for culling infected animals. Therefore, this may not be the preferred strategy in the Indian situation. Moreover, in calculating losses for test and slaughter, we assumed that animals will be consumed after slaughter in accordance with the WHO guidelines [28]. However, it may not be feasible to do so or may increase the risk of spread of infection. Therefore, it would be more sensible to adopt a ‘test and euthanasia’ strategy in which the infected animals are euthanized and their carcasses burnt or buried and not consumed. This strategy is likely to have a greater acceptance among the community which is very essential for the success of any control program. However, this would increase the cost of the test and slaughter program as it will result in a complete loss of slaughtered animal instead of just a loss of 20% considered in the scenario. Thus the actual cost of the third scenario may be higher than we estimated.
In this study, BCRs for three inventions for cattle were estimated to range from 3·16 to 10·62 and for buffaloes from 3·79 to 21·27. Similar estimates have been obtained in some other studies conducted around the world. The strategy of vaccinating 3–6 month old female bovine, male ovine and female ovine followed by compulsory slaughter after attaining the target prevalence have been advocated in Turkey [10], where BCR was estimated to be 2·26 [10]. A BCR of 3·2 has been estimated for control of brucellosis in Nigeria [29]. A national serological survey and risk based vaccination using S19 and an awareness program was found to have a BCR of 6·8 for control of brucellosis in Nepal [9].
Note that we only considered scenarios for control of the disease; eradication was not considered feasible in the current circumstances. The disease is highly prevalent and endemic in India; therefore, it would be unrealistic to achieve eradication. Further, India is a vast country with significant movement and intermixing of animals. Moreover, eradication would definitely require test and slaughter but religious and cultural beliefs would impede implementation of any such program due to limited community support. However, once the prevalence reduces below 2% after 20 years, there may be greater community support for eradication as well as test and slaughter/euthanasia as discussed before. Therefore, it would be wise to revisit this question sometime in the future.
In this work realistic inputs of vaccination coverage aimed to also adjust for the reduced vaccine efficacy due to vaccine failure as well as problems associated with cold chains, which were not directly accounted for. Assumed vaccination rates ranged from 70% to 50% to cover different vaccine efficacies that have been used in previous studies [25, 30]. Undoubtedly, real-time data of vaccine efficacy could further improve the predictive ability of our model. However, in our sensitivity analysis we considered the realistic fact that herds that were not covered once were more likely to remain uncovered due to issues associated with inability to reach them or farmers’ will to cooperate. However, even better NPV and BCR will be achieved if the vaccination coverage/efficacy is improved.
It has been reported that 53.6% of the bovine (cattle and buffalo) population receives foot and mouth disease (FMD) vaccination in India [31, 32]. Therefore, our assumption of 50% vaccination coverage is quite realistic. However, there will be pockets of low (<10%) and high coverage (90%+) areas. Many factors such as poor infrastructure, lack of knowledge and veterinary personnel availability are responsible for poor adoption of vaccines in India [33]. A farmer’s perceptions such that vaccination could lead to decrease in milk yield, swelling and fever also decrease vaccine coverage [33]. Low community acceptance, vaccine stock outs at the local level and timeliness of vaccine also affect the vaccine coverage [34]. These factors could affect the benefits but could not be accounted in the current study.
The advantage of using the S19 vaccine is that immunity induced is long-lasting and has been reported to be effective till fifth pregnancy [10, 35]. However, there are a number of concerns with using this vaccine. The major concern is the common occurrence of the needle stick injuries and the accidental inoculation in veterinary personnel while participating in Brucella vaccination programs [36]. The rates of accidental exposure ranging from 6·7% to 46% have been reported [37]. The needle stick injuries have been reported to cause a low virulence human brucellosis [38]. To avoid needle stick injuries, research should be conducted in the use of a safety vaccinator as used for other vaccines such as Gudair® vaccination in Australia [39] and animals should be properly restrained before vaccination.
Additionally, the S19 vaccine could interfere with the recommended diagnostic tests and may cause abortion in the pregnant animals [35, 40]. Moreover, the vaccine cannot be used for male [41] or infected animals [42]. Therefore, there is a need for the development of a better vaccine that can differentiate infected from vaccinated animals (DIVA). RB51 vaccine can be used instead of S19 as it allows serological differentiation between naturally infected and vaccinated animals but it is not currently available in India and is considered to have a lower efficacy than S19.
It is worth mentioning here that the cost and benefit analyses evaluated in this manuscript only pertain to the effect of the disease on the domestic animal population. The benefits to the human population would be over and above the benefits discussed here but were beyond the scope of this study. It is well known that humans get infected while handling infected animals. Therefore, various studies have shown that the disease is prevalent among occupational groups such as veterinary personnel, laboratory workers, livestock farmers and abattoir workers in India [43–45]. We have recently shown that the disease causes a loss of 177 601 (95% UI 152 695–214 764) DALYs at the rate of 0.15 (95% UI 0.13–0.17) DALYs per thousand persons every year [46] and an annual median loss of Rs 627.5 million (US $ 10.46 million) in India [46]. Complete eradication of the disease will save these losses but further studies are required to investigate the real impact of the control strategies discussed in this manuscript on the human population.
|
10.1371/journal.pcbi.1006055 | Computational analysis of the oscillatory behavior at the translation level induced by mRNA levels oscillations due to finite intracellular resources | Recent studies have demonstrated how the competition for the finite pool of available gene expression factors has important effect on fundamental gene expression aspects. In this study, based on a whole-cell model simulation of translation in S. cerevisiae, we evaluate for the first time the expected effect of mRNA levels fluctuations on translation due to the finite pool of ribosomes. We show that fluctuations of a single gene or a group of genes mRNA levels induce periodic behavior in all S. cerevisiae translation factors and aspects: the ribosomal densities and the translation rates of all S. cerevisiae mRNAs oscillate. We numerically measure the oscillation amplitudes demonstrating that fluctuations of endogenous and heterologous genes can cause a significant fluctuation of up to 50% in the steady-state translation rates of the rest of the genes. Furthermore, we demonstrate by synonymous mutations that oscillating the levels of mRNAs that experience high ribosomal occupancy (e.g. ribosomal “traffic jam”) induces the largest impact on the translation of the S. cerevisiae genome.
The results reported here should provide novel insights and principles related to the design of synthetic gene expression circuits and related to the evolutionary constraints shaping gene expression of endogenous genes.
| Each cell contains a limited number of macromolecules and factors that participate in the gene expression process. These expression resources are shared between the different molecules that encode the genetic code, resulting in non-trivial couplings and competitions between the different gene expression stages. Such competitions should be considered when analyzing the cellular economy of the cell, the genome evolution, and the design of synthetic expression circuits. Here we study the effect of couplings and competitions for ribosomes by performing a whole-cell simulation of translation of S. cerevisiae, with parameters estimated from experimental data. We demonstrate that by periodically changing the mRNA levels of a single gene (endogenous or heterologous) or a set of genes, the translation of all S. cerevisiae genes are affected in a periodic manner. We numerically estimate the exact impact of the mRNA levels periodicity on the translation process dynamics, as well as on the dynamics of the free ribosomal pool and the way it is affected by parameters such as the codon composition of the oscillating gene, its initiation rate and mRNA levels. Furthermore, we show that the codon compositions of synthetically highly expressed heterologous genes that are expected to oscillate must be carefully considered. For example, synonymous mutations resulting in “traffic jams” of ribosomes along the fluctuated mRNAs may cause significant fluctuations of up to 50% in the steady-state translation rates of all genes.
| During the gene expression process various macromolecules (e.g. ribosomes, RNA polymerase (RNAP), transcription factors, elongation factors, spliceosome, transfer RNA (tRNA) molecules, etc.) process the genetic material (DNA, mRNA, pre-mRNA) in order to generate proteins [1]. The number of gene expression macromolecules and factors in the cell is finite; for example, there are about 200,000 ribosomes and 30,000 RNAP-II molecules in the S. cerevisiae cell [2, 3]. Thus, this limited resource budget induces competition between the different molecules/regions encoding the genetic material, resulting in non-trivial correlations and couplings between the different gene expression stages, and between the processed genetic material molecules.
Some previous studies have suggested that such competition should be considered when designing synthetic gene expression circuits [4–9], and that they significantly affect the evolution of genomes [10]. For example, [9] considered a stochastic model to analyze the competition of two types of mRNAs (two genes) for the limited ribosomal resource, where the total number of mRNAs and ribosomes fluctuate randomly. It was shown that the strength of the couplings (or cross-talk) between the translation of the two protein types strongly depends on whether the ribosomes are underloaded (i.e., there are more ribosomes than mRNAs) or overloaded (i.e., there are more mRNAs than ribosomes).
Specifically, it was also suggested that the competition for the limited main resources in transcription (RNAP [11]) and translation (ribosomes [12]) is a primary factor in the cellular economy of the cell. The competition for the available resources, which leads to an indirect coupling between expressions of different genes, might be one of the reasons why levels of genes, mRNAs, and proteins in the cell do not necessarily correlate [4, 10, 12–16].
The expression levels of large sets of genes and relevant gene expression factors are fluctuating or oscillating in different physiological conditions (e.g. cell cycle [17–21]). In addition, there are many cases of oscillating genes that are significant (up to hundreds of genes oscillating with a ratio of up to about three folds between highest to lowest mRNA levels) in all domains of life [22–34]. Furtheremore, various synthetic circuits and cell free systems include oscillators [35–42]. The couplings, due to competition, may link the oscillations related to one gene expression stage (e.g. transcription) to oscillations in a different gene expression stage (e.g. translation). In this study, we suggest for the first time that finite intracellular resources induce non-trivial and significant coupling between different gene expression stages (transcription and translation) in endogenous and heterologous genes. For example, increased mRNA levels in one gene affects the translation levels of all other genes. To this end, we perform a whole-cell simulation of translation [43, 44], which captures fundamental properties of translation, with parameters estimated from experimental data that enables us to comprehensively quantify these effects for the first time. This type of information is currently not available experimentally, and we believe that our results are expected to reflect well the reality.
We specifically demonstrate by Monte Carlo simulations that by periodically changing the mRNA levels of a single gene or a set of genes, i.e. by periodically modifying the transcription process, the translation of all S. cerevisiae genes are affected in a periodic manner, with the same periodicity as the mRNA levels periodicity. Importantly, we numerically estimate, for the first time, the exact impact of the mRNA levels periodicity on the translation process dynamics, as well as on the dynamics of the free ribosomal pool and the way it is affected by parameters such as the codon composition of the oscillating gene, its initiation rate and mRNA levels.
We utilize a large-scale, whole-cell computational model for simultaneous mRNA translation and competition for ribosomes to study the effect of mRNA levels fluctuation on the translation process [43–45]. The model considers all the fundamental properties of translation such as the different decoding times of codons and their order, the excluded volume interactions between ribosomes, the finite pool of ribosomes shared by all mRNAs, initiation rates, etc. [43–45]. The dynamics in this model is expressed by a set of ordinary differential equations describing the time evolution of the ribosomal occupancies in the different positions along the mRNAs, and the time evolution of the free ribosomal pool. It was shown that this computational model provides predictions with high correlation with protein levels and ribosome density measurements (see more details in the Materials and methods section).
We use the model to simulate translation of the S. cerevisiae genome including a heterologous green fluorescent protein (GFP) gene (with different codons compositions), while periodically modifying the mRNA levels of the GFP gene or of a subset of the endogenous genes. The competition for the limited, shared ribosomes, results in an indirect coupling between the translation processes of the different genes. We measure the coupling effect on the free pool of ribosomes, and on the translation rates and ribosomal densities of the different translation process as a function of the oscillating mRNA (or mRNAs) parameters and mutations.
Fig 1(a) depicts our study flow diagram. We consider the S. cerevisiae genome and a GFP gene, while periodically controlling the transcription (i.e. the mRNA levels) of one or more S. cerevisiae genes or the GFP gene. All mRNAs are then simultaneous translated, while competing for the ribosomal resource. We then measure different translation parameters, such as translation rate and ribosomal density of all genes, and the free ribosomal pool (i.e. the number of free ribosomes). The block diagram of the model we use for simultaneous translation and competition is depicted in Fig 1(b). Please refer to the Materials and methods section for a detailed description of the model. Finally, Fig 1(c) shows an example of the translation parameters behavior when periodically controlling the mRNA levels of the GFP gene. The figure depicts the free ribosomal pool, and the GFP translation rate and mean ribosomal density as a function of time. It may be observed that these oscillate with a common periodicity.
It is important to mention that all the parameters used in the computational model (e.g. codon compositions, codon decoding times, mRNA levels, and number of ribosomes) were inferred based on experimental measurements of S. cerevisiae, and based on known properties of the GFP. Specifically, the S. cerevisiae genome consists of m ≔ 6310 protein-encoding genes with ORFs ranging from as low as 25 codons to as high as 4911 codons (see S4 Fig). The GFP gene ORF consists of 240 codons (see more details in the Materials and methods section).
Table 1 lists the parameters used throughout the simulations and their source.
Let Lh denote the oscillating gene nominal mRNA levels and α its initiation rate. Recall that the oscillating gene (or genes) can be either an endogenous gene or the GFP heterologous gene. We periodically change the mRNA levels of the oscillating gene as follows:
lh(t)=Lh(1+Asin(2πtT)), (1)
where ℓh(t) denotes the oscillating gene mRNA levels at time t, A ∈ [0, 1) is the normalized amplitude, and T is the period time. Let LT denote the total number of (S. cerevisiae) mRNAs in the cell (i.e. LT ≔ 60, 000), and
L ˜ h ≔ 100 L h L T ,
the oscillating gene nominal mRNA levels in percentage of LT.
Let Ri denote the average steady-state translation rate of gene i, R a m p i denote its steady-state amplitude, and
R a i ≔ 100 R a m p i R i ,
denote the steady-state translation rate amplitude of gene i in percentage of its average steady-state translation rate Ri. Let
R ¯ a ≔ ∑ i = 1 m R a i m ,
denote the average (over all genes) steady-state translation rate amplitude (in percentage).
In the same manner let (see Fig 1(a))
See an example of these parameters in S5 Fig. The variance parameters provide indication of how the individual genes amplitude vary relative to the average. A large [small] variance implies that the genes amplitudes are widely [closely] scatter relative to the average.
As the intent in this paper is to analyze the impact of oscillations on the translation process, due to the shared, limited ribosomal resource, we believe that quantifying the above parameters under different conditions is essential in understanding the impact on the translation process.
In the following sections we present numeric measurements of the impact of periodically modifying the mRNA levels of endogenous genes or the GFP gene (or its mutations) on the above parameters over all the S. cerevisiae genes.
At the first step, we aim at evaluating the impact of fluctuating mRNA levels of a S. cerevisiae endogenous gene set on the translation of the entire S. cerevisiae transcriptome. Fig 2 panels (a) and (b) depict the results as a function of the number of oscillating endogenous genes. In the figure we plot both the effect of the average/typical oscillating gene set, and the effect of the oscillating set with maximal mRNA levels (see the Materials and methods section for more details).
As can be seen, oscillating the mRNA levels of a typical large set of 1, 000 S. cerevisiae genes with normalized amplitude A = 1/2 is expected to typically induce an amplitude of about 9% on the rest of the genes translation rates; the maximal effect of a set of 1, 000 S. cerevisiae genes is very high and close to 50%. During the life cycle of a cell large sets of genes may fluctuate/oscillate together (e.g. due to a common regulatory mechanism) at the transcription level and the results reported here demonstrate that these oscillations should have non-negligible effect on the rest of the genes at the translation levels. Note that the results for za and ρ ‾ a when oscillating a typical gene set are very similar (the solid-line for za cannot be distinguished from the solid-line for ρ ‾ a).
One phenomena that involves large scale gene expression oscillation is the cell cycle process. Ref. [17] identified 800 protein-encoding transcripts in S. cerevisiae that are cell cycle regulated, i.e. genes whose transcript levels vary periodically during the cell cycle process. These genes are involved in different cell cycle related functions such as cell cycle control, DNA replication, DNA repair, budding, glycosylation, nuclear division and mitosis. We evaluate the effect of oscillating these genes on the translation of the rest of the genes as a function of the normalized amplitude A ∈ [0.1, 0.9]. These are depicted in Fig 2 panels (c) and (d). It may be noticed that the amplitudes increase linearly with A, and that the amplitude of the free ribosomal pool and the translation amplitudes are very similar. Note that the variance of the steady-state mean density amplitude hardly change as a function of A, whereas the variance of the steady-state translation rate amplitude increases from zero to about 0.12 for A = 0.9. This suggests that the steady-state mean density amplitudes of all S. cerevisiae genes vary much less than the corresponding steady-state translation rate amplitudes. Since ρ a i measures the average of the steady-state density amplitudes of gene i, it is indeed expected that its variance over all genes will be less than the variance of the steady-state translation rate amplitudes over all genes.
Next, we aim at understanding the effect of oscillating the mRNA levels of a heterologous gene on the free ribosomal pool, and on the translation rate and ribosomal density of the endogenous genes. Note that there are many synthetic systems where the mRNA levels of a single heterologous gene occupy dozens of percentages of the total number of mRNAs in the cell (see, e.g., [49, 50]). This analysis should specifically provide some intuition related to the effect of synthetic gene expression oscillation circuits on the translation of the rest of the genes. (Note that there are many examples of synthetic genes with oscillatory mRNA levels [36, 51–57]). It should also teach us about the effect of fluctuations in the expression levels of highly expressed heterologous genes on the expression levels of the rest of the genes. To this end, we add to our whole-cell model a heterologous GFP gene with periodically varying mRNA levels.
Fig 3 depicts the average steady-state translation rate amplitude (R ‾ a) and mean density amplitude (ρ ‾ a) for different (typical) values of GFP nominal mRNA levels L ˜ h and initiation rates α for A = 1/2, T = 16, and z ‾ = 30 %. It may be seen that both R ‾ a and ρ ‾ a increase with both α and L ˜ h. This is expected since increasing α or L ˜ h increases the dynamic assignment of ribosomes to the GFP mRNAs, which in turn increases the impact of ribosomes assignment to the S. cerevisiae genes via the shared pool. For example, for L ˜ h = 30 %, R ‾ a [ρ ‾ a] ranges from about 2.5% [2.5%] to about 13.5% [14%]. Another observation is that the impacts on R ‾ a and ρ ‾ a are very similar. This suggests that by measuring the periodic amplitude of the translation rates at steady-state one can reasonably conclude the average amplitude of the mean ribosomal densities at steady-state.
Fig 3 also depicts R ‾ a and ρ ‾ a as a function of A ∈ (0, 1/2] for α = 0.8, L ˜ h = 20 %, T = 16, and z ‾ = 30 %. It may be noticed that the translation rate and ribosome density increase linearly with A ∈ (0, 1/2]. Similar observations were made for several other values of L ˜ h and α. We conclude that highly expressed heterologous genes can have an effect of up to about 20% on the amplitude of the translation rate and ribosome density of the rest of the endogenous genes. This should be considered when designing the properties of a synthetic circuit. Note that by the analysis done in the previous section, oscillations of large number of endogenous genes should also affect the heterologous genes.
In this section different synonymous substitutions are introduce to the heterologous GFP gene to study their effect, separately, on the translation of the endogenous genes. The goal here is to evaluate the effect of the coding region (and thus the induced ribosomal density and translation rate) on translation oscillation. In brief, we consider various variants of the GFP coding region; all of them code the same GFP protein but with different codons (a detailed description of each synonymously mutated GFP can be found in the Materials and methods section). The mutated GFP genes considered are:
Table 2 lists the steady-state translation rate R and mean densities ρ of each mutated GFP modeled to include initiation rate equals to 0.8 (which is the median initiation rate of the S. cerevisiae genome [47]). The table also lists two metrics (η and η ˜) for ranking the codon decoding times of the coding region (named decoding time measure (DTM)). The DTM provides a score of how fast the ORF can be decoded; a value of zero means that it is composed of the fastest synonymous codons, and a larger value of DTM indicates that slower codons are used in the ORF. Specifically, in η all codons contribute equally to the DTM, whereas in η ˜ the codon impact on the DTM increases as we move closer to the 3’-UTR end of the gene (see the Materials and methods section for more details).
The following may be concluded from Table 2:
Fig 4 depicts the translation normalized statistics as a function of the nominal mRNA levels L ˜ h, for α = 0.8 and α = 3.2, for each of the GFP mutated genes, when translated (separately) with the S. cerevisiae gene pool, for A = 1/2, T = 16, and z ‾ = 30 %. Each data point in the figure represents the corresponding statistics per 600 mutated GFP mRNAs (1% of the total S. cerevisiae mRNA levels), i.e. we divide the statistics values by the corresponding L ˜ h and multiply by 600. This represents the impact on the translation process per a unit of 600 GFP mRNAs (a normalized measure can then be used to compare results between different values of L ˜ h).
We first observe that the normalized statistics increase with α for each L ˜ h value. This is obviously expected since large values of α imply high periodic variations of assigned ribosomes to the GFP mRNAs, and thus also to the S. cerevisiae genes mRNAs (due to the shared pool), and so we expect the amplitudes of the free pool, translation rates and mean densities to increase. We also note that the statistics variations over the different mutations increase with α. For example, for L ˜ h = 25 %, the normalized za varies between 0.55% and 0.67% (in case of α = 3.2), and between 0.35% and about 0.4% (in the case of α = 0.8). This is expected since, for example, a low value of α means that the initiation is the rate limiting factor, and in this case the GFP ORF mutations (affecting the elongation rates) less affect the parameters.
In addition, it may be seen that the normalized statistics maintain a particular ranking for different L ˜ h values: they achieve their maximal values when oscillating the GFP_HIGH_RD mutation, are reduced when oscillating the GFP original gene, and achieve their minimal values when oscillating the GFP_LOW_RD mutation. For example, for L ˜ h = 10 % and α = 3.2, the normalized R ‾ a is about 0.57% when oscillating GFP_HIGH_RD, is about 0.55% when oscillating GFP, and is about 0.45% when oscillating GFP_LOW_RD. This correlates with the mean steady-state ribosomal densities of these mutations, as well as with their non-homogeneous DTMs (η ˜). This suggests that mRNAs with “traffic jams” at steady-state (i.e. mRNAs that occupy large number of ribosomes at steady-state) have a substantial impact on the translation of the other genes via the shared ribosomal pool.
The normalized za, R ‾ a and ρ ‾ a seem to slightly decrease with L ˜ h, implying that the non-normalized parameters increase sub-linearly with L ˜ h. Oscillating the mRNA levels of the GFP gene increases and decreases periodically the assigned number of ribosomes to the GFP mRNAs, which in turn decreases and increases periodically the amount of free ribosomes, respectively. This affects the actual initiation rate to the mRNAs. However, due to the finite, shared pool of ribosomes, the oscillation effect caused by an increase in mRNA levels admits a linear region which is eventually saturated (similar to most physical systems). Finally, it may be observed that the corresponding variances increase with both L ˜ h and α, indicating, as expected, that for large oscillating mRNA levels, and/or initiation rates, the variations of the amplitudes over all genes increase. Note that the variance values are few order of magnitudes less than the corresponding average values; for example, for L ˜ h = 20 % and α = 3.2, R ^ a [ρ ^ a] is about 0.7% [0.03%] of the corresponding average values.
In general, both the steady-state translation rate and the mean density of the mutated or the original GFP gene affect the parameters. For example, the effect of the gene GFP_SPD_TR on the statistics is less severe than the effect of the gene GFP_MDN_RD, even-though GFP_SPD_TR consumes more (by 25%) ribosomes at steady-state (see Table 2). However, the steady-state translation rate of GFP_SPD_TR is larger (by about 26%) than the steady-state translation rate of GFP_MDN_RD, thus ribosomes in the GFP_SPD_TR mutation case are released faster to the pool and thus are available more for translating other genes.
In addition, we can observe a ‘diminishing marginal utility’ effect: the results depicted in Fig 4 suggest that oscillating a larger number of mRNAs of the mutated GFP gene decreases the amplitude of the free pool and of the genes translation rate and mean density per GFP mRNA level. This effect is partially due to the limited and shared ribosome pool.
Fig 5 depicts the translation normalized statistics when oscillating the mutated GFP_HIGH_RD gene for several values of the average steady-state free ribosomal pool z ‾. It may be seen that za decreases with z ‾, whereas R ‾ a and ρ ‾ a are slightly affected by z ‾. For example, for L ˜ h = 30 % and α = 0.8, the normalized za decreases from about 1.4% for z ‾ = 10 % to about 0.4% for z ‾ = 30 %, whereas both the normalized R ‾ a and ρ ‾ a hardly vary and are equal to about 0.32% and 0.34%, respectively. One possible explanation for this is as follows: As z ‾ decreases (i.e. as less ribosomes are free thus more are assigned to the mRNAs) the effective initiation rates to the mRNA increases. This increases the oscillation amplitude induced by the GFP_HIGH_RD mRNAs, and thus the relative effect on z ‾ increases (recall that za denotes the free pool oscillation amplitude relative to z ‾). On the other hand, an increase in the effective initiation rates increases both the steady-state translation rates and mean densities of all the S. cerevisiae mRNAs, and so the effect on R ‾ a and ρ ‾ a is small. However, as suggested by Fig 5, the corresponding variances increase slightly as z ‾ decreases, implying that the amplitude variations over all genes do not change much as z ‾ decreases from 30% to 10%. Note that, again, the variance values are few order of magnitudes less than the corresponding average values (for example, for L ˜ h = 20 %, R ^ a [ρ ^ a] is about 1.0% [0.05%] of the corresponding average values).
The results depicted in Fig 5 suggest that the fluctuations of the translation rates and mean ribosomal densities are hardly affected by the affinity of ribosomes to the mRNA molecules (this affinity may be controlled by initiation efficiency, for example). However, the fluctuations of the free ribosomal pool increase as more ribosomes are translating the mRNA molecules.
As another example, Table 3 depicts the (non-normalized) statistics when oscillating the mutated GFP_HIGH_RD gene for A = 0.35, T = 16, α = 0.8, L ˜ h = 20 %, and for several values of the average free ribosomal pool at steady-state z ‾. The same conclusions can be derived here as well.
In summary, the current subsection teaches us that when designing highly expressed heterologous genes that are expected to fluctuate/oscillate we should carefully choose their codons composition: to induce low effect on the other genes we should minimize the ribosome density, and on the other hand high ribosomal density results in large effect on the other genes. As was demonstrated here the exact profiles that maximize/minimize ribosome densities are not simply the ones with optimal/slowest codons along the coding region, respectively; thus, it is important to develop models and algorithms for engineering and manipulating ribosome density of endogenous and heterologous genes.
The results reported here with the heterologous gene may be further validated experimentally in the future using in-vitro and/or in-vivo systems with oscillating GFP proteins [51]. However, we believe that with the current experimental approach it should be challenging to directly study the coupling we reported here due to the following reasons. First, oscillating endogamous systems probably includes various effects and feedbacks that may “cancel” or blur the phenomena presented here. Second, in order to be able to measure, with the current techniques, the effect reported here large portion of the mRNA molecules in the cell should oscillate. Finally, this study analyzes oscillations during the translation stage. Thus, to study them one should directly measure translation rate; the conventional experimental approaches (e.g. RNA-seq, ribo-seq or approaches based on quantitative mass spectrometry) measure variables that are expected to be related/correlated with the translation rate but are not the actual translation rate.
Our results should be specifically considered when designing large intra-cellular circuits with many components/genes. In such cases, among others, the oscillation in transcription levels of some parts of the circuit should affect the other part of the circuit. We provide here some initial guidelines related to this topic. First, if we are not interested in cross-talk between the different oscillating genes we should engineer their transcript to minimize the induced oscillations (e.g. designing codon profiles that minimize ribosome density and if possible decrease their initiation rate). Second, in some cases we may want to design genes that induce oscillations on the rest/other genes; in these cases, we will design them accordingly (e.g. high initiation rate and ribosome density). Third, our (or similar) models can be used to estimate potential “noise” due to oscillation cross-talk. These estimations can be considered when designing the circuit and assuring its performance.
The goal of this study is to understand and carefully quantify the impact of oscillations over wide range of conditions and parameters (e.g., large range of A, L ˜ h, α, z ‾, and different GFP mutations). This is important, as the severity of the oscillations impact (in terms of its phenotypic or biological-significant effect) is, in general, gene and condition specific. It might depend on the function of the genes (e.g. structural genes, transcription factors, signaling proteins, etc.), the exact condition (e.g. initiation rate, mRNA levels), the organisms type, etc.
The computational model used in this study is deterministic, enabling rigorous analysis of its properties using tools from systems and control theory. In addition, it was shown to admit high correlation with the stochastic TASEP model of translation (e.g. see [45]), and furthermore using it to simulate large-scale translation with competition is simple. The processes in the cell are stochastic in nature, and future study may employ stochastic whole cell models to study the effect of oscillations in a “noisy” environment. In S6 Fig we provide initial results that suggest that noise in the model parameters should not affect our conclusions.
It is important to emphasize that the results reported here are relevant also in cases where the time scales of translation and cell cycle differ. Note that it has been suggested that translation of cell cycle related genes is regulated by periodically varying tRNA levels [18]. This implies, among others, that the time scales are quite similar. Specifically, the translation time in general can be longer than the cell cycle period. For concreteness, consider the case of S. cerevisiae. The cell cycle period in S. cerevisiae is less than 87 minutes [58]. Cell cycle period can be much shorter in eukaryotes; for example, it was reported that the duration of cell cycle in early embryo of the fruit fly D. melanogaster is only eight minutes [59]. The translation rate in S. cerevisiae was estimated to be higher than 0.956 codons per second (the slowest codon is CUU) [60] with average rate over all codons of 10 codons per second (in mouse the average codon translation rate was estimated to be about five codons per second [61]). In practice, this rate can be much slower due to strong folding of the mRNA molecule and interaction of the translated amino-acid peptide with the exit channel of the ribosome [62, 63]. In S. cerevisiae the ORF length is between 75 and 14, 733 nucleotides. The longest gene corresponds to an upper bound on the translation time of a gene, which is about 82 minutes (assuming a lower bound on translation rate of one codon per second, which may be lower in practice), an estimated translated time of this protein based on mean codon translation time is 8.2 minutes. In mammals the mean codon decoding time is five codons per second, and the longest human protein (Titin–TTN), which consists of 33, 000 amino acids, corresponds to estimated translation time of 110 minutes. This suggests that periodically varying mRNA levels in cell cycle related genes may be similar to the time scale of mRNA translation.
We will conclude with the main lessons from the analysis performed here based on our whole-cell computational model: 1) Competitions for limited resources in the cell lead to indirect couplings between the gene expression stages, and these couplings must be considered when analyzing the cellular economy of the cell; 2) A whole-cell computational model of translation that takes into account fundamental properties of translation, with parameters estimated based on experimental measurements, can comprehensively quantify the effect of oscillations on the ribosomal densities and translation rates of all genes; 3) Careful considerations must take place when designing highly express heterologous genes that are expected to fluctuate, as their codon compositions and translation initiation rates may have high effect on all genes translation rate. We demonstrate specific cases with high and low effect on fluctuations; 4) Quantitative estimation (based on parameters estimated from experimental data) of the magnitude of these oscillations in endogenous and heterologous genes is provided here; and 5) The conclusions reported here in general should also be relevant to other aspects of gene expression and/or intracellular phenomenon. For example, when considering oscillations in tRNA levels, the number of DNA copies of a virus, intracellular transport factors, etc.
We first sort the S. cerevisiae genes according to mRNA levels and evaluate the steady-state mean density and translation rate amplitudes as a function of the number of oscillating genes chosen sequentially from the sorted list of genes, starting from the gene with the largest mRNA levels (dashed-lines in Fig 2, panels (a) and (b)). For example, when using p genes with oscillating mRNA levels, the p genes with the largest mRNA levels are used. This provides a bound on the maximal oscillating amplitudes when any arbitrary p S. cerevisiae genes are oscillating.
The “typical” selected genes were chosen randomly from the S. cerevisiae gene pool. Here the oscillation amplitudes and variances for each number of “typical” oscillating genes set is averaged over 30 repetitions. The results of oscillating these genes are depicted, using solid-lines, in Fig 2, panels (a) and (b).
The assumption in this study is that large set of genes can be regulated (oscillate) independently of the ribosomal pool. This is motivated by: 1) The regulations at the translation and transcription stages are not tightly coupled [64]; 2) There may be delays between the two stages [65]; 3) In the case of heterologous genes (and the corresponding promoters and gene expression circuits) that are engineered by design there is no reason to assume that the ribosome pool is also regulated.
Ref. [17] identified 800 protein-encoding transcripts in S. cerevisiae that are cell cycle regulated. We evaluate the parameters when oscillating 770 of these genes, as we lack mRNA measurements for 30 of the reported 800 cell cycle related genes. The 770 cell cycle genes used are listed in S1 Table. Table 4 lists the 30 genes we lack mRNA measurements for.
The ribosome flow model network with pool (RFMNP) [43] is a deterministic computational model for large-scale simultaneous mRNA translation and competition for ribosomes. It is based on combining several ribosome flow models with input and outputs (RFMIOs) [45, 66], interconnected via a pool of free ribosomes. Each gene is modeled by a single RFMIO, and all the RFMIOs are sharing the same pool of ribosomes. The dynamics of the system is expressed by a set of ordinary differential equations that describes the time evolution of the ribosomal densities in the different RFMIOs and the free pool. In this paper we utilize the RFMNP to simulate a whole-cell S. cerevisiae simultaneous translation and competition for ribosomes. Each of the S. cerevisiae gene (and the GFP gene) is modeled by a single RFMIO. We next describe in details the RFMIO and the RFMNP.
The ribosome flow model (RFM) [45] is a deterministic mathematical model for mRNA translation that can be derived by a mean-field approximation of an important model from statistical physics called the totally asymmetric simple exclusion process (TASEP) (see, e.g., [67] and [68]). In the RFM, mRNA molecules are coarse-grained into n consecutive sites of codons. The state variable x i ( t ) : ℝ + → [ 0 , 1 ], i = 1, …, n, describes the normalized ribosomal occupancy level at site i at time t, where xi(t) = 1 [xi(t) = 0] indicates that site i is completely full [empty] at time t. The model includes n + 1 positive parameters that regulate the transition rate between the sites: the initiation rate into the chain λ0, the elongation (or transition) rate from site i to site (i + 1) λi, i = 1, …, n − 1, and the exit rate λn (see Fig 6).
The dynamics of the RFM with n sites is given by n nonlinear first-order ordinary differential equations:
x ˙ 1 = λ 0 ( 1 - x 1 ) - λ 1 x 1 ( 1 - x 2 ) , x ˙ 2 = λ 1 x 1 ( 1 - x 2 ) - λ 2 x 2 ( 1 - x 3 ) , x ˙ 3 = λ 2 x 2 ( 1 - x 3 ) - λ 3 x 3 ( 1 - x 4 ) , ⋮ x ˙ n - 1 = λ n - 2 x n - 2 ( 1 - x n - 1 ) - λ n - 1 x n - 1 ( 1 - x n ) , x ˙ n = λ n - 1 x n - 1 ( 1 - x n ) - λ n x n . (2)
If we let x0(t) ≔ 1 and xn+1(t) ≔ 0, then (2) can be written more succinctly as
x ˙ i = h i - 1 ( x ) - h i ( x ) , i = 1 , & , n , (3)
where hi(x) ≔ λixi(1 − xi+1). This can be explained as follows. The flow of ribosomes from site i to site i + 1 at time t is λixi(1 − xi+1). This flow increases with the density at site i, and decreases as site i + 1 becomes fuller. This corresponds to a “soft” version of a simple exclusion principle. Since the ribosomes have volume, the input rate to site i decreases as the number of ribosomes in that site increases. Note that the maximal possible flow from site i to site i + 1 is the transition rate λi. Thus, Eq (3) simply states that the change in the density at site i at time t is the input rate to site i (from site i − 1) minus the output rate (to site i + 1) at time t.
The ribosome exit rate from site n at time t is equal to the protein translation rate at time t, and is denoted by R(t) ≔ λn xn(t).
Denote by x(t, a) the solution of (3) at time t ≥ 0 for the initial condition x(0) = a. Since the state-variables correspond to normalized occupancy levels, we always assume that a belongs to the closed n-dimensional unit cube C n ≔ { x ∈ R n : x i ∈ [ 0 , 1 ] , i = 1 , & , n } . Let int(Cn) denote the interior of Cn. Ref. [69] showed that the RFM is a tridiagonal cooperative dynamical system [70], and that this implies that (2) admits a unique steady-state point e = e(λ0,…λn) ∈ int(Cn), that is globally asymptotically stable, that is, limt→∞ x(t, a) = e, for all a ∈ Cn (see also [71]). In particular, the translation rate converges to the steady-state value R ≔ λnen. We denote by
ρ ≔ ∑ i = 1 n e i n
the steady-state mean ribosomal density along the mRNA.
The RFM can be extended into a single-input single-output (SISO) control system, by defining the translation rate as the output, and by introducing a time-varying input control u : ℝ + → ℝ + representing the flow of ribosomes from the “outside environment” into the mRNA (which is related to the rate ribosomes diffuse to the 5’end (in eukaryotes) or the RBS (in prokaryotes) of the mRNA). This is referred to as the RFM with input and output (RFMIO) [66]. Thus, the equation for the change in the density at site 1 in the RFMIO becomes
x˙1=λ0u(1−x1)−λ1x1(1−x2),
and all other equations for x ˙ i, i = 2, …, n, are the same as in the RFM. The RFMIO can then be written in a compact-form as
x˙=f(x,u),y=λnxn,
(4)
where y denotes the output.
In this study, each S. cerevisiae gene is modeled by a RFMIO, where each RFMIO site contains 10 consecutive codons (the ribosome footprint is assumed to be about 10 codons wide).
In [43], a network of m RFMIOs interconnected via a pool of free ribosomes (called the RFM network with pool (RFMNP)) was introduced for analyzing large-scale translation while competing for the available, limited ribosomal resource. Competition for the available ribosomal resource leads to indirect coupling between the different mRNAs. For example, if more ribosomes bind to a certain mRNA molecule then the pool of free ribosomes in the cell is depleted, and this may lead to lower initiation rates in the other mRNAs.
Let z(t):ℝ+→ℝ+ denote the free ribosomal pool occupancy at time t. For an RFMNP with m RFMIOs, let nj, i = j, …, m, denote the jth RFMIO dimension, yj ≔ Rj(t) its output rate at time t, and λ0j,…,λnjj its transition rates. The input to the jth RFMIO is uj = Gj(z) where the function Gj(⋅):ℝ+→ℝ+ satisfies: (1) Gj(0) = 0; (2) Gj(z) is strictly increasing on ℝ+; and (3) for all z > 0 sufficiently small Gj(z) is linearly proportional to z. Typical examples are Gj(z) = z, and Gj(z) = aj tanh(z/bj) with aj, bj > 0 (see S1 Fig and [43] for more details). Thus, the RFMNP is given by
x˙1=f(x1,u1),y1=λn11xn11,⋮x˙m=f(xm,um),ym=λnmmxnmm, (5)
and
z˙=∑j=1myj-∑j=1mλ0j(1-x1j)Gj(z). (6)
Eq (6) implies that the change in the free pool, as a function of time, is the sum of all output rates of the RFMIOs (input flow to the free pool) minus the total flow of ribosomes that bind to the mRNA molecules (output flow from the free pool). The RFMNP is then a dynamical system with (1+∑j=1mnj) state-variables. Since the RFMNP is a closed system, the total number of ribosomes H(t)≔z(t)+∑j=1m∑i=1njxij(t) is conserved, that is H(t) ≡ H(0) for all t ≥ 0.
It was proven in [43] that for any given number of total ribosomal pool H(0), the RFMNP admits a unique steady-state point that depends on the rates and H(0) but not on the initial conditions. Furthermore, if one or more of the RFMIOs rates are time-periodic functions, with a common minimal period T > 0, then the RFMNP entrains to the periodic excitations in the λijs, i.e. every state-variable converges to a periodic solution with period T. This also means that each of the translation rates and mean densities converge to periodic solutions with period T. Thus, we do not need to evaluate different values of T, and the value T = 16 is used throughout this paper (e.g. using T = 20 instead yields the same behavior but with periodicity T = 20).
In this paper we simulate the RFMNP while periodically changing the mRNA levels of the heterologous GFP gene or several endogenous genes. Assume, for example, that the GFP mRNA levels are changing (periodically) between a minimal value of β1 > 0 and a maximal value of β2 > β1. It is straightforward to verify that this is equivalent to an RFMNP with β2 copies of the gene GFP, while periodically changing the initiation rates of these copies. Thus, [43] provides a rigorous proof to the periodicity we observed at steady-state in the state-variables. Finally, the parameters of the model used here are based on [72]; see more details below.
We use ribo-seq data to infer the codon decoding rates [73], and normalize these rates so that the median codon elongation rate of all S. cerevisiae mRNAs becomes 6.4 codons per second [48]. This holds for all endogenous genes and the GFP. The ribo-seq data and the decoding rates are used also for inferring the initiation rates. The ribo-seq data and mRNA levels were taken from [74]; the number of S. cerevisiae ribosomes used in the simulation is 200,000 [75], with 60,000 mRNAs [46], scaled according to the mRNA levels from [74]. Thus, the correlations between the predicted ribosome densities from our model and measured ribosome densities in the analyzed conditions are very high (correlation coefficient r > 0.7 for sites size of 10 codons) and is similar to the correlations between two experimental replications in the field [76]. Note that the large-scale measurements of mRNA levels and ribosome profiling suggest that almost all genes have certain mRNA levels and ribosome densities; this suggests that most of the genes are transcribed/translated at the same time but at (possibly extremely) different rates/levels (the differences among genes can be very significant: up to four orders of magnitudes).
Let q ≔ 10 denote the number of codons per RFMIO site. Given an S. cerevisiae gene ORF consisting of K codons (excluding the stop codon), we model it using RFMIO with n sites as follows. The mRNA is divided into (n + 1) pieces: the first piece contains (q − 1) codons (that are also related to later stages of initiation [14]), pieces 2 to n contain each q non-overlapping codons, and the last piece contains between q/2 and 3q/2 codons. For example, for q = 10 and K = 146, the first piece contains 9 codons, pieces 2 to 14 contain each 10 codons, and piece 15 contains 7 codons, thus n = 14. The first piece corresponds to λ0, and pieces 2 to n + 1 correspond to λ1 to λn, respectively, as described next.
The initiation rate (that corresponds to the first piece) is estimated based on the ribosome density per mRNA level, as this value is expected to be approximately proportional to the initiation rate when initiation is the rate limiting factor [45, 77]. We apply a normalization that sets the median initiation rate of all S. cerevisiae mRNAs to 0.8 [47].
The RFMIO rates, per S. cerevisiae gene, are then set as follows:
Let τi denote the decoding time of codon i in the ORF, and let ψ(i) denote the minimum among the decoding times of codon i and its synonymous mutations. Define the decoding-time measure (DTM) of a gene by
η≔∑i=1K(τi-ψ(i))wiK, (7)
where K denotes the number of codons in the ORF (excluding the stop codon), and wi > 0, i = 1, …, K, is the weight given to the non-negative cost (τi − ψ(i)). The DTM then provides a score of how fast the ORF can be decoded; a value of zero means that the ORF is composed from the fastest synonymous codons, and a larger value of η indicates that slower codons are used in the ORF. One might expect that in general η should be inversely proportional to the steady-state translation rate. However, since η doesn’t provide information about the distribution of the decoding time costs along the ORF, this might not always hold. For example, a slow codon in the middle of the ORF can impact the steady-state translation rate more than a slow codon in the boundaries. Another possible interpretation of η is in describing the “speed-budget” relative to the optimum (η = 0 corresponds to the fastest possible decoding times). Thus, two genes with similar DTMs correspond to the same speed-budget.
In the case where w1 = … = wK = 1, the DTM is referred to as the homogeneous DTM. A monotone-increasing weights describes the hypothesis that slower codons toward the 3’ UTR increase ribosomal “traffic jams” on the mRNA, resulting in larger number of ribosomes on the mRNA at steady-state (see S2 Fig).
The GFP protein sequence is from gi:1543069. Recall that the GFP gene ORF consists of 239 codons (excluding the stop codon). The mutated GFP genes are generated by performing the following synonymous substitutions relative to the GFP gene (see also S3 Fig):
|
10.1371/journal.pcbi.1005045 | The Computational Properties of a Simplified Cortical Column Model | The mammalian neocortex has a repetitious, laminar structure and performs functions integral to higher cognitive processes, including sensory perception, memory, and coordinated motor output. What computations does this circuitry subserve that link these unique structural elements to their function? Potjans and Diesmann (2014) parameterized a four-layer, two cell type (i.e. excitatory and inhibitory) model of a cortical column with homogeneous populations and cell type dependent connection probabilities. We implement a version of their model using a displacement integro-partial differential equation (DiPDE) population density model. This approach, exact in the limit of large homogeneous populations, provides a fast numerical method to solve equations describing the full probability density distribution of neuronal membrane potentials. It lends itself to quickly analyzing the mean response properties of population-scale firing rate dynamics. We use this strategy to examine the input-output relationship of the Potjans and Diesmann cortical column model to understand its computational properties. When inputs are constrained to jointly and equally target excitatory and inhibitory neurons, we find a large linear regime where the effect of a multi-layer input signal can be reduced to a linear combination of component signals. One of these, a simple subtractive operation, can act as an error signal passed between hierarchical processing stages.
| What computations do existing biophysically-plausible models of cortex perform on their inputs, and how do these computations relate to theories of cortical processing? We begin with a computational model of cortical tissue and seek to understand its input/output transformations. Our approach limits confirmation bias, and differs from a more constructionist approach of starting with a computational theory and then creating a model that can implement its necessary features. We here choose a population-level modeling technique that does not sacrifice accuracy, as it well-approximates the mean firing-rate of a population of leaky integrate-and-fire neurons. We extend this approach to simulate recurrently coupled neural populations, and characterize the computational properties of the Potjans and Diesmann cortical column model. We find that this model is capable of computing linear operations and naturally generates a subtraction operation implicated in theories of predictive coding. Although our quantitative findings are restricted to this particular model, we demonstrate that these conclusions are not highly sensitive to the model parameterization.
| For more than a century, neuroscientists have worked to refine descriptions of cortical anatomy, either differentiating or consolidating models of cortical circuits [1]. The notion that a fundamental neuronal circuit performs a canonical computation in neocortex, that can be generalized across species and areas, is of fundamental value to both experimental and theoretical neuroscientists. Douglas and Martin provided evidence for such a canonical microcircuit in the cat striate cortex, as well as a descriptive model of its structure [2, 3].
The fundamental building block of circuits on this scale is the cell type specific population. For example, the Douglas and Martin microcircuit model implicated distinct cell types and cortical laminae in its function. Each individual population in the circuit might perform linear or nonlinear transformations on its inputs, depending on the parameterization of the model [4]. The individual cells that make up the population might be spatially segregated (i.e. distinguished by layer) or might be intermingled, and distinguished by genetically defined cell type or projection pattern. The microcircuit can then be conceptualized as a modular collection of populations, with scale and composition dependent on function. Whole brain regions are assembled from ensembles of microcircuits that together perform its overall function, the clearest example being orientation columns in V1. Over time, the microcircuit model can be refined, constrained by including cell type specific parameterizations, synaptic properties, detailed microcircuit anatomy, and other relevant experimentally measured data of a particular cortical area.
When taken together, the cumulative result of multiple recurrently connected canonical circuits might perform the complex nonlinear computations necessary to implement models of higher-order cognitive function. Furthermore, many theoretical models of cortical processing involve hierarchical arrangements of processing stages, and the evidence for such a hierarchical organization, particularly in the visual system, is generally accepted (for example, [5]). Informed by the seminal work of Hubel and Wiesel [6] in the perception of orientation, the catalogue of algorithms for which there exists models relying on a staged, hierarchical implementation has grown significantly. Beyond perception, hierarchical theories include invariant object recognition (for example [7]; see [8] for a review), selective visual attention (see [9] for a review) and models of Bayesian inference via predictive coding (for example [10]).
In order to perform any of these hierarchical computations, individual elements within the hierarchy must perform an intermediate stage of processing. It is hypothesized that these intermediate stages implement a local canonical computation, and their hierarchical arrangement subserves (or even defines) a global information processing stream. In this study, we examine the computational properties of the Potjans and Diesmann [11] cortical column. The main focus of that study was the construction of a realistic computational model of cortex. Here we ask what type of computation this model might subserve as a candidate canonical model of cortical processing. Based on its properties, we then speculate about the role of such a processing unit in an abstract hierarchical computational scheme.
We find that simultaneous excitation to L2/3 and L4 offset in their effects on L5, in essence performing a subtractive computation between two step inputs. Additionally, we find that the model possesses a linear computational regime under the condition that incoming inputs do not preferentially target inhibitory or excitatory populations within a layer. We then examine the response of the model to sinusoidal inputs, again finding evidence of linear computation. In the discussion, we relate these findings to the role of such a processing element in light of theories of hierarchical computation.
The Potjans and Diesmann [11] cortical column model is composed of 8 recurrently connected homogeneous populations of neurons, totaling approximately 80,000 neurons and .3 billion synapses. Each neuron receives background Poisson input, and is recurrently connected to neurons in other populations via a population-specific connection probability matrix derived by combining data and methods from several studies. In our study both network and single neuron parameters from [11] are used.
The population statistic approach used in Iyer et al. [12] assumes synapses that instantaneously perturb the voltage distribution of the postsynaptic population. Therefore, we assume that the fast kinetics of synapses in the Potjans and Diesmann cortical column (τs = .5 ms) can be well-approximated by the DiPDE formalism (For a discussion regarding the effect of non-instantaneous synapses, see [12], “Methods: Non-instantaneous synapses”). As a consequence, the coupling of these shot-noise synapses is instantaneous, perturbing the voltage distribution directly by a constant .175 mV for excitatory synapses, and -.7 mV for inhibitory synapses. These values are computed from the total charge resulting from a single synapse of weight w in [11] using their notation (See [12] for additional details):
Δ v = Q C m = 1 C m ∫ 0 ∞ I ( t ) d t = w C m ∫ 0 ∞ exp ( - t / τ s ) d t . (1)
Connection probabilities, synaptic weight distributions, and delay distributions are taken directly from [11] (with the exception of the L4e → L2/3e connection probability, which was doubled to .088, following [13]). This was done to define equal synaptic strength for all excitatory connections (The original strength for this one connection was doubled relative to other projections), while maintaining roughly the same overall projection strength. The connection probability was multiplied by the size of the presynaptic population to parameterize an effective multiplier (in-degree) on the incoming firing rate from a presynaptic population. Because they minimally impact the firing rate dynamics of the leaky integrate-and-fire model, refractory periods were simplified from 2 ms to zero.
The only significant deviation from the Potjans and Diesmann model was a decrease in the mean background firing rate across all populations by a factor of 8.54, and subsequent increase in the synapse strength of these connections by an equal amount. This modification leaves the mean synaptic input from background unchanged from the original model, but increases the variance of this stochastic input. After this change, the intrinsic oscillations of the original NEST model are significantly damped (but not completely eliminated; see Fig 1). The matched DiPDE model does not exhibit intrinsic oscillations, although in general population density models are capable of exhibiting this phenomenon [14].
Each population is initialized to a normal distribution of membrane voltages, with a mean at the reset potential and standard deviation of 5 mV. Before application of any additional input (i.e., step or sinusoidal drive), background excitation is applied to each population as specified in [11], and simulated for 100 ms to reach a pre-stimulus steady state. When driving the model, additional layer-specific excitatory stimulus is input into the target layers(s), and simulated for an additional 100 ms. For step inputs, the difference of the final steady-state less the pre-stimulus steady-state (i.e. after discarding the initial start-up transient dynamics) define the layer-specific firing rate output perturbation.
In this study, all simulations of the model were performed using a numerical simulation of the displacement partial integro-differential equation (DiPDE) modeling scheme proposed in [12] with a time-step of.1 ms. At this temporal resolution, a 200 ms DiPDE simulation requires 31 seconds running on a 2.80 GHz Intel Xeon CPU. The corresponding NEST simulations [15, 16] included in Fig 1(c) and 1(d) require 402 seconds each (single processor), and results from 100 of these simulations are averaged to obtain the mean firing rate pictured. In each of these 100 averaged NEST simulations, connectivity matrices and initial values for voltages were randomized. The population density approach in computational neuroscience seeks to understand the statistical evolution of a large population of homogeneous neurons. Beginning with the work of Knight and Sirovich [17] (See also [18, 19]), the approach typically formulates a partial integro-differential equation for the evolution of the voltage probability distribution receiving synaptic activity, and under the influence of neural dynamics. Neuronal dynamics typically follow from the assumption of a leaky integrate-and-fire model. We implement a numerical scheme for computing the time evolution of the master equation for populations of leaky integrate-and-fire neurons with shot-noise current-based synapses (For a similar approach, see [20]).
τ m d v d t = - v + Δ v ∑ i δ ( t - t i ) v > v t h ⇒ v → v r (2)
Here τm is the membrane time constant, v is the membrane voltage, Δv is the synaptic weight, vth is the threshold potential, and vr is the reset potential (here taken to be zero for simplicity). Extending [12], each population receives input from both background Poisson input and recurrent connections from each cortical subpopulation. We emphasize that this is not a stochastic simulation; for example, the background Poisson drive is not a realization of a Poisson process, but rather the effect of a Poisson-like jump process on the evolution master equation. At each time step, a density distribution representing the probability distribution of membrane voltages for each population is updated according the differential form of the continuity equation for probability mass flux J(t, v) (Here p(t, v) is the probability distribution across v at time t on (−∞, vth); see [21] for more information):
∂ p ∂ t = - ∂ J ∂ v (3)
The voltage distribution is modeled as a discrete set of finite domains (See Fig 2). Synaptic activation of input connections drive the flux of probability mass between nodes, while obeying the principle of conservation of probability mass. As a result, a numerical finite volume method is an ideal candidate for computing the time evolution of the voltage density distribution, and we numerically solve Eq 3 with a finite volume method.
The spatial (voltage) domain
D = [ v m i n , v θ ] ⊂ R (4)
is subdivided into a set of non-overlapping subdomains
V = { v i ⊂ D } . (5)
Each subdomain contains a control node pi that tracks the inflow and outflow of probability mass due to synaptic activation and passive leak. At each time step, pi is updated by considering probability mass flow resulting from synaptic activation from all presynaptic inputs as well as leak; for simplicity we will describe the update rule assuming a single presynaptic input. Additionally, we will assume a single synaptic weight, although in general this approach works equally well for a distribution of synaptic weights. Under these assumptions, the discretized version of Eq 3 can be formulated as:
d p i d t = - Δ J i Δ v i (6) Δ J i = f i + 1 2 - f i - 1 2 (7) = ( j ( s , i ) - - j ( l , i ) + ) - ( j ( s , i ) + - j ( l , i ) - ) . (8)
Here f i ± 1 2 denotes flux across the right or left subdomain boundary, js denotes flux resulting from the input population (via synaptic activation), and jl denotes flux from the leak; the superscript is a convenience that denotes the overall sign (i.e. inflow or outflow) of the contribution of the term to pi.
Synaptic activation contributes j(s) to the overall flux by displacing probability mass (pΔv) with a transition rate λin, the presynaptic firing rate. By directly computing the probability mass flux as Δt → 0 over the subdomain boundary (while enforcing probability mass conservation), the contribution of passive leak j(l) to the overall flux can be formulated as a transition rate that increases exponentially with time constant τm as the voltage of the subdomain boundary being crossed increases. To summarize, the flux contributions to the ith subdomain are:
j ( s , i ) + = p k Δ v k λ i n (9) j ( s , i ) - = p i Δ v i λ i n (10) j ( l , i ) + = p i + 1 v i + 1 2 τ m (11) j ( l , i ) - = p i v i - 1 2 τ m (12)
Here the synaptic influx j ( s , i ) + depends on pk, the probability mass in subdomain vk located a distance w = vi − vk (the synaptic weight) from vi. In the special case of i = 0 and w > 0, the node that acts as the reset value for probability mass that exceeds the spiking threshold vθ (i.e. the boundary condition) receives probability mass from all nodes less than w from vθ. Because these updates result from a linear update from probabilities, the entire time evolution can be formally represented as:
d p d t = ( L + S ) p (13)
where leak and synaptic input contributions have been separated into two separate discrete flux operator matrices. At this step, it is trivial to include additional synaptic inputs S0, S1, …Sm, yielding a formal solution over a single time step Δt:
p ( t + Δ t ) = exp Δ t L + ∑ s = 0 m S s p ( t ) (14)
for some initial probability distribution p(t). At each time step, the synaptic input matrices Sk are updated to reflect the changes in firing rate of the presynaptic populations (if necessary).
Probability mass that is absorbed at threshold and inserted at the reset potential defines the fraction of the population that spiked; after normalization by the discrete time step Δt, this defines the output firing rate. The output firing rate provides the rate of a Poisson process that drives any recurrently-connected postsynaptic populations. Probability mass flux through the boundary vθ into the subdomain at i = 0 defines the instantaneous firing rate of the population, computed as:
λ o u t ( t ) = ∑ s = 0 m j ( s , 0 ) + Δ t (15)
Recurrent coupling between simulated populations is accomplished by assigning λout of the presynaptic population to λin of the postsynaptic population.
The source code for DiPDE is released as an open source python package under the GNU General Public License, Version 3 (GPLv3), and is available for download at http://alleninstitute.github.io/dipde/. The package includes an example implementation of the cortical column model analyzed in the main text, absent any inputs in excess of background excitation.
In this section we describe the repertoire of computations caused by step inputs over and beyond background excitation (See Fig 1(c) and 1(d) for an example simulation, compared to 100 averaged leaky integrate-and-fire (LIF) simulations) into a coarse-grained population-statistical version of the Potjans and Diesmann cortical column model (See Fig 1(a) for a visual summary of projections in the column model; for a complete model description see [11], Tables 4 and 5). The targeting of cell types (i.e. target specificity) has important consequences for the responses caused by incoming inputs. We examine the consequences of three types of target specificity, summarized in Fig 1(b) for incoming excitatory projections into a given layer within the column. The excitatory and inhibitory target specificity regimes excite their respective cell types, while the balanced regime does not preferentially target either subpopulation. Unless otherwise specified, the step input has a firing rate of 20 Hz, and models a convergent connection with 100 independent presynaptic sources per target neuron.
Fig 3 provides an overview of output perturbations evoked by step input into a given layer, under each target specificity condition. In effect, this provides an at-a-glance summary of the catalogue of computations that the cortical column can perform, given a 20 Hz step pulse excitatory input into a single layer.
We find that the effect of driving L2/3 under any specificity condition has a depressing effect on the activity in L5. In contrast, when driving L4 or L5, activity across almost every population in the network increases or decreases when driving the excitatory or inhibitory subpopulation, respectively. Fig 3 also demonstrates that under balanced target specificity (yellow), the effects of inputs into L2/3 and L4 are nearly equal-and-opposite with respect to the output of L5. We summarize this comparison across all output layers in Fig 4 which additionally plots the combined effect of inputs simultaneously into L4 and L2/3. This plot demonstrates that these two inputs approximately offset; we explore this observation further in the next section.
We note that the input layers involved in this subtraction are implicated in bottom-up vs. top-down comparisons in the theory of hierarchical predictive coding [22]. Also conspicuous is the output population reporting this subtraction; L5 pyramidal neurons provide the dominant cortical output, including the pons, striatum, superior colliculus, and to value encoding dopaminergic neurons in the VTA or SNc [23] where subtraction errors might skew reward expectations (see Discussion for further details).
Linear computations are characterized by simultaneously exhibiting homogeneity (i.e. multiplicative scaling in the sense of a linear map) and additivity with respect to inputs. The previous section examined output perturbations across layers and target specificity profiles of a single strength (20 Hz firing rate). In this section, we first examine the effect of linearly increasing the strength of the input, testing the homogeneity of the system. Fig 5 extends Fig 3 by providing a summary across an increasing range of input strengths.
Under balanced target specificity (middle column of panels), the magnitude of each population response exhibits a scaling behavior, linear in the input magnitude. In contrast, when neurons are targeted with a cell type specific bias, the response of certain subpopulations can be nonlinear. The clearest example of the nonlinear influence of the inhibitory subpopulation occurs when driving layer 5. Through both an increase in direct self-inhibition, and indirect reduction of self-excitation via the L5e subpopulation, excitatory drive into L5i can paradoxically decrease activity, an effect described previously in inhibition-stabilized recurrent networks [24]. Eventually this effect reverses when the L5e activity is completely inhibited.
Fig 6 demonstrates that, likewise, additivity is violated (somewhat, as the points deviate from the identity line) when preferentially targeting inhibitory neurons. Each point in the figure depicts the result of driving two separate layers with a 20 Hz firing rate input, and considering the perturbation in firing rate of each subpopulation (specified in the legend). For a given target specificity condition, two independent simulations are run, for each of the two input layers; the sum of the perturbation they evoke is plotted on the vertical axis. The output resulting from a single simulation with two equal inputs into each input layer, is plotted on the horizontal axis. When a point lies along the identity line, this implies additivity.
This figure implies a conclusion similar to the homogeneity study above: as the target specificity moves from excitatory to inhibitory, the firing rate computation performed on laminar inputs by the cortical column changes from linear to weakly nonlinear. In the previous section, we demonstrated that balanced 20 Hz firing rate inputs to L2/3 and L4 approximately offset each other in the output evoked in L5. The homogeneity and additivity demonstrated above indicate that L5 will actually reflect a subtraction operation on these two inputs. We return to this point in the discussion.
Given the amount of recurrent connectivity in the model, its linear response to step inputs under balanced target specificity might seem surprising. However, it is known that balanced networks can exhibit linear responses to external inputs (See, for example, [25]). Although the model parameterization is taken from the literature, we also investigated the sensitivity of this linear response to perturbations in model parameters. By perturbing the connection probability matrix (Table 5 “Connectivity” in [11]), we defined 1000 alternative models. Specifically, each entry in the matrix was multiplied by a normally distributed random number with unit mean, and standard deviation taken as 5% of the entry (negative values were thresholded to zero).
The homogeneity of response to each new model was assessed by linearly extrapolating the perturbation resulting from a 10 Hz firing rate step input from the results obtained from a 5 Hz step input. The absolute value of the prediction error:
Δ F = ( F 10 - F 0 ) - 2 · ( F 5 - F 0 ) (16)
quantifies the difference between the extrapolated value, and the true value obtained by direct simulation of a 10 Hz firing rate input. Here F indicates the firing rate after reaching steady-state, and the subscript indicates the strength of the step input. Intuitively, this quantity will be zero when a linear extrapolation can predict the data (i.e. a linear relationship between inputs and outputs). Nonzero values indicate the failure of a linear extrapolation, and thus a nonlinear dependence of the output firing rate on the input over the regime of 0–10 Hz perturbations.
S1 Fig shows a stacked histogram of this prediction error for the 1000 perturbed models, across all combinations of target specificity, laminar drive, and output population. Under balanced target specificity (middle column), the prediction error is reliably smaller, particularly when layers 4 and 5 are targeted (middle two rows). This implies that the linear relationship between inputs and outputs under balanced input of the original cortical column model is insensitive to small perturbations in the connection probability matrix.
A similar result holds for additivity predictions, shown in S2 Fig. For the same perturbed models, the additive prediction error is defined as the sum of output responses in two layers from two different simulations, minus the output resulting from driving the two layers in the same simulation. Again, the model under balanced target specificity is less sensitive to perturbations than when cell types are selectively driven. Therefore, we conclude that the observation of linear responses in model output in the previous section is not a result of fine tuned parameters.
In the previous section, we demonstrated that target specificity can determine the linearity of the model response under step inputs. To further investigate the linearity of the transformation that the column applies on its inputs, we next consider sinusoidal drive above and beyond background drive (See Fig 1(d) for an example simulation, compared to 100 averaged LIF simulations). S3 Fig summarizes the nonlinear distortion in each populations response under a 5 Hz peak amplitude sinusoidal drive. Only responses with a peak amplitude greater than .05 Hz are plotted. Total harmonic distortion (THD) compares the power present in the harmonics of the driving frequency in the sinusoidal input signal that perturbs a subpopulation above and beyond the background firing rate:
T H D ( f ) = ∑ i = 2 ∞ V i 2 V 1 (17)
Here Vi is the power spectral density (PSD, [26]) of the ith harmonic of the principal (driving) frequency. This figure reinforces the conclusion from the previous section, that the target specificity of the sinusoidal drive can affect the nonlinearity of transformations resulting from population-level processing. In particular, balanced drive minimizes the harmonic distortion imposed by the dynamics within the model. In contrast, inhibitory drive into layer 5 produces nonlinear responses throughout the column, in agreement with observations about the homogeneity of responses to step inputs (cf. Fig 7)
The low THD of the output signals from balanced drive indicate that the firing rate y(t) of a population in the column model can be approximately modeled as a linear filter on the input signal x(t) plus a baseline x0:
y ( t ) = x 0 + ∫ 0 ∞ x ( t - τ ) h ( τ ) d τ (18)
Fig 8 provides a numerically computed description of three examples of this linear filtering, resulting from balanced drive from L2/3→L5e, L4→L5e, and L4→L23e. Of all possible input/output pairs, these examples show the least signal attenuation from the amplitude Ain of the input signal x(t) to the amplitude Aout of the output signal y(t) (i.e. the largest impact on changes to subpopulation firing rate). Clearly evident in the first two figures are first-order lowpass filters, similar to feedforward systems found in [4] with significantly higher synaptic weights (relative to threshold). These filters both have a cutoff frequency near 15 Hz, implying a corresponding RC time constant near the membrane time constant (10 ms) of neurons in the system. Interestingly, this observation is in agreement with the very general prediction of predictive coding theories, that high frequencies should be attenuated when passing from superficial to deep layers [22] (See Discussion). Transmission from L4 to L2/3 is band-passed near the 10–30 Hz range.
In this study, we examine what input/output transformations a popular model of a cortical column performs on layer-specific excitatory inputs. Transformations are defined as perturbations to the steady-state mean firing rate activity of subpopulations of cells in response to step and sinusoidal inputs in excess of background drive. Because the mean firing rate is a population-level quantity, we use a population statistic modeling approach, by numerically computing the population voltage density using DiPDE (http://alleninstitute.github.io/dipde/), a coupled population density equation simulator (See Numerical Methods). This approach enables a fast, deterministic exploration of the stimulus space and model parameterization. Our approach begins with a data-driven model as a starting point, and then examines the computations its dynamics subserve, as opposed to fitting a model to a preselected set of dynamical interactions resulting from assumptions about cortical computation. Our goal is to discover robust evidence for theories of cortical function using knowledge about structure (synthesized by Potjans and Diesmann [11] into their cortical column model), while limiting biases and a priori functional assumptions.
We consider three discrete regimes of input specificity: excitatory preference, no preference (i.e. balanced, in which both excitatory and inhibitory cells in any one layer receive the same external input), or inhibitory preference. We find that balanced target specificity results in output perturbations that scale linearly with input strength and combine linearly across input layers. In contrast, selective targeting of a particular cell type (especially the inhibitory subpopulation) leads to nonlinear interactions. Additionally, we find that equal, simultaneous, and balanced inputs into L2/3 and L4 are offset in their effect on the L5 firing rate; combining this with the observation of linearity implies that perturbations in L5 activity represent a subtraction from L4 activity of L2/3 activity. The inhibitory effect of L2/3 input on L5e output appears to be largely mediated by L2/3 interneurons inhibiting L5 pyramids (c.f. [27], their Fig 4) while the excitatory effect of L4 on L5 is a network effect resulting from multiple projection pathways. We conclude that the cortical column model implements a subtractive mechanism that compares two input streams and expresses any differences in the mean activity of L5. While this computation can be implemented via other inputs, this combination is interesting because no target cell type specificity is required.
How does this observation of a mechanism for subtraction relate to existing theories of cortical processing? Predictive coding [10, 28] postulates a computation that compares an internal model of the external environment to incoming sensory signals, in order to infer their probable causes [29, 30]. The subtractive dynamics supported by the cortical column model could accomplish this. However, this would imply that sensory signals are represented dynamically in one layer, an environmental model in the supragranular layer, and that their functionally relevant difference is relayed by the infragranular layer (layer 5). The internal granular layer (layer 4) is the obvious candidate for incoming environmental evidence, given its specialized role in receiving input from the primary sensory thalamus. Similarly, the role of the infragranular layer in driving subcortical structures involved in action (basal ganglia, colliculus, ventral spinal cord) seem compatible with the proposition of layer 5 representing the output of a comparison operation. Although more speculative, this leaves the supragranular layer responsible for generating the internal environmental model, which seems reasonable given its abundance of intracortical projections and increased development in higher mammals.
These speculative roles of the various cortical layers conform to abstract models of canonical microcircuits (See, for example, [31]). This is especially true when placed in a hierarchy of processing stages, for example in hierarchical predictive coding (hPC) [22]. In this framework, sequential processing stages generate top-down predictions, and pass bottom-up prediction errors, at each level in the hierarchy. In primates, the laminar segregation of these streams is easily aligned with the anatomical characterization from Felleman and Van Essen [5], with feedforward connections targeting L4, and feedback connections avoiding L4. In rodents, the relation between lamination and hierarchy is less clear [32]. Although the central theme of distinct populations of forward-projecting neurons targeting L4 vs. backward projecting neurons avoiding L4 in the visual system seems conserved [33], these distinct populations are not segregated by layer, but instead intermingled [34]. Therefore, future experimental attempts to establish connections between hierarchically defined visual processing regions and theoretical models may require projection-target-segregated (or perhaps genetically-segregated, if projection markers can be established), as opposed to laminae-segregated, cellular subpopulations.
An additional connection between the hPC model and the results of the simulations in this study is presented in Fig 8. Here the response of the deep layer of the model to stimulation in either of the two superficial layers is well characterized by a linear low-pass filter. Interestingly, this filtering is a prediction of hPC, where high frequencies should be attenuated when passing from superficial to deep pyramidal cells [22]. The low-pass filtering prediction arises from the hypothesis that cortex is performing a form of Bayesian filtering, by attempting to update an estimated quantity using noisy measurements. These noisy estimates by their nature have higher-frequency content than the uncorrupted “true” quantity being estimated, and so the appearance of a smoothing transform is not surprising. However, it is surprising that our model, formulated without assuming any underlying computation (especially not Bayesian filtering), performs this smoothing at a dynamical stage precisely where the anatomically-informed hPC model requires it. Taken together, the convergence of experimental, anatomical, theoretical, and simulation evidence is striking.
As mentioned above, neurons in the infragranular layer project both cortically and subcortically. Based on the hPC model, and because of the model’s ability to compute a subtraction between inputs to the granular and supragranular layers, we have speculated this computation could represent an error signal between reality and expectation. What would this conclusion imply for the subcortical projections? Watabe-Uchida et al. [23] found that dopaminergic neurons in the ventral tegmental area and substantia nigra pars compacta receive sparse input from the deep layers of cortex (for example their Fig 5). Because of the well established role of these midbrain structures in valuation, motivation, and reinforcement learning, these authors suggest that these dopaminergic neurons might “calculate the difference between the expected and actual reward (i.e., reward prediction errors).” While speculative, it is possible that this prediction is calculated cortically and relayed either directly or indirectly [35].
There are a number of concrete steps that can be taken to strengthen the relationships between model, theory, and experiment. A more comprehensive model parameterized by experimental data with additional layers and cell types could be combined with matched optogenetic in vivo and in silico perturbation experiments. These manipulations could validate model predictions, suggest refinements, and test specific conclusions related to theories of population-based cortical processing, for example the functional role of different classes of genetically defined interneuron populations. Such models might also suggest a reinterpretation of how different cell populations contribute to the computation of error signals, or suggest new canonical computations carried out by population-level activities. Either way, in our view, the population density modeling approach will continue to provide a valuable tool for quickly exploring the dynamical consequences of population level computational models.
|
10.1371/journal.pgen.1003732 | Meiotic Recombination Initiation in and around Retrotransposable Elements in Saccharomyces cerevisiae | Meiotic recombination is initiated by large numbers of developmentally programmed DNA double-strand breaks (DSBs), ranging from dozens to hundreds per cell depending on the organism. DSBs formed in single-copy sequences provoke recombination between allelic positions on homologous chromosomes, but DSBs can also form in and near repetitive elements such as retrotransposons. When they do, they create a risk for deleterious genome rearrangements in the germ line via recombination between non-allelic repeats. A prior study in budding yeast demonstrated that insertion of a Ty retrotransposon into a DSB hotspot can suppress meiotic break formation, but properties of Ty elements in their most common physiological contexts have not been addressed. Here we compile a comprehensive, high resolution map of all Ty elements in the rapidly and efficiently sporulating S. cerevisiae strain SK1 and examine DSB formation in and near these endogenous retrotransposable elements. SK1 has 30 Tys, all but one distinct from the 50 Tys in S288C, the source strain for the yeast reference genome. From whole-genome DSB maps and direct molecular assays, we find that DSB levels and chromatin structure within and near Tys vary widely between different elements and that local DSB suppression is not a universal feature of Ty presence. Surprisingly, deletion of two Ty elements weakened adjacent DSB hotspots, revealing that at least some Ty insertions promote rather than suppress nearby DSB formation. Given high strain-to-strain variability in Ty location and the high aggregate burden of Ty-proximal DSBs, we propose that meiotic recombination is an important component of host-Ty interactions and that Tys play critical roles in genome instability and evolution in both inbred and outcrossed sexual cycles.
| Meiosis is the cell division that generates gametes for sexual reproduction. During meiosis, homologous recombination occurs frequently, initiated by DNA double-strand breaks (DSBs) made by Spo11. Meiotic recombination usually occurs between sequences at allelic positions on homologous chromosomes, but a DSB within a repetitive element (e.g., a retrotransposon) can provoke recombination between non-allelic sequences instead. This can create genomic havoc in the form of gross chromosomal rearrangements, which underlie many recurrent human mutations. It has been thought that cells minimize this risk by disfavoring DSB formation in repetitive elements, partly based on studies showing that presence of a Ty element (a yeast retrotransposon) can suppress nearby DSB activity. Whether this is a general feature of Tys has not been evaluated, however. Here, we generated a comprehensive map of Tys in the rapidly sporulating SK1 strain and examined DSB formation in and around all of these endogenous Ty elements. Remarkably, most natural Ty elements do not appear to suppress DSB formation nearby, and at least some of them increase local DSBs. These findings have implications for understanding the relationship between host and transposon, and for understanding the impact of retrotransposons on genome stability and evolution during sexual reproduction.
| Meiosis is the specialized cell division that halves the genome complement to produce gametes for sexual reproduction. During meiosis, homologous recombination is induced by programmed DNA double-strand breaks (DSBs) made by the topoisomerase-like Spo11 protein in a reaction in which Spo11 attaches covalently to 5′ strand termini of the DSB [1]. Endonuclease cleavage releases Spo11 from the DSB ends in a covalent complex with a short oligonucleotide [2]. Subsequent resection of DSB 5′ strand ends generates 3′ single-stranded tails, which are substrates for proteins that search for a homologous DNA duplex and effect the templated repair of the break [3].
DSBs in single-copy sequences usually induce recombination between allelic segments on homologous chromosomes, which promotes pairing and accurate segregation of homologs and increases genetic diversity in gametes. However, eukaryotic genomes are replete with repetitive elements that share high sequence identity. A DSB formed in a repeat can induce recombination between non-allelic DNA segments, which can in turn result in chromosome rearrangements such as duplications, deletions, inversions or translocations [4]–[6]. In humans, such non-allelic homologous recombination (NAHR, also referred to as ectopic recombination) in the germ line contributes to non-pathogenic structural variation [7] and is linked to numerous genomic disorders [5]. NAHR is thus a driving force in genome evolution and a source of genome instability. Meiotic DSBs are distributed non-randomly across genomes [8], [9], so the propensity toward NAHR depends strongly on how likely it is that Spo11 cuts in and around repetitive elements [6].
A major class of repetitive element in S. cerevisiae comprises the Ty elements, ∼6-kb retrotransposons related to mammalian retroviruses [10]. Each contains an internal region encoding Gag- and Pol-like proteins required for retrotransposition, flanked by ∼330-bp long terminal repeats (LTRs). The S288C strain (source of the yeast reference genome) contains 50 Ty elements in five distinct families: 31 Ty1, 13 Ty2, 2 Ty3, 3 Ty4 and 1 Ty5 [11]. S288C also contains a much larger number of solo LTRs or LTR fragments, which likely arise from homologous recombination between the LTRs of full-length Tys [11]. The predominant families, Ty1 and Ty2, exhibit high sequence identity: >90% in pairwise comparisons within families and >70% between Ty1 and Ty2 [11], [12]. Because of their sequence similarity and dispersed distribution, Ty elements are potent sources of gross chromosomal rearrangements. Numerous studies have documented Ty-mediated NAHR induced by DSBs or replication errors in vegetatively growing cells [e.g.], [ 12]–[15].
Comparatively little is known about Ty recombination provoked by Spo11-generated DSBs in meiosis. A URA3-marked Ty2 element inserted in the HIS4 promoter caused a >13-fold reduction in DSBs at this site, which is normally a strong DSB hotspot [16]. An open (nucleosome-depleted) chromatin structure is an important determinant of Spo11 hotspots [17]–[19]. The HIS4 promoter, like most yeast promoters, displays hypersensitivity to DNase I digestion of chromatin, but the inserted Ty (which is itself resistant to nuclease digestion), converted the local chromatin structure to a nuclease-resistant state [16]. Thus, a Ty can suppress DSB formation nearby, possibly via spreading of a closed chromatin structure into the surrounding region [16]. However, although this element mimicked a spontaneous Ty insertion [20], it is in an unusual position since Tys most often integrate near tRNA genes and only rarely into RNA pol II promoters [11], where most DSB hotspots occur in yeast [19]. By direct restriction mapping and Southern blotting on chromosome III in the rapidly and efficiently sporulating S. cerevisiae SK1 strain, two novel Tys were identified [21]. DSBs were not detected within or adjacent to these Tys, but it remained unknown whether DSBs are infrequent in or near other natural Tys. Also, Ty elements can differ widely from one another in many of their behaviors. For example, the few Tys examined to date undergo NAHR at dissimilar frequencies, ranging from ∼10−5 to ∼10−2 per meiosis [4], [22], [23], and expression of Tys and Ty-adjacent genes varies substantially between individual elements [10], [24]. Thus, it is presently unknown whether local DSB suppression can be extrapolated to be a general feature of natural Ty elements.
Deep sequencing of the Spo11 oligos that are byproducts of DSB formation provided a high resolution DSB map and suggested that DSBs are moderately suppressed within Tys on average [19]. However, average behavior does not reveal the extent of variation between different sites. Here, we examine DSB formation in and around endogenous Ty elements in SK1 and explore how the presence of natural Ty elements affects local DSB frequency.
Full-length Ty1 and Ty2 elements were previously mapped in SK1 by microarray hybridization of genomic DNA containing Ty sequences [25]. Twenty-five Ty-containing regions were identified, but spatial resolution (ranging 1–21 kb) was not high enough for us to assess nearby DSBs. The Saccharomyces Genome Resequencing Project (SGRP) generated an SK1 genome assembly from shotgun sequencing combined with phylogenetic comparisons [26]. The number of Ty-containing reads led to an estimate that retrotransposons are ∼2% of the SK1 genome vs. >3% for S288C, implying SK1 has ∼30 Tys. However, the SGRP assembly and its subsequent refinement [27] did not compile all Ty elements, identify Ty families, or reveal precise Ty positions, because repetitive elements pose a computational challenge in genome assembly [26], [28], [29]. Moreover, the SK1 strain sequenced by SGRP is a homothallic, prototrophic strain (HO, LYS2, URA3, LEU2) related to an ancestor of the strains most widely used in meiosis research [30].
To overcome these limitations, we took a multipronged approach to precisely map all full-length Ty elements in the Kleckner laboratory-derived SK1 lineage. We identified a total of 30 Tys and fine-mapped their positions, in most cases to single-nucleotide resolution (Figure 1 and Table 1).
First, we asked whether SK1 has Tys that are present in S288C. The SGRP data consist of paired-end sequence reads with an average insert of 4–5 kb. These reads can reveal structural differences between a reference genome and the DNA source if the distance between mapped pairs is substantially larger or shorter than the average, or if orphans are present, where one read maps but its mate fails to map or maps to a different genomic region and/or multiple locations (Figure 2A). From inspection of SGRP read maps, we found only one S288C element that was also present in SK1: YCLWTy5-1 at the left end of Chr III (Figure 1, Figure 2B and Table 1). This is the only full-length Ty5 family member in either strain, although there are Ty5 solo LTRs and LTR fragments in both (data not shown). Ty5-family insertions are found preferentially near telomeres and silent mating type loci [reviewed in 11]. YCLWTy5-1 contains mutations rendering it nonfunctional for transposition [31], so this is an ancient Ty present in the last common ancestor of these strains. The remaining 49 S288C Tys are not present in SK1 (Figures S1A, S1B and data not shown). This manual inspection also identified eight S288C Ty sites for which SK1 has one or more Ty elements nearby, subsequently confirmed by PCR (11 Tys total; Figure S1B, Table 1 and data not shown). These novel Tys are at different positions in SK1 than in S288C and often of a different family or in opposite orientation, thus are independent integration events.
Second, we evaluated Ty1 and Ty2 sites mapped by Gabriel et al. in the Kleckner lineage [25]. Using SGRP sequence patterns plus PCR and sequencing of genomic DNA, we validated and fine-mapped 22 SK1-specific elements (Figure S1C and data not shown), but 3 sites showed no evidence of a Ty in SGRP data. Two of these reflect differences between SGRP and Kleckner SK1 strains: a spontaneous ura3 mutation selected during derivation of the Kleckner strains and caused by Ty integration [30], [32] (Figure 2C); and a Ty1 on Chr XIII (Figure 2D and data not shown). The latter is likely a de novo, unselected integration that passed through the bottlenecks of strain derivation, demonstrating the potential for occult differences between otherwise isogenic strains. The third discrepancy is a single Ty incorrectly assigned to two separate sites on Chr IV. S288C contains a tandem duplication of similar genes encoding a hexose transporter (HXT6 and HXT7) [33], but SK1 has only one HXT copy in this region and lacks the intervening sequence (Figure 2E and data not shown). This structural difference caused the microarray hybridization data to artifactually give two peaks from a single Ty when projected onto S288C sequence space.
Third, we used an unbiased approach to ensure that all Ty elements were identified, using SGRP data and a paired-end genomic sequence library from NKY291, a Kleckner-lineage haploid. We retrieved sequence pairs in which one mate matched non-LTR parts of Tys, then mapped the non-Ty mate on the S288C genome (277 SGRP reads and 4,963 NYK291 reads, <1% of the total from each). Tys appear as clusters of reads pointing from both directions at the insertion site (Figures 3A and 3B). We identified all of the Tys described above, and also found an additional element on Chr II, present in both libraries and confirmed by PCR (data not shown). On average, 8.6 SGRP reads tagged each SK1-specific Ty or cluster of Tys, and the read counts matched a Poisson distribution (Figure 3C). Thus, we estimate the probability to be <0.0002 that a Ty was missed because of chance failure to recover supporting reads. The NKY291 library provided even more reads identifying each Ty (mean 158.2, range 30–304), so it is highly likely we identified all of the Tys in SK1.
SK1 Tys showed both conserved and non-conserved features with their S288C counterparts. Ty1 and Ty2 are the predominant families, as in S288C, and only one each of Ty3 and Ty5 are present (Figure 3D). Of the Ty1 or Ty2 elements that could be typed by established criteria [11] or mapped by a prior study [25], 21 are Ty1 and 5 are Ty2 (Figure S1D and Table 1; two could not be typed with available data). Since we did not determine the entire sequence of the Ty elements, it is unknown which are capable of autonomous transposition. No Ty4 element was found and none of the SGRP or NKY291 reads matched Ty4 internal sequences, but Ty4-derived solo LTRs are present (data not shown). Thus the Ty4 family is extinct in SK1.
Most LTR-retrotransposons generate sequence duplication at the integration site [34]. Among SK1 Ty elements whose insertion sites were precisely mapped, >95% showed perfect target sequence duplication with a good match to the consensus for elements in S288C (Table 1 and Figure S1E).
Ty integration can be potentially deleterious, by inactivating or altering expression of neighboring genes [35], [36]. However, obviously deleterious insertions are relatively rare in S288C [11]. Selection may account for some of this pattern, but target site bias is also a major factor: ∼90% of Ty1–Ty4 insertion sites in S288C (including solo LTRs) are near RNA pol III-transcribed genes such as tRNAs [11], mediated by interaction of Ty integrase with factors required for RNA pol III transcription [35]. Similarly, most SK1 Ty1, Ty2, and Ty3 elements (26 of 29) are near tRNA genes (Table 1), and one of the exceptions (at ura3) was selected because it conferred a desirable phenotype.
We previously showed that Spo11 oligo counts covary linearly with DSB levels, so the frequency of mapped Spo11 oligos is a proxy for DSB frequency [19]. To assess global trends for DSB formation near Ty elements, we compiled densities of Spo11 oligos within 0.5, 1, and 2 kb windows on both sides of each SK1 Ty (Figure 4A and Table S1). These densities varied widely between different Ty insertion sites, covering 80 to 500-fold ranges, depending on window size. Many Ty-flanking regions differed substantially from genome average, both hotter and colder. There was no obvious distinction between Ty families, in that the five elements unambiguously identified as Ty2 showed 33-fold variation in local Spo11 oligo density, and overlapped extensively with densities for Ty1 elements (p = 0.25, Wilcoxon rank sum test) (Table S1).
The mean Spo11 oligo density near Ty elements was higher than genome average, irrespective of window size (Figure 4A). However, since Tys are not randomly positioned, genome average may not be the most informative comparison. Although SK1 does not have full-length Ty elements where most of the Tys in S288C are found, integration bias with respect to tRNA genes was similar in the two strains. We reasoned that S288C integration sites can be viewed as potential integration sites in SK1, i.e., that S288C sites provide a good negative control for correlations between DSBs and Ty presence. In three window sizes analyzed, Spo11 oligo densities around these control sites varied as widely as for bona fide Ty integration sites (Figure 4A). However, while the density ranges overlapped, the values were consistently higher around SK1 Ty elements than around control sites, with mean Spo11 oligo densities 2.3–2.7-fold higher around the SK1 Tys (Figure 4A). We conclude that natural Ty insertion sites display a great degree of individual variability with respect to local Spo11 activity, comparable to the variability that would be seen for similar genomic locations without a Ty present. Moreover, these data do not provide evidence that Ty presence invariably causes DSB suppression nearby, and instead raise the possibility that Tys may tend to increase the local likelihood of DSB formation.
DSBs are preferentially formed at RNA pol II promoters [8], [37]. Intergenic regions between divergent transcription units, i.e., containing two promoters, tend to be somewhat hotter on average than intergenic regions between tandemly oriented genes, i.e., with just one promoter, while intergenic regions between convergent transcription units tend to be much colder than either [19]. All SK1 Ty elements, except the one in ura3, are in intergenic regions. When Ty elements were divided according to type of intergenic region, the local Spo11 oligo densities mirrored the trends seen for all intergenic regions genome-wide: Tys in divergent regions tended to have more Spo11 oligos mapped nearby than Tys in tandem regions, and both tended to be hotter than Tys in convergent regions (p = 0.0337, one-way ANOVA; Figure 4B). These findings imply that Ty elements do not necessarily override the intrinsic DSB-forming potential of the intergenic regions where they reside.
Spo11 oligo patterns were confirmed by direct detection of DSBs near a subset of Ty elements. Since meiotic DSBs are transient in wild type, DSBs were detected in repair-deficient mutants. Sae2 is required for removal of Spo11 from DSB ends, so sae2 mutants accumulate unresected DSBs that can be precisely mapped [2], [38]–[40]. However, these DSBs can differ quantitatively from wild type in a region-specific manner, for unknown reasons [41]. Dmc1 is a meiosis-specific strand exchange protein; dmc1 mutants can remove Spo11 and generate ssDNA tails, but are unable to carry out further recombination steps and thus accumulate hyper-resected DSBs that migrate faster on agarose gels [42]. Wild-type DSB distributions appear to be more faithfully represented in dmc1 mutants [41], [43]. Genomic DNA was purified from meiotic cultures of these mutants, restriction digested, and DSBs were detected by Southern blotting and indirect end-labeling (Figures 4C–4F). We chose four sites for physical analysis, reflecting a range of local Spo11 oligo distributions. As detailed below, all four showed good agreement between DSBs and Spo11 oligo maps, both quantitatively and spatially (Figures 4C–4G).
TyPEX25-CAR1 had the highest Spo11 oligo density nearby because of a strong hotspot immediately adjacent to its 5′ LTR (Figure 4C and Table S1). This hotspot was among the hottest 0.5% of all hotspots compiled previously [19]. A much weaker hotspot was also present adjacent to the 3′ LTR. TyEST3-FAA3 also had a strong hotspot near its 5′ LTR (Figure 4D). This hotspot was again within the hottest 0.5%, but was relatively wide. A weaker hotspot was present on the 3′ side of this Ty, close to a tRNA gene and the EST3 promoter (discussed further below). Both TyPEX25-CAR1 and TyEST3-FAA3 are in intergenic regions containing a tRNA gene between divergently transcribed genes (Figures 4C and 4D). In both cases, the region next to the 5′ LTR carries the strong hotspot even though the region next to the 3′ LTR also contains a promoter. These two loci demonstrate that presence of a Ty can be compatible with very high DSB activity nearby.
TyCGR1-SCW11 showed weak DSB levels adjacent to the 5′ LTR (Figure 4E) as well as within the Ty, discussed below. This Ty is in an intergenic region containing a tRNA gene between convergent genes. A modest DSB and Spo11 oligo hotspot was also observed ∼2 kb away in the SCW11 promoter (Figure 4E). TyURA3 also had a weak DSB hotspot nearby (Figure 4F). This hotspot was in the ura3 promoter, coinciding with the 5′ LTR side of the Ty. Trace numbers of Spo11 oligos mapped in the ura3 coding sequence adjacent to the 3′ LTR, but the corresponding DSB signal was too weak to be detected (Figure 4F and data not shown). TyCGR1-SCW11 and TyURA3 exemplify a situation in which presence of a Ty correlates with low DSB levels nearby, but do not speak to whether the Ty causes the low DSB activity.
Spo11 oligo mapping showed that meiotic DSBs occur within Ty elements [19], but individual Tys could not be evaluated. Physical analysis revealed a modest DSB hotspot inside TyCGR1-SCW11 (Figure 4E). DSBs overlapped the 5′ LTR and a region ∼1.8 kb from the 5′ end of the Ty, inside the Gag coding sequence. DSB signal was not detected near the 3′ end when the Southern blot was reprobed from the opposite side of the restriction fragment (data not shown), thus DSBs are more frequent near the 5′ end for this Ty. The Ty element that disrupts ura3 also showed evidence of DSBs near its 5′ end, but at a level too low to be quantified (Figure 4F, inset). We did not observe discrete DSB signals inside either TyPEX25-CAR1 or TyEST3-FAA3 (Figures 4C and 4D), so these Tys lack hotspots above the limit of detection by Southern blotting (∼0.01% of DNA). Infrequent, relatively disperse DSBs would not be detected in this analysis. These results show that Tys differ significantly from one another in terms of number and location of internal DSBs. Interestingly, break levels in the flanking regions do not necessarily correlate with levels inside the Ty.
Open chromatin structure provides a window of opportunity for Spo11-dependent DSB formation [37]. To investigate the relationship between DSBs and chromatin structure at Ty elements, intact nuclei were prepared from meiotic cultures of wild-type cells and partially digested with micrococcal nuclease (MNase). DNA was extracted and digested with appropriate restriction enzymes, and MNase cleavage sites were identified by Southern blotting and indirect end-labeling (Figure 5). MNase digestion of purified genomic DNA was examined in parallel. Nucleosomal DNA is relatively resistant to MNase cleavage (Figure 5A). For example, the SCW11 promoter showed a broad band of preferred MNase digestion indicative of a nucleosome-depleted region (NDR) typical of many yeast promoters, flanked by ladders of bands from cleavage in the linkers between positioned nucleosomes upstream and downstream of the promoter (Figure 5B, lanes 2–3). As expected, the DSB hotspot in the SCW11 promoter corresponded to the MNase-hypersensitive NDR (Figure 5B, lanes 2–3 vs. lane 5).
TyCGR1-SCW11 showed dispersed MNase cleavage inside, with two prominent MNase-hypersensitive zones toward its 5′ end, one of which corresponded to the DSB hotspot within this Ty (Figure 5B, lanes 2–3 vs. 5). Within each hypersensitive zone a weak banding pattern could be seen, suggesting a modest tendency for nucleosomes to occupy certain preferred positions in subpopulations of cells. Within the Ty element, 28.3% of DNA was cleaved (4.7% per kb), compared with 30.7% of DNA cleaved between the 5′ LTR and the end of CWH41 (11.8% per kb). Thus, this Ty overall is only about two-fold more resistant to MNase than the intergenic and genic regions flanking it.
In contrast, TyPEX25-CAR1 appeared less sensitive to MNase compared to flanking genic regions. Whereas 17.3% of DNA was cleaved in the intergenic region between the 3′ LTR and the start of PEX25 (43% per kb), 33.3% of DNA was cleaved within TyPEX25-CAR1 (5.6% per kb). TyPEX25-CAR1 did not show prominent hypersensitivity toward its 5′ end (Figure 5C, lanes 2–3). Instead, it showed a broad region of modest hypersensitivity at its 3′ end, suggestive of an array of weakly positioned nucleosomes extending into the flanking intergenic region. These results show that chromatin structure can vary between individual Ty elements. Importantly, MNase-hypersensitive sites indicative of NDRs were present at both the strong DSB hotspot in the CAR1 promoter and the weaker hotspot in the PEX25 promoter flanking TyPEX25-CAR1 (Figure 5C, lanes 2–3 vs. 5). Thus, presence of a Ty close by need not result in elimination of the open chromatin structure typical of promoters and DSB hotspots.
To test whether natural Ty elements directly affect adjacent DSB formation, we individually deleted two Tys and compared DSB patterns with and without these elements present. As a control, we quantified DSBs in the same cultures at the YCR048W hotspot on Chr III; DSBs at this hotspot were similar between the parental and Ty deletion strains (Figure 6E).
Remarkably, a strain lacking TyEST3-FAA3 experienced ∼2–3 fold fewer DSBs in the FAA3 promoter region than the parental strain carrying this Ty (hotspot i in Figures 6A and B). Results were similar irrespective of which side of the genomic restriction fragment was probed. Although DSB levels were different, their distribution within the hotspot was unchanged (Figure 6B). The other hotspots in the probed region were affected little if at all in the strain lacking the Ty (hotspots ii, iii, and iv in Figures 6A and 6B).
In a strain lacking TyCGR1-SCW11, the weak DSB signal near the 5′ end of the Ty element became undetectable (hotspot v, Figure 6C), and the hotspot in the SCW11 promoter showed 2.3-fold lower DSBs than the parental strain (hotspot vi, Figure 6C). The weak hotspots on the other side of the Ty insertion site were essentially unchanged (hotspots vii–ix, Figure 6D). As expected, the DSB signal inside the retrotransposon was not observed in the Ty-deletion strain (hotspot x, Figure 6C), but no new DSB signal arose in its place as would have been expected if presence of the Ty were suppressing an otherwise active DSB site.
These findings do not support the hypothesis that Ty elements invariably suppress meiotic DSB formation in their vicinity. Instead, we conclude that at least some Ty insertions cause an increase in DSBs nearby.
Prior analyses of nucleotide variation demonstrated that SK1 is genetically distant from S288C [26], [44]. Accordingly, we find that the catalogs of full-length Ty elements are completely different in these strains, except for an ancient and immobile copy of Ty5. Full-length Tys are prone to loss by LTR-LTR recombination [45]. S288C does not have full-length Tys or solo LTRs where Ty elements reside in SK1 (data not shown), suggesting that transposition of the SK1 Tys occurred after SK1 and S288C diverged from their last common ancestor. While SK1 does not have full-length Tys at the same sites as in S288C, we did not comprehensively map solo LTRs, so it is possible that some S288C Ty elements predate divergence of the strains and were lost in SK1 by LTR-LTR recombination. It will be interesting to identify if any solo LTRs are shared between SK1 and S288C. Such LTRs would be “fossils” of ancestral transposition events, and comparison of their features with those of younger LTRs or Tys may illuminate how host-Ty element relationships have evolved.
In principle, the deep sequencing approach we used for Ty mapping should be broadly applicable to repetitive elements of any type in any organism. Indeed, while this work was in progress, others independently used a similar method to identify new transposon insertions in Drosophila [46]. This approach, combined with growing libraries of whole-genome, paired-end sequencing data from widely divergent S. cerevisiae strains, will facilitate assembly of complete genome sequences and also permit genealogical analysis of Ty insertion site diversity.
Chromosomal rearrangements can arise in vegetatively growing cells as a consequence of NAHR between Ty elements [12]–[15], [47]–[49]. Ty location and orientation dictate the degree of susceptibility to rearrangement, the structures of rearranged chromosomes, and whether the outcome is deleterious, neutral, or advantageous. For example, in S288C, deletion of HTA1-HTB1 (one of two gene pairs encoding histones H2A and H2B) causes pleiotropic defects that both promote and select for amplification of the separate HTA2-HTB2 locus [47]. Amplification occurs via NAHR between two flanking Ty elements in direct repeat orientation near the centromere of Chr II (see Figure 1). These Ty elements are not present in the W303 strain, so facile amplification of HTA2-HTB2 is not possible and deletion of HTA1-HTB1 is lethal in this strain [47]. SK1 also lacks similarly positioned Tys (Figure 1), so we anticipate that deletion of HTA1-HTB1 would be lethal in this strain too. Moreover, closely juxtaposed Ty elements in inverted orientation can create fragile sites predisposed to chromosome rearrangement [14], [50]. SK1 has no instances of closely spaced, inverted, full-length Ty pairs, but the Ty fragment TyEXG2-YDR262W-2 is juxtaposed in inverted orientation to full-length TyEXG2-YDR262W-1 on Chr IV, and TyNCE103-YNL035C-2 is inserted in inverted orientation into TyNCE103-YNL035C-1 on Chr XIV. These are thus candidates for fragile sites in this strain. More generally, these scenarios (histone gene amplification and Ty-associated fragile sites) illustrate the importance of Ty maps in different strains because the particular details of Ty element distribution are critical for understanding the influence of these retrotransposons on genome instability and evolution of genome structure.
Ty-mediated NAHR also occurs during meiosis [4], [22], [23]. We show here that four individual Ty elements experience different frequencies of DSBs inside. To our knowledge, this is the first direct detection of meiotic DSBs in Ty elements, confirming the inference from Spo11 oligo mapping that significant numbers of DSBs occur within Tys [19]. We detected a total internal DSB frequency of at least 0.1–0.3% of DNA in TyCGR1-SCW11. Assuming at most one DSB per four chromatids in a given cell, this frequency predicts that 0.4–1.2% of meiotic cells experience a DSB within this Ty element alone. This number is small on a per-cell basis, but becomes substantial when considered from the perspective of a population of cells or over many generations. Furthermore, we previously showed that ∼0.28% of Spo11 oligos map to Ty-derived sequences, indicating that one in every 2–3 meiotic cells experiences a DSB in a Ty or solo LTR, assuming an average of ∼160 DSBs per cell [19]. Excluding LTRs, ∼0.1% of Spo11 oligos map to Ty-internal sequences, which predicts a DSB frequency of 1.1–2.5% of DNA summed over all Ty elements, based on linear regression of Spo11 oligo counts vs. DSB levels (see Materials and Methods). This estimate is higher than the total DSB frequency observed in the four Tys assayed here, so it is likely that other Ty elements experience a significant number of DSBs as well.
Based on copy number compiled here, we estimate that Tys account for ∼1.5% of genomic DNA, not including rDNA or the contribution of solo LTRs. In turn, this suggests that DSBs within Tys are ∼15-fold suppressed relative to genome average since only ∼0.1% of total Spo11 oligos came from Ty-internal sequences. However, genome average includes many strong DSB sites, such as promoters, that are structurally and functionally dissimilar from the inside of a Ty, which is principally coding sequence. Genome wide, coding sequences account for only ∼11.5% of Spo11 oligos but occupy ∼69.4% of the genome. Thus, on average, Tys are only ∼2–3-fold colder than the typical open reading frame.
Our findings have implications for understanding behavior of outcrossed yeast strains: as a consequence of different Ty distributions, any DSB within a Ty would lack a recombination partner at the allelic position, so such DSBs are most likely repaired from the sister chromatid, by NAHR, or by single-strand annealing between 5′ and 3′ LTRs (which deletes the Ty-internal sequence leaving behind a solo LTR). It will be interesting to determine whether large-scale differences in Ty distributions contribute to reduced ability of hybrids to produce viable spores [51], [52], in turn contributing to reproductive barriers between strains. Our findings also have implications for inbred strains, including diploids produced by homothallic strains: DSBs within Tys have potential to provoke NAHR even if there is a Ty present at the allelic position on the homologous chromosome. Such NAHR may contribute to sequence homogenization and co-evolution of Ty elements. Moreover, our results provide a framework for studying mechanisms that act after DSB formation to minimize the risk of deleterious chromosome rearrangements [6].
Chromatin structure may play an important role in DSB formation within Ty elements, as suggested by the observation of MNase hypersensitivity at the 5′ end of TyCGR1-SCW11 where DSBs are formed. Our findings show that different Tys can have different chromatin architecture. In a similar vein, relative transcription levels of Ty1 elements in vegetatively growing S288C were found to differ by ∼50 fold [24]. Thus, Ty elements can differ greatly from one another, precluding generalization of a one-size-fits-all pattern from any given element.
Although DSBs wholly within Tys have greater potential to instigate NAHR, breaks in unique sequences near Ty elements may also be at risk because DSB resection generates recombinogenic ssDNA for significant distances (up to a kb or more) from the Spo11 cleavage site [53]–[55]. We find here that DSB levels vary substantially in regions flanking different Ty elements and that presence of a Ty does not invariably cause suppression of adjacent DSB activity. These findings are counter to predictions from prior analysis of a Ty in the HIS4 promoter [16], further highlighting the individual variability of Ty elements.
We propose that differences between the studies reflect aspects of host-transposon interactions that evolved to minimize deleterious effects of retrotransposition. The Ty at HIS4 mimicked a spontaneous Ty integration that disrupted HIS4 expression (Ty917) [20], [22]. It was inserted ∼70 bp upstream of HIS4, moving the TATA box and upstream activator sequence ∼6 kb away from their normal position and eliminating the DNase I hypersensitivity of the HIS4 promoter [16]. The altered chromatin structure was interpreted as spreading of closed chromatin from the Ty into surrounding regions [16], but an alternative interpretation is that Ty917 is simply an insertional mutation that compromises the cis-acting elements defining the HIS4 promoter NDR, thereby disrupting both promoter activity and Spo11 access. In this view, the effect of Ty917 on DSB formation is context dependent and intimately tied to its deleterious effect on a host gene.
In contrast to Ty917, most naturally occurring Ty elements are found near tRNA or other RNA pol III-transcribed genes, likely targeted there via interactions of integration complexes with RNA pol III transcription machinery [11], [56], [57]. Ty elements inserted near (and especially upstream of) tRNA genes will tend to be distant from regulatory regions of other adjacent genes because the mean distance between tRNA genes and their upstream neighbors (excluding Ty and LTR sequences) is ∼500 bp larger than the distance from tRNAs to downstream genes or the average size of intergenic regions genome-wide [58]. Thus, while Ty integration site preference may have evolved to prevent deleterious mutations [11], [36], [59], it has the additional consequence that Ty elements tend to avoid the very RNA pol II promoters where most meiotic DSBs are formed, and tend not to impinge on promoter properties that favor Spo11 activity, such as transcription factor binding and nucleosome depletion. Our direct analysis of chromatin structure and DSB formation around TyPEX25-CAR1 supports this view. The correlation between DSB levels and the class of Ty-bearing intergenic region (Figure 4B) also supports this idea by implying that DSB frequency is substantially influenced by the local DSB-forming potential of the neighborhoods where Ty elements reside.
We were surprised to find that deletion of two Ty elements in different genomic contexts caused decreased DSB formation nearby. Thus, at least some Tys stimulate adjacent DSB formation, and our genome-wide analysis suggested this may be a fairly general property. The mechanism behind this effect is as yet unclear. Both Ty deletions showed an apparent polarity in that the regions where DSB levels were most affected were adjacent to the 5′ LTRs. Although sample size is too small to know if this is a general pattern, it may indicate that adjacent DSB formation is modulated by properties of Ty 5′ LTRs, which in some cases carry promoter activity and contain binding sites of transcriptional activators [e.g., 24]. Alternatively, it may be that DSB stimulation is not a unique property of the Ty itself, but instead is simply a consequence of a structural change in the chromosome. Indeed, there are numerous examples where heterologous DNA insertions generate new DSB hotspots [reviewed in 8], although such insertions rarely, if ever, cause enhanced activity of natural, promoter-associated hotspots nearby.
Regardless of the mechanism, this finding has implications for inheritance of Tys across sexual cycles. The chromosome that experiences a DSB is the recipient of genetic information from its homologous partner, in part because of net degradation of the broken chromosome by DSB resection and resynthesis using the intact partner as the template [60]. As a consequence of this gene conversion bias, an allele with a higher propensity toward DSB formation will tend to be under-transmitted during meiosis. Thus, elevated DSB frequency near Tys might tend to favor elimination of Ty copies by meiotic recombination in diploids heterozygous for the Ty insertion. In principle, this tendency could affect new Ty insertions in a diploid or inbred population, as well as older Ty insertions in outcrosses between diverged strains. Our findings thus raise new questions about retrotransposon-host relationships and the roles of the intersection between Ty elements, meiotic recombination initiation, and NAHR.
Yeast strains are listed in Table S2. Ty elements were deleted by two-step gene replacement, resulting in precise replacement of each Ty element with a diagnostic restriction site (see legend to Tables S2 and S3). Other alleles were introduced by genetic crosses or by one-step gene replacements using standard methods. All gene replacements were confirmed by Southern blotting. A whole-genome mate pair library was prepared according to manufacturer's recommendations (Roche) from genomic DNA purified from a vegetative culture of NKY291, and sequenced on the Roche 454 platform. The NKY291 library had an average sequence length of 173 bp and an average insert size of 2.8 kb (16-fold coverage in 653,261 sequence pairs). Sequence data are available at http://cbio.mskcc.org/public/SocciN/SK1_MvO/Data/GCL0188__454__PE_3k/
To evaluate presence of Tys from S288C or previously identified in SK1 [25], SK1-derived sequence reads from the SGRP were viewed in the genome browser provided by the Sanger Institute (http://www.sanger.ac.uk/research/projects/genomeinformatics/sgrp.html). When read alignment patterns were indicative of Ty presence, partial DNA sequences of the Tys were deduced from contigs assembled from these reads and used to determine Ty orientation and family, by comparison to exemplars of Ty families from S288C. Ty insertion sites were mapped by identifying SGRP reads overlapping boundaries between Tys and flanking genomic sequence. If no overlapping reads were present, PCR products spanning the Ty-element-containing region were partially sequenced to determine the precise insertion sites.
For systematic Ty mapping, the SK1 mate pair libraries from SGRP and from our sequencing of NKY291 were mapped against a compilation of non-LTR portions of S288C Ty elements. Mapping was performed using LastZ on the Galaxy server (http://main.g2.bx.psu.edu/). Mate pairs of reads that aligned with Ty-internal sequence were then mapped onto the S288C genome using LastZ, and reads that mapped to multiple positions were discarded. Candidate Ty insertion sites identified from the remaining reads were validated by manual inspection of sequence alignments and/or PCR of genomic DNA. In addition to the full-length Tys and large Ty fragments listed in Table 1, this analysis identified three small (∼120–180 bp) non-LTR Ty fragments at ∼805 kb on Chr IV, ∼78 kb on Chr VIII, and ∼338 kb on Chr XVI (data not shown). How these insertions arose is uncertain, but because they are so short, they were not considered as Tys in this study.
Synchronous meiotic cultures were prepared essentially as described [61]. Cells were harvested from a single culture of SKY4121, two independent cultures of SKY4151 and of SKY4153, and single cultures of SKY4188, SKY4189, SKY4191 and SKY4192 at 6 hr in meiosis, and genomic DNA was isolated in low melting temperature agarose plugs, digested with appropriate restriction enzymes, electrophoresed on agarose gels, and analyzed by Southern blotting and indirect end-labeling, as described previously [19], [61]. Restriction enzymes and probes are as follows and primers used to prepare probes are in Table S3: TyPEX25-CAR1, BamHI, PEX25 probe; TyEST3-FAA3, Bsu36I, DOT5 or EPS1 probe; TyCGR1-SCW11, BamHI, RPS24A or CWH41 probe; TyURA3, BamHI, GEA2 probe; YCR048W hotspot, BglII, RCS6 probe. Hybridization signal was detected and quantified with Fuji phosphor screens and ImageGauge software. DSB frequency was determined as the percent of radioactivity in DSB fragments relative to total radioactivity in the lane. Signals from the spo11-Y135F strain were used to subtract background.
The large difference in size of the parental-length restriction fragments between Ty-containing and Ty-deleted loci (experiments in Figures 6A–6D) could be expected to cause differences in Southern blot transfer efficiencies, which could lead to incorrect estimates of relative DSB levels. To account for this, we applied the following strategy. First, Ty+ and TyΔ samples were run on the same gel, transferred together, and hybridized together to the appropriate probe for the Ty locus. The membranes were then stripped and re-hybridized to probes from different loci to serve as loading controls: YCR057C probe for the blots shown in Figures 6A and 6B and YKL182W probe for the blots in Figures 6C and 6D (Table S3). We used the loading controls to correct DSB estimates by assuming that there was “missing signal” from the Ty+ lanes because of less efficient transfer of Ty-containing DNA fragments. From this analysis, we estimated that the parental bands in the Ty-containing strains were transferred at ≥75% the efficiency seen with the Ty-deletion strains (data not shown).
Meiotic culture of wild-type diploid, SKY41, was prepared as described above. Intact meiotic nuclei were prepared 4 hrs after induction of sporulation by spheroplasting, hypotonic lysis, and centrifugation on sucrose step gradients, as described previously [62]. Nuclei were quantified by fluorometry with Hoechst 33258 dye. A volume of nuclear suspension containing 4 µg DNA was diluted with an equal volume of ice-cold 10 mM Tris-HCl, pH 8.0, 5 mM MgCl2 and 1 mM Pefabloc. Nuclei were collected by centrifugation and resuspended in 90 µl of 10 mM Tris-HCl, pH 8.0, 2.5 mM CaCl2, 3.5 mM MgCl2 on ice. Ten µl of appropriate concentration of MNase (Worthington) was added, digestion was performed for 5 min at 37°C, then terminated by addition of 0.4 ml of 62.5 mM EDTA, 125 mM Tris-HCl, pH 8.0, 0.625% SDS and 5 µl of 20 mg/ml proteinase K. Samples were incubated at 58°C for 2 hrs to overnight. DNA was extracted twice with phenol∶chloroform∶isoamyl alcohol (25∶24∶1) and once with chloroform, then precipitated with isopropanol with 10 µg of glycogen and dissolved in 10–20 µl of dH2O. As a control, genomic DNA was purified from vegetatively growing cells and treated with MNase, followed by purification as above. DNA from MNase-treated nuclei or naked DNA was digested with BamHI, electrophoresed on agarose gels, and analyzed by Southern blotting and indirect end-labeling using the CWH41 probe (TyCGR1-SCW11) or PEX25 probe (TyPEX25-CAR1) (Table S3).
For the analysis in Figure 4A, groups of closely neighboring Tys in SK1 (between EXG2 and YDR262W on Chr IV, and between NCE103 and YNL035C on Chr XIV) were treated as single Tys. Furthermore, the Ty5 at the left end of Chr III was excluded, as DSBs are known to be suppressed in subtelomeric regions [19], [41], [43]. Therefore, Spo11 oligo counts were determined in 27 Ty-bearing regions in SK1 (Table S1). As controls, we used the coordinates of S288C Ty elements. We excluded S288C Ty positions within 2 kb of SK1 Ty elements, closely neighboring Tys were considered as a single element, and YCLWTy5-1 was excluded as above. Spo11 oligo densities adjacent to 37 control sites were determined.
To estimate the genome-wide percentage of DNA broken in Tys, we summed Spo11 oligos that mapped to non-LTR Ty sequences. Excluding LTRs means that we are underestimating DSBs associated with full-length Tys, but this is necessary because we cannot distinguish Spo11 oligos from LTRs flanking Ty elements from those originating within solo LTRs or LTR fragments. Using our previously defined regression relationship [19], we converted Spo11 oligo counts to DSB frequency, yielding estimates in dmc1 and sae2 background of 2.0% and 0.85%, respectively. Since the prior study used the spo11-HA strain, which forms DSBs at a reduced frequency of ∼80% of a SPO11+ strain [63], we therefore estimate the Ty DSB frequency to be ∼2.5% in dmc1 and 1.1% in sae2 in the SPO11+ background.
|
10.1371/journal.pcbi.1000158 | Modeling ChIP Sequencing In Silico with Applications | ChIP sequencing (ChIP-seq) is a new method for genomewide mapping of protein binding sites on DNA. It has generated much excitement in functional genomics. To score data and determine adequate sequencing depth, both the genomic background and the binding sites must be properly modeled. To develop a computational foundation to tackle these issues, we first performed a study to characterize the observed statistical nature of this new type of high-throughput data. By linking sequence tags into clusters, we show that there are two components to the distribution of tag counts observed in a number of recent experiments: an initial power-law distribution and a subsequent long right tail. Then we develop in silico ChIP-seq, a computational method to simulate the experimental outcome by placing tags onto the genome according to particular assumed distributions for the actual binding sites and for the background genomic sequence. In contrast to current assumptions, our results show that both the background and the binding sites need to have a markedly nonuniform distribution in order to correctly model the observed ChIP-seq data, with, for instance, the background tag counts modeled by a gamma distribution. On the basis of these results, we extend an existing scoring approach by using a more realistic genomic-background model. This enables us to identify transcription-factor binding sites in ChIP-seq data in a statistically rigorous fashion.
| ChIP-seq is an apt combination of chromosome immunoprecipitation and next-generation sequencing to identify transcription factor binding sites in vivo on the whole-genome scale. Since its advent, this new method has generated much excitement in the field of functional genomics. Proper computational modeling of the ChIP-seq process is needed for both data scoring and determination of adequate sequencing depth, as it provides the computational foundation for analyzing ChIP-seq data. In our study, we show the characteristics of ChIP-seq data and present in silico ChIP sequencing, a computational method to simulate the experimental outcome. On the basis of our data characterization, we observed transcription factor binding sites with excessive enrichment of sequence tags. Our simulation results reveal that both the genomic background and the binding sites are not uniform. On the basis of our simulation results, we propose a statistical procedure using the more realistic genomic background model to identify binding sites in ChIP-seq data.
| Gene expression is carefully regulated in all living cells. Only a fraction of the genes in a genome are expressed to various degrees under a given condition or in a particular cell type. The main control of such regulation occurs at the transcription level: the RNA polymerases transcribe genes following binding of trans-acting transcription factors to cis-acting regulatory DNA sequences within genes or in their vicinities. To determine the biological functions of transcription factors, it is imperative to identify their binding sites and target genes in the genome.
Currently the most commonly used high-throughput method for identifying transcription factor binding sites (TFBSs) is chromatin immunoprecipitation followed by microarray hybridization (ChIP-chip) [1]–[3]. In this method, the transcription factors are cross-linked to DNA under the test condition. After the genomic DNA is isolated and fragmented by sonication, an antibody specific to the transcription factor of interest is used to isolate the transcription factor and the DNA fragments which it binds. Following chromatin immunoprecipitation, the protein–DNA crosslink is reversed and the DNA fragments are hybridized to a tiling microarray. After the signal quantification, the DNA fragments enriched by the binding of the transcription factor are identified—in terms of both genomic sequence and location—by the oligonucleotide tiles that give significantly high relative signals on the microarray [4].
Instead of using microarrays to identify the sequences of the immunoprecipitated DNA fragments, new methods have recently been developed to take advantage of the fast-maturing next-generation massively parallel sequencing technologies. In one such method, ChIP-PET [5], paired-end ditags (PETs) derived from both ends of the immunoprecipitated DNA fragments are sequenced and mapped to the genome. In a newer method, ChIP-seq [6],[7], immunoprecipitated DNA fragments are directly sequenced at one end for ∼30 bp, and the short sequence reads are then mapped to the reference genome. The apt combination of ChIP and next-generation sequencing technology has generated much excitement in the field of functional genomics. Comparing with ChIP-chip, whose usability for large mammalian genomes is limited by serious cross-hybridization at high genomic resolution, these sequencing-based methods offer not only direct whole-genome coverage but also low analytical complexity, high signal-to-noise ratio, and sensitivity that increases with sequencing depth. The current trend in high-throughput molecular biology laboratories is to migrate from ChIP-chip to ChIP sequencing to identify transcription factor binding sites in vivo.
Proper computational modeling of ChIP-seq process is needed for both data scoring and determination of adequate sequencing depth, as it provides the computational foundation for analyzing ChIP-seq data. Here we show the characteristics of ChIP-seq data and present in silico ChIP sequencing, a computational method to simulate the experimental outcome. Our simulation results reveal that both the genomic background and the binding sites are not uniform. Such nonuniformity in the background will have important implications in ChIP-seq data analysis and binding sites identification.
ChIP-seq data are generated in a straight-forward manner, by high-throughput sequencing and subsequent sequence alignment. Because Illumina/Solexa 1G Genome Analyzer generates a very large number of short sequence reads, ChIP sequencing is currently done mainly with this sequencing platform. This could change in the future, however, as other high-throughput sequencing technologies may become better suited. Here we briefly describe the procedure of ChIP sequencing with the Solexa platform. The immunoprecipitated DNA fragments are sequenced from one end for approximately 30 bp. These short sequence reads are aligned to the human reference genome, and only uniquely mapped reads (typically 60–80% of all sequence reads) are retained for the downstream analysis. Based on size selection after gel electrophoresis prior to sequencing, the retained reads are elongated into longer tags by directional extension to the mean length of the size selected DNA fragments and then transformed into profiles of the number of overlapped DNA fragments at each nucleotide in the reference genome [6].
For our analysis, we link overlapping tags into tag clusters (Figure 1), each of which is characterized by y, the number of tags it contains, and indexed by a and b, its start and end genomic locations. Thus, by definition a tag cluster is a genomic site continuously covered by one or more sequence tags can be characterized in two different ways. One type of characterization is to set a and b to the boundaries of the cluster and y to the number of all tags in it, while the other is to identify the peak of the overlap in the cluster first and then to set a and b to the start and the end positions of the peak and y to the height of the cluster. We term tag clusters characterized by these two methods as ‘outer clusters’ and ‘inner clusters’ respectively and use ‘outer clusters’ in our analysis. Suppose there are M tag clusters, ChIP-seq data after preprocessing are defined by the matrix T, whose row m, (am, bm, ym), characterizes tag cluster m (m = 1,…, M). The main goal of our ChIP-seq data analysis is to identify tag clusters that are transcription factor binding sites by determining a threshold on the tag count to separate the DNA-binding signals from the background noise.
To identify transcription factor binding sites in ChIP-seq data, we assess the statistical significance of each tag cluster found in the actual data by assigning it a P-value as the result of the test of the null hypothesis that its tag count is generated by a null distribution, which is the distribution of the tag count on the genomic background alone. This null distribution is generated by placement of sequence reads onto the genomic background in the absence of binding sites. It is critical to simulate the correct background, as the null distribution generated from it is used to assign P-values to all actual tag clusters.
The simulation starts with the removal of sequence gaps and repeats from the genomic region—the entire genome or a part of it—under consideration. It is followed by the random placement of n sequence tags, corresponding to the same number of uniquely-mapped sequence reads from the experiment, onto the genomic background, whose distribution of the sampling weight on the nucleotide level could be either uniform or non-uniform. After the tag placement, suppose that N tag clusters are identified in the simulated data and the largest one contains C tags, thus the null distribution of the cluster tag count is given by the number of tag clusters on each tag count level, 1, 2, …, C: .
Given this null distribution, for tag cluster m (m = 1, 2, …, M) identified in the experimental data we calculate its associated P-value, Pm, for the test of the null hypothesis that it is part of the background asin which ym is the tag count of tag cluster m from the experimental data and kc is the number of tag clusters on tag count level c in the simulated data. In essence this is a permutation test and Pm can be calculated to arbitrary accuracy as the number of simulation increases. To control the type I error in this set of M hypothesis tests, we first adjust the P-values so that they directly reflect the controlled false discovery rates [8], and then choose the lowest tag count that gives a low FDR (e.g., less then 0.05) as the threshold. Tag clusters with at least this tag count are identified as the binding sites.
For our simulation of ChIP sequencing (Figure 2), we use the lengths of human chromosomes as specified in the NCBI v36/hg18 human genome assembly. We first remove all sequence gaps as defined in the UCSC genome browser annotation database. Because only uniquely mapped sequence reads are used in ChIP-seq data analysis, we also remove positions covered by repetitive sequences identified by RepeatMaster, and then randomly place without overlap a chosen number of transcription factor binding sites, each of which was assumed 500 bp long, onto the genome. After the placement of binding sites, the genome (excluding removed sequence gaps and repeats) is effectively partitioned into the floating fixed foreground (binding sites) and the background.
The process of the chromosomal immunoprecipitation and the subsequent unique mapping and extension of sequence reads can be simulated by randomly placing uniquely mapped sequence tags onto the chromosome, according to certain sampling weight at each nucleotide position. Such weights are generated first for the background nucleotide positions and then for those in the binding sites. For a uniform background, every nucleotide position in the background is given one as its sampling weight. For a varying background, if we assign each nucleotide position a different weight, given the large size of the human genome it becomes computationally prohibitive to sample the background many times as the simulation requires. Instead, we partition the background into adjacent blocks of nucleotide positions. After testing different block sizes ranging from 500 bp to 5 kb, we find they all give practically identical simulation results. In the end, we choose 1 kb as the block size. Every adjacent 1-kb block in the background is given a random weight drawn from a pre-specified underlying distribution and all nucleotide positions in a block are assigned the same weight. For the background variation, we assume that most of the background has a low sampling weight as most of the background is not enriched in the immunoprecipitation (the working principle of ChIP) but a few places of it have relatively high weights, comparable to some binding sites. Based on this assumption, we use a gamma distribution, Gamma(s,c), which skews to the right, for the distribution of sampling weight on the background.
To specify the sampling weights in the binding sites, we first calculate w̅b, the average sampling weight at each nucleotide position in the background, and multiply it by the enrichment coefficient t, to obtain w̅f = t⋅w̅b, the average sampling weight at each nucleotide position in the binding sites. The ChIP enrichment at different binding sites is, however, different and can be estimated by the fold increase of tags placed in the foreground over those placed in the background in the simulation. Given w̅f and the number of nucleotides in the binding sites, we calculate Wf, the total amount of sampling weight in the binding sites, and then distribute it to each binding site either evenly or varyingly according to a certain distribution. For the intersite variation, we use a power-law distribution generated by a “preferential attachment” procedure. If a tag is placed in a binding site, the current sampling weight of this site, wk, is updated by a linear function as wk = w+r·k·w, in which w is its initial sampling weight, k is the number of tags placed at this site, and r is the weight increase coefficient. For each binding site, we also distribute the amount of its sampling weight to each nucleotide position according to a symmetric binomial or an equilateral triangular profile. We test various combinations of values for s, c, and r, the two free parameters in our simulation method, and find s = 1, c = 20, and r = 1.5 produce simulated data that give overall best fit to the actual data.
We implemented our ChIP-seq simulation method in R and wrote several auxiliary programs for text processing in Perl. The whole software package with source code and documentation is available for download at http://www.gersteinlab.org/proj/chip-seq-simu.
For our analysis and simulation of ChIP-seq data, we used the dataset generated from STAT1 DNA binding under IFN-γ stimulation by Robertson et al. [6]. Of the initial 2,915,382 sequence reads obtained in their experiment, 2,025,931 (69.5%) could be uniquely mapped to the unmasked NCBI v36/hg18 human reference genome. After the genomic mapping, we extended the length of mapped sequence reads from 27 to 174 bp, the estimated average length of the size selected DNA fragments [6], and identified 1,264,752 STAT1 tag clusters on the whole genome level.
While the majority (1,149,405, >90%) of these tag clusters comprise only one or two tags, a relatively small number (661) of them contain large numbers of tags (50 and more, the outer-overlapping count) and consequently show high stacking peaks (the inner-overlapping count) in their profiles (Figure 3A). For example, the most prominent STAT1 tag cluster appears immediately upstream to the centromere of chromosome 1. With a peak height of 472 tags, it comprises 1,733 tags in its ∼1.6 Kb genomic footprint. Indeed, a closer examination revealed that the tag count c follows a power-law distribution:where the degree exponent γ = 2.97 (R2 = 0.9955, P-value<2×10−16) for the outer count and 3.44 (R2 = 0.9976, P-value<2×10−16) for the inner count, respectively (Figure 3A and 3B).
We also examined tag counts on individual human chromosomes separately to check for possible discrepancies in their distributions on different chromosomes. The plots in Figure 3C and 3D show that over all the tag count on individual chromosomes and on the genome as a whole follows the same power-law distribution, and there is considerable variation among different chromosomes in the distribution at high counts.
In our simulation of the ChIP-seq process, we use either uniform or varying sampling weights on the genomic background and among the binding sites for the tag placement. The four simulated datasets generated from the resultant combinations of the background and the inter-site distributions fit the actual data in very distinct ways (Figure 4). The goodness of fit is assessed by the fit of the simulated distribution to the actual one in the range of small to high tag counts.
The four combinations of the background and the inter-site distributions can be seen as a gradual increment in the overall simulation complexity: from a simple model that assumes uniformity in both the background and the binding sites to one that assumes variation in either of them and to the most complex one that assumes variation in both of them. The simplest model assumes that the tag placement is identical everywhere on the background and also identical among the binding sites. Data generated from this model give a distribution of tag counts that is a very poor fit to the actual one: not only is there a depletion of tag clusters with small to medium tag counts due to an excess of single tags being placed onto the genome, but also clusters with large tag counts are completely absent (Figure 4A and see the Table S1 for the quantification of the goodness of fit).
The slightly more complicated second model assumes identical binding sites but a varying background for tag placement. The simulated data fit the actual distribution well at small to medium (1 to ∼5) tag counts but there is still a complete absence of clusters with large tag counts (Figure 4B). Contrary to the second model, the third model assumes a uniform background but varying binding sites for tag placement instead. Using this model we see an inversion in the simulation result: tag clusters with small to medium tag counts are depleted in the simulated data while clusters with large tag counts are generated (Figure 4C). Finally, we use a model that assumes variation both in background and among binding sites for tag placement. It generates data that give the best fit to the actual distribution of the tag count in its whole range (Figure 4D).
To identify STAT1 binding sites, we can assess the statistical significance of each tag cluster found in the actual data using a null distribution of tag counts derived from a background model. For the initial assessment, we used a simple background model that assumes equal probabilities for random tag placement at every available nucleotide position in the genome and combined 500 independent replicates of background simulation to generate such a null distribution. After assigning P-values and adjusting them for multiple testing to control the false discovery rate, we set five and above, which corresponds to an FDR<0.05, as the threshold on the tag count and identify 32,763 STAT1 binding sites.
In light of the simulation results, we can reassess the statistical significance of each tag cluster found in the actual data by using the varying-background model and combining 500 independent replicates of background simulation to generate the null distribution of the tag count. As before, we assign P-values to tag clusters found in the actual data by using this null distribution and adjust them for multiple testing to control the false discovery rate. At the same FDR level (<0.05), we set thirteen and above as the threshold on the tag count and identified 5,858 STAT1 binding sites from the initial ∼3-million sequence reads.
Using the full sets of reads, we identified 28,434 and 5,307 STAT1 binding sites with and without IFN-γ stimulation respectively (Table S2). In their study, Robertson et al found 41,582 and 11,004 sites in these two datasets. The reduction in both of our numbers reflects a more stringent threshold for peak calling, which was set by the more realistic varying-background model. Moreover, the proportionally greater decrease in the number of sites without stimulation reflects the limitation of STAT1 as a transcription factor without IFN-γ stimulation. To demonstrate the validity of the threshold change, we performed a STAT1 motif analysis in the peaks that are between the thresholds set by the uniform background and the varying background models. Using Meta-MEME [9] with blocksize = 128,205 characters, background = peaks.bg (nucleotide frequencies estimated from the input peak sequences), and E-value<1, we are able to identify significant STAT1 motifs (as defined in TRANSFAC [10] and JASPAR [11]) in 6.1% of those peaks. This result suggests that the threshold increase greatly boosts the specificity at a very small expense of the sensitivity.
Four distributions of tag counts are plotted in Figure 5: two actual distributions generated by experiments with and without IFN-γ stimulation and two null distributions derived from the uniform- and the varying-background models. Compared with either null, there is a significant increase of the number of tag clusters with high tag counts in the observed stimulated distribution. For example, there are 661 tag clusters with 50 or more tag counts in the actual data but none in the simulated data generated with either background model. While the number of tag clusters strictly decreases monotonously as the tag counts increases in the null distribution, there is a long tail on the right of the actual distribution given by the enrichment of clusters with high tag counts. Moreover, we also observe significant differences between the simulated datasets generated with two background models alone. First, comparing with 903,832 tag singletons in the actual data, there is an enrichment of tag singletons in all simulated background datasets. However, this increase is much more pronounced in the datasets generated with the uniform-background model (∼150%) than with the varying-background model (∼115%). Second, on average there are only three tag clusters with nine or more tag counts in the data simulated with the uniform-background model but over 2,000 with the varying-background model.
To check how closely the varying-background model models the background in the actual experimental data, we compared the distribution generated under this model with the actual one without the IFN-γ stimulation. In response to the stimulation, STAT1 binds to numerous promoter elements to upregulate interferon stimulated genes. Without the stimulation, the role of STAT1 as a transcription factor is limited. Given such a difference in the DNA binding of STAT1 in the presence or absence of IFN-γ, we expect the distribution of tag counts from the experiment without stimulation should be a distribution dominated by a significant background with a small right tail from its limited DNA binding. This is exactly what we see in Figure 5, where the good fit between the distribution simulated under the varying-background model and the actual unstimulated one is striking and shows the validity of the varying-background model. Considering these observations (Figure 5) in the light of the full simulation results presented in the previous subsection (Figure 4), we conclude that as the genomic background is varying it is better captured by the varying model than the uniform one.
We generate synthetic ChIP-seq datasets under simulation models with various assumptions for the binding sites and the genomic background. By comparing the simulated dataset with the actual one, we assess the goodness of the assumptions made in each simulation and thus can gain insight into the actual ChIP-seq data generating process: the closer the simulated dataset is to the actual one, the closer the assumptions are to the real process.
We use the uniform and the varying models for both the background and the binding sites in our simulation. In Figure 5, marginal comparisons show that the model with a varying (non-uniform) weight distribution for either the background or the binding sites generates substantially better simulated data. When the background and the binding sites are considered together, the simulated datasets generated with various combinations of the background and the binding-site models show striking differences in their quality. The data simulated with the uniform-weight models used for both the background and the binding sites show practically no fitting to the actual data except for the general trend (Figure 5A). When the varying-weight model is used for either the background or the binding sites, there are substantial improvements to the fit in different ranges of the tag count (Figure 5B and 5C). However, when the varying-weight models are used for both the background and the binding sites, not only is the fit the best but also there is a general agreement between the simulated and the actual data (Figure 5D).
These simulation results clearly show that neither the binding sites nor the background is uniformly presented in ChIP-seq data. Due to the inherent random noise in the experiment, binding sites are unlikely to contain the same number of mapped sequence tags. Not all the variance in the number of sequence tags mapped to binding sites could be explained by random noise, which should be counted by the uniform-site model as the simulation itself is intrinsically a stochastic process. Because DNA segments containing the binding sites are enriched by immunoprecipitation, the variance should also reflect the different DNA-binding affinity that a transcription factor has for its binding sites. Such variation could be the result of differences in either the nucleotide sequences of the binding sites [12] or the local chromotin modification status [13].
Perhaps more importantly, our simulation results also reveal that there is a substantial variation in the tag placement on the genomic background. Obviously, such background variation cannot be explained by the uniformity of background currently assumed in ChIP sequencing. Instead, our results suggest a varying background that is mildly fluctuating and contains some “hot” spots with relatively high ChIP enrichment comparable to some binding sites. The presence of such background ‘hot’ spots in the ChIP-seq data may be caused by preferential sequencing particular to the sequencing protocol/platform used in the experiment. Their enrichment through immunoprecipitation is precluded, however, as the background DNA segments are not bound by the transcription factor. Our inference of a varying genomic background not only raises questions about both biology and technology involved in ChIP sequencing but also has important practical implications to the analysis of ChIP-seq data as it provides a better background model (see next subsection for explanation).
To examine our simulation results more closely, we plot in Figure 6 the actual tag count distribution and the simulated ones generated under different background and site models with the enrichment coefficient t = 10 only (the blue lines in Figure 5A–D) because as seen in Figure 5D at this enrichment level the simulated data give the best fit to the actual ones. Based on the fitting of different simulated distributions to the actual one, the range of the tag count in the actual data can be divided into four sections with low, medium, high, and ultrahigh tag counts respectively.
As marked by the dashed circles and lines in Figure 6, the three section boundaries are defined by the divergence of the simulated distribution based on the varying-background and uniform-site model from the actual distribution (the green and the black lines), the convergence of the simulated distribution based on the uniform-background and varying-site model from the actual distribution (the purple and the black lines), and the divergence of the simulated distribution based on the varying-background and varying-site model from the actual distribution (the orange and the black lines). Based on the models used to generate these simulated distributions, we can also infer the genomic identities of tag clusters found in the actual data. Tag clusters with low and high (including ultrahigh) tag counts are almost certain to be background and binding sites, respectively. Because there is a mixture of signals, the true identities of the clusters with medium tag counts are much less certain, and thus some form of thresholding is necessary. Figure 6 also shows that the part of the tag count that has a power-law distribution is supported by the background or the binding sites or both at low, high, and medium counts respectively. The right tail, diverged from the power-law distribution (Figure 4), occupies the ultra-high count section.
Reported in two recent studies [6],[7], ChIP sequencing is a newly-developed high-throughput method for genome-wide mapping of in vivo protein–DNA association. In these two studies, two different analytical methods were used to identify transcription factor binding sites. In the first study [7], a list of sites ‘known’ to be bound (the positives) and unbound (the negatives) by the transcription factor being studied is first compiled. Given this ‘gold standard’, the sensitivity and the specificity of the experiment at each threshold on the sequence read per region are then calculated. And finally a threshold is chosen to give both high sensitivity and high specificity. In the second study [6], a background model is first used to simulate the sequence read placement unto the genome in the absence of binding sites. The false discovery rate, defined as the ratio of the number of peaks at and above a peak height threshold in the simulated data to that at and above the same threshold in the actual data, is then calculated at each peak height as the threshold. And finally a threshold on the peak height is chosen to give a stringent FDR. For easy reference in our later discussion, we name the former the “known-sites” method and the latter the “background-simulation” method.
The known-sites method has the advantage in giving the sensitivity and the specificity of a particular ChIP-seq experiment at a chosen threshold. Its applicability is, however, problematic since it requires a “gold standard,” a list of true positives and true negatives. Conceptually, the validity of such a ‘gold standard’ is questionable given the dynamic nature of protein–DNA association—i.e., under different conditions a transcription factor has different DNA-binding profiles. Operationally, this method is also difficult to use. The prerequisite functional “gold standard” is rarely available, let alone a good one. Moreover, the “known” positives are biased towards binding sites with high enrichment of sequence tags, and as the majority of the genome is not bound by a transcription factor ever, it is an open question how many “true negatives” should be included in the calculation. That is, given the huge preponderance of negatives, it is very difficult to build a correctly balanced gold standard, which is essential for training an effective classifier [14].
Instead of using a “gold standard” to identify binding sites in ChIP-seq data, the background-simulation method uses a background model to simulate how sequence reads are distributed in a genome in the absence of binding sites. Since this method does not assume any prior knowledge about the binding sites of the transcription factor under investigation, it avoids major difficulties encountered by the known-sites method. In their study, Robertson et al used a background model that implicitly assumes uniform tag placement everywhere on the background. However, our simulation results show that the data generated by this uniform-background model agree poorly with the actual experimental data. Based on our further analysis, we can generate a better null distribution by using a more realistic, varying-background model that assumes most of the background is not enriched but at a few places it has a high enrichment level on a par with some binding sites.
In our analysis we estimated the background and the foreground together from the ChIP-seq sample data alone. However, if the negative control data from the experiments without immunoprecipitation are available, the estimation of the background becomes simpler as such experiments give a direct empirical estimate of the ChIP-seq background. Because our method can simulate the background alone, the negative control data can thus be easily accommodated. First the control data are used to estimate the parameters of the varying background model. The fitted model is then used to generate the null distribution of the tag count. And finally this null distribution is used to score the ChIP-seq data.
We also make improvement to the usage of the null distribution in the background-simulation method. In the study of Robertson et al, the false discovery rate is defined as the ratio of the number of peaks at and above a threshold in the simulated data to that at and above the same threshold in the actual data. The implicit assumption behind this definition is that the peaks identified in the simulated data are false positives and the number of them is equal to the number of false positives in the actual data. The first half of this assumption is reasonable, but the second half is unwarranted. For direct comparability, the same number of uniquely mapped sequence tags as contained in the actual data is used to simulate the null distribution on the background. Due to the finiteness of this number and the presence of binding sites (the true positives) in the actual data, the number of the peaks identified in the simulated data will be greater than the number of false positives in the actual data at any threshold. This discrepancy is more pronounced at lower thresholds. In fact, at low thresholds there could be more peaks in the simulated data than in the actual data. When this happens, the false discovery rate exceeds one, which is nonsensical. Instead of using the null distribution in such an ad hoc manner, we use it to assign each tag cluster found in the actual data a P-value to assess its statistical significance. We then adjust the P-values of the multiple-hypothesis tests to control the false discovery rate.
|
Subsets and Splits